Paper Radar · today in AI

★ MolDesignBench: Evaluating LLM-based Agent for Scenario-grounded Molecular Design

Thu, 24 Sep 2026 00:00:00 +0000

99% · Evaluation 99%, Agents 96%

Real-world molecular design remains challenging for large language model (LLM)-based agents. It requires them to interpret design contexts, satisfy multiple constraints, identify infeasible specifications, and reason over multi-step tool outputs. Existing benchmarks do not capture this complexity, focusing instead on explicit and narrow constraints, only feasible problems, and single-path solutions. To address this gap, we propose MolDesignBench, a scenario-grounded benchmark that more closely reflects real-world molecular design for evaluating tool-augmented LLM agents. MolDesignBench comprises 2K generation and optimization instances that combine implicit requirements embedded in design narratives with explicit property and functional-group constraints, including infeasible cases, and require the effective use of 17 specialized chemistry tools. Experiments across diverse frontier LLMs reveal low success rates--with the best achieving only $\sim43$\%--and frequent failures in implicit-constraint reasoning, infeasibility detection, and tool reasoning. The corresponding fine-grained failure-mode analysis identifies implicit constraint interpretation and infeasibility detection as the primary bottlenecks, establishing MolDesignBench as a rigorous testbed to guide future research on chemical agents. The benchmark, tool interface, and evaluation code are publicly available.

★ Uncheatable Eval: Dynamic Compression-Based Evaluation of Language Models

Kaifeng Tan, Yudong Li, Linlin Shen — Thu, 24 Sep 2026 00:00:00 +0000

99% · Evaluation 99%

Modern large language models are pretrained on massive datasets, making it difficult to prevent benchmark data from entering their training sets and undermining the reliability of evaluation results. Reliable evaluation is particularly challenging for base models, whose limited instruction-following ability complicates task-based assessment. We introduce Uncheatable Eval, a dynamic benchmark that regularly collects newly published text to evaluate base language models and reduce the risk of data contamination. Drawing on the relationship between a model's predictive ability and its ability to compress data losslessly, we use compression rate to evaluate how well models predict new text. We evaluate 80 models across 14 text categories, study how compression changes with context length, and examine the correlation between compression rate and zero-shot MMLU accuracy. Our results yield three main findings: (1) compression performance follows a consistent scaling trend with model size; (2) attention-based, hybrid, and recurrent models differ in how their compression performance changes as more context becomes available; and (3) lower compression rates are strongly associated with higher zero-shot MMLU accuracy. Code is available at https://github.com/Jellyfish042/uncheatable_eval.

★ PASTABench: Proactive Assessment of Sequential Trajectories for Agent Safety

Jiapeng Sun, Yujin Zhou, Han Zhu, Pengcheng Wen, Jiayi Zhou, Sirui Han — Thu, 24 Sep 2026 00:00:00 +0000

99% · Evaluation 99%

As Large Language Models (LLMs) evolve into autonomous agents that alter real-world states, ensuring operational safety across multi-step workflows has become a critical challenge. While recent work has moved beyond single-turn evaluation toward multi-turn paradigms, key limitations persist: step-level methods treat actions in isolation, missing how risks accumulate, while trajectory-level evaluations operate post-hoc, offering no opportunity for timely intervention. To address these limitations, we formalize Decoupled Proactive Safety Monitoring along three dimensions: whether to intervene, when to intervene, and what the risk is. We introduce PASTABench, a benchmark of 1,139 multi-turn trajectories spanning 5 risk categories and 13 subcategories. We further propose the Optimal Intervention Window (OIW), anchored by annotated Earliest-Signal and Trigger turns, to quantify intervention timeliness. Evaluation of 16 LLMs reveals that proactive intervention remains largely unsolved, with the best model achieving only 40.74% optimal-timing interventions. Fine-grained diagnosis further uncovers pervasive lexical overfitting: competitive safety scores of smaller models mask keyword hypersensitivity rather than genuine risk comprehension, as their proactive capability largely collapses once hazard vocabulary is neutralized.

★ Recognized but Not Produced: A Generation Benchmark for Culturally Specific Kinship Terms

Sahil Pardasani, Madhusudan Singh — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Current literature evaluates large language models (LLMs) on multilingual kinship understanding using multiple choice benchmarks, treating it as a recognition problem. We instead prompt five open weight LLMs to generate kinship terms in three non Western languages (Hindi, Tamil, and Korean) across two communicative tasks and pair this with a matched option-supported selection baseline. On identical relation language cells, GPT OSS120B selects the correct term in 90.67% of 75 valid cells but produces an accepted term in 36.00% of the corresponding attempts; Llama 3.370B shows the same pattern (77.92% versus 24.24%). Since the four-option condition displays the candidate terms and does not require script production, the difference is interpreted as an evaluation format gap rather than direct proof that lexical knowledge is intact. On explicitly specified L3 prompts, accuracy varies sharply, from GLM-5.1 at 72.29% to Llama-3.370B at 24.24%. The paternal-lineage advantage is language specific; it is large in Hindi but weak or reversed in Korean, while Tamil shared-term pairs provide a control for measurement variation. These results show that culturally specific kinship generation remains difficult even when the relationship is explicitly stated and motivate generation-based evaluation alongside multiple-choice testing.

★ Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court

Felix Ringe — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Judicial reasoning remains challenging for large language models (LLMs) to analyze. This paper contributes a sentence-level benchmark for evaluating the ability of LLMs to classify interpretive canons as articulated by Larenz in the tradition of Savigny. Our contributions are threefold. First, we operationalize this conception of interpretation as classification criteria. Second, we provide a dataset of decisions of the German Federal Constitutional Court annotated at the sentence level. Third, we report baseline evaluations of four LLMs from three model families under expert hand-written prompts, compared against prompts optimized with Genetic-Pareto (GEPA). Mean F1 over the seven binary subtasks clusters between 70.4 and 79.2 across models, with grammatical interpretation usually the easiest canon to identify and systematic interpretation usually the hardest; under the tested configuration, GEPA-optimized prompts do not systematically outperform the hand-written ones, suggesting that the expert prompts provide a meaningful baseline.

★ Count Evidence, Not Sentences: Tempered Evidence Fusion of LLM Judgments for Long-Text Value Measurement

Yuhe Wu, Rui Qian, Guangyu Wang, Yuran Chen, Yuanchao Zhu, Junjie Yang — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Large language models (LLMs) are increasingly used to measure public value orientations from long social media posts, yet such posts often mix background, quotations, concessions, and only a few stance-bearing sentences. Existing approaches either ask the model to predict a document-level label directly, which can be overconfident, or aggregate sentence-level predictions by majority or soft voting, which treat uncertain and decisive sentences as equally informative. We formulate long-text value measurement as a decision-fusion problem and propose Tempered Evidence Fusion (TEF), a training-free rule that weights each sentence's log-odds by its normalized information gain, as derived from a generalized Bayesian posterior. This makes the fused score nearly vanish for uncertain sentences while preserving the Bayes-optimal weight of decisive evidence. We further introduce Multi-event Insight Network Dimensions (MIND), a benchmark of 8,358 Chinese and English posts spanning five years of public events and six value dimensions. On MIND, TEF outperforms the strongest baseline among Direct, Majority Vote, and Soft Vote by an average of 4.5 accuracy points and 4.6 macro-F1 points across five LLMs and two languages. MIND dataset and code are available at https://github.com/Kzczc/ICASSP2027-TEF.

★ CAVEAT: Towards Robust Computer-Use Agents in Incentive-Misaligned Environments

Yuxuan Li, Will Epperson, Wesley Deng, Zezhou Huang — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Agents 95%, New model 33%

Computer-use agents (CUAs) increasingly act on behalf of users online. What happens when the environments they operate in have incentives that do not align with the user's? In online marketplaces, for example, platforms may favor some products over others, potentially steering agents away from the user's objective. Existing CUA benchmarks cover cooperative settings or explicit attacks, but do not test whether agents preserve user objectives when the environment itself has a stake in the outcome. We introduce CAVEAT, a controlled benchmark spanning nine marketplace environments and a taxonomy of eight common steering mechanisms. Across five model families, agents purchase the user-optimal product in 78.6% of matched-control episodes but only 17.3% when steering mechanisms are enabled. Larger models and increased reasoning improve robustness, but substantial failures persist. Our trajectory analysis and targeted ablations identify three points where steering enters the decision process: (1) agents distort the user's priorities, (2) prematurely narrow the set of alternatives they consider, and (3) commit before resolving decision-relevant evidence. Guided by this diagnosis, we develop CAVEAT-Harness, which directly targets these failure modes and raises user-optimal purchasing by 55.0%. Targeted post-training further improves a smaller open model. These results establish incentive robustness as a distinct challenge for delegated agents, diagnose how it fails, and show that targeted interventions can substantially improve it.

★ Guides That Cause Actions: An Offline Study of Guide-Action Mutual Reinforcement in Multimodal Web Agents

Chengguang Gan, Yunhao Liang, QingHao Zhang, Shiwen Ni — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Agents 97%

Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.

★ Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models

Kian Shamsaie, Iman Modarressi — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.

★ Cross-Lingual Legal QA for Vietnamese Labour Law: Retrieval, Translation, and Verifier-Guided Correction

Nguyen Minh Chi, Mo El-Haj, Nguyen Ha Thanh, Dawn Knight, Paul Rayson — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Hallucination 89%

Cross-lingual legal question answering must retrieve statutes across languages while preventing unsupported legal claims. We introduce a bilingual evaluation suite of 231 Vietnamese--English question--answer pairs from Vietnamese labour law. Of these, 75 are additionally annotated for five challenging legal reasoning phenomena. We evaluate a verifier-guided pipeline that decomposes answers into claims, checks citation reachability and entailment, and corrects citation failures and contradictions. We also introduce six automatic diagnostics for faithfulness to retrieved evidence, covering citations, modality, exceptions, procedures, conclusions, and evidential support. Experiments show that learned-sparse retrieval performs poorly for English-to-Vietnamese retrieval (R@5~=~0.032), whereas dense retrieval reaches 0.358 and slightly outperforms hybrid retrieval. Translation placement has no statistically detectable effect on these automatic diagnostics in our controlled comparison and supporting sensitivity analyses. Verifier-guided correction improves citation preservation by $0.022$--$0.034$ at the system level but produces no reliable gains in the remaining dimensions. Human evaluation further shows that the automatic diagnostics do not fully align with human judgements of answer quality.

★ Forecast Workflow Bench: Evaluating Language-Model Decisions with Budgeted Forecast Tools

Shunya Nagashima — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Agents 81%

Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints.

★ Not What You Meant: Can LLMs Follow a Specified Negation Semantics?

Qiming Bao, Agnieszka Mensfelt, Michael J. Witbrock, Kostas Stathis — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Negation does not carry a uniform interpretation across domains. In legal, regulatory, and medical reasoning, the intended interpretation depends on the reading in force -- open- versus closed-world, two- versus three-valued, and credulous versus skeptical. We study which reading of negation large language models adopt by default and whether they can override that preference when a different reading is explicitly specified. To this end, we introduce NAFBench, a procedural generator of solver-certified instances spanning four semantic viewpoints: SLDNF, well-founded semantics (WFS), and credulous and skeptical reasoning under stable-model semantics. The generator emits ground normal logic programs with controlled depth, width, and cycle structure. Each program is solved under all four viewpoints using SWI-Prolog, a well-founded semantics solver, and clingo, yielding up to four divergent labels. The programs are then verbalized into natural language under multiple framings and rule orderings that leave the answer invariant. The results expose a consistent gap. Across open-source models, following a specified negation semantics remains unsolved: the strongest models score 59--74% across the four semantic viewpoints, while the weakest score 31--67%. All models are order-sensitive on more than half of logically identical rule shufflings, while the two weaker models frequently overcommit on well-founded "undefined." Two frontier models reach 100% on the main fixed-complexity evaluation set, and a third, o4-mini, is near-perfect, falling only to 81% on well-founded "undefined." Delegating reasoning to a solver, fine-tuning on certified traces, or forcing an explicit three-valued verdict each partly closes the gap.

★ MWE-ECL: Recoverable Long-Range Context Does Not Always Override Local Lexical Priors

Wei He, Aline Villavicencio, Rodrigo Wilkens, Zhenyun Deng — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%

Long-context evaluations often test whether a model can recover distant evidence, but recoverability does not guarantee behavioral influence. We test the prediction that a distant discourse anchor can remain explicitly recoverable yet fail to change the locally preferred reading of a familiar multiword expression; such failures should concentrate when the model's no-anchor default conflicts with the anchor, while prior-correct decisions remain largely preserved. We introduce Multiword Expression Effective Context Length (MWE-ECL), a bilingual diagnostic whose matched anchor-retrieval, no-anchor prior, and interpretation prompts measure explicit recoverability, model-observed defaults, and anchor-conditioned decisions, respectively. Across eight English deployment panels on a shared 0-128K grid, retrieval-control accuracy on prior-conflict items is 0.989-1.000, prior-conflict override spans 0.806-1.000 (0.809-1.000 after conditioning on correct retrieval), and preservation of prior-correct decisions remains 0.977-1.000. A same-call control querying retrieval and interpretation in one prompt reproduces the gap for DeepSeek V4 Pro (1.000 retrieval versus 0.900-0.920 interpretation), showing that separate invocations are not its sole explanation; smaller or absent gaps in the other two models bound its generality. For DeepSeek V4 Flash, separate prompt-fit tests retain perfect retrieval with lower interpretation at 512K and 1M, while foil-consistent cues shift the no-anchor prior far more than retrieval; cross-model cue effects are heterogeneous. A separately reported 10-family Chinese subset shows similar descriptive gaps, but imperfect retrieval for some models prevents an integration-only attribution. MWE-ECL therefore evaluates whether explicitly recoverable distant context changes a competing local semantic decision.

★ When Context Misleads: In-context Learning with Jurisdiction in Large Language Models

Pei-lin Li, Qingle Liu, Junyang Feng, Siyu Li, Sunqi Fan, Xin-Sheng Chen — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, New model 91%, Hallucination 79%

In-Context Learning (ICL) has become a cornerstone of modern LLM deployment. However, existing ICL post-training methods have a critical blind spot: they excel at extracting patterns from demonstrations while often neglecting context authority, the ability to determine whether contextual information should govern the final answer. To benchmark this capability, we introduce FakeContextBench, which contains pseudoscientific claims across seven domains. Our evaluation of commercial and open-source models shows that large-scale pre-training alone is insufficient for reliable context-authority discrimination. Moreover, prevalent ICL fine-tuning methods can increase susceptibility to misleading context, reducing reality accuracy by up to 14.95 percentage points relative to the base model. To address this trade-off, we propose Jurisdiction In-Context Learning (J-ICL), a post-training framework that incorporates context validation into the training objective. Across four model backbones, J-ICL improves ICLEval by an average of 5.84 percentage points and reality accuracy by 9.20 points over the corresponding base models. It also raises the Reality Rate by an average of 18.09 points relative to MetaICL and Symbol Tuning. These results demonstrate that ICL capability and resistance to deceptive context can be improved together. The benchmark is available at https://github.com/peilin717/FakeContext-Bench.

★ Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents

Jiaxing Li, Lei Song, Rui Dong, Youyong Kong — Thu, 24 Sep 2026 00:00:00 +0000

98% · Agents 98%, Evaluation 32%, New model 31%

Small and medium-sized language models offer cost-effective executors for tool-using agents, making them attractive for local and large-scale deployment. However, in long-horizon and stateful environments, they often make structural errors such as missing required observations, performing premature writes, repeating failed calls, and violating action preconditions. These errors can lead to incorrect state updates, policy violations, and costly or irreversible consequences, making reliable tool execution a critical deployment challenge. Existing fine-tuning approaches require substantial data and computation, while flat memory may retrieve failed actions without preserving their causal context or safety conditions. In this paper, we propose FRESH, a Failure-aware Retrieval framework over Experience-Structured Heterogeneous graphs, which transforms historical successes and failures into structured external experience for tool-using agents. By explicitly modeling the dependencies among tasks, actions, errors, repairs, and execution conditions, FRESH helps frozen language models reuse reliable strategies, avoid recurring failures, and make safer decisions in stateful tool interactions. Experiments on $\tau$-Bench and AppWorld with multiple open-source models show that FRESH consistently improves task success and tool-use reliability over no-memory agents and representative memory-based baselines.

★ Can LLMs Catch a Rigged Backtest? A Clean-Control Calibration Benchmark

Makar Ulesov, Vladislav Smirnov, Omar Ibrahim, Arsenii Bobovnikov — Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Hallucination 43%

Backtest auditing is a calibration problem: high flaw recall is not useful when the model falsely flags matched clean strategies. We build a 96-item paired benchmark in which every flawed backtest has a clean control that holds strategy, dates, code style, labels, and reporting scaffold fixed while changing one methodology detail. A deterministic scorer separates flaw recall, clean-control false positives, evidence localization, and fix relevance. Over 1440 cached audits from four text endpoints, the primary DeepSeek auditor reaches 100.0\% closed and clean-aware code recall, but open prompts over-flag 93.8\% of clean code controls, and clean-aware all-three specificity is 87.5\% even where recall saturates. A clean-aware warning drops DeepSeek code false positives from 20.8\% (95\% CI 11.7--34.3) to 0.0\% (0.0--7.4) at unchanged recall, while the budget anchor still flags 38/48 clean controls under the same prompt. Reporting recall alone would rank three of these four models identically; reporting the clean-control rate separates them by 79 points.

★ StudentBench: AI and human tutoring yield equivalent GRE learning gains

Thu, 24 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Efficiency 43%

Artificial intelligence offers an unprecedented opportunity to augment human capabilities, yet progress at the frontier has focused primarily on advancing model capabilities. We introduce StudentBench, a suite of AI teaching evaluations and a public platform that enables large-scale data collection with over 175,000 student-AI messages to study whether large language models (LLMs) produce learning gains equivalent to human tutoring. Using StudentBench, we measured learning gains on Quantitative and Verbal GRE questions across 2,383 human participants receiving AI tutoring, human tutoring, or no tutoring. We establish that AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains (p = .015), and in five of the seven GRE domains, the best performing AI tutor surpassed the human tutor, on average. In a second study, expert human tutors compared LLM-generated lesson plans and practice problems through 2,028 pairwise rubric evaluations. Together, the two studies clearly separate AI tutors across: (1) lesson planning, (2) practice-problem creation, (3) conversational pedagogy, (4) cost, and (5) engagement. Surprisingly, one AI tutor achieved learning gains equivalent to human tutoring (p = .044) at 918 times lower cost (USD 0.0052 for AI versus USD 4.81 for human, per percentage point gained). For Quantitative GRE sessions, faster AI replies correlated with more student messages, more messages with more correct practice, and more correct practice with larger learning gains (all p < .002). The StudentBench platform is freely available at https://studentbench.org.

★ TwinCheck: Evidence-Grounded Negative-Twin Verification for Stateful Tool Agents

Jiaxuan Dai, Tianyi Huang — Thu, 24 Sep 2026 00:00:00 +0000

97% · Agents 97%, Evaluation 82%

A single locally plausible tool call can derail an otherwise successful agent trajectory. Suspicion alone does not justify intervention, because the replacement itself can introduce the very failure verification is meant to prevent. We introduce TwinCheck, an inference-time verification policy that considers replacement only when the trace satisfies an evidence condition tied to a trace-local failure hypothesis. It constructs a trace-grounded counterfactual alternative, a negative twin, and replaces the agent's proposal only if the twin passes structural checks and the pairwise verifier prefers it in both candidate orders. For paired evaluation, exact replay holds the agent's parsed responses and actions fixed until the first accepted replacement, separating intervention effects from resampling. In the primary analysis of 159 multi-turn BFCL V4 tasks with complete exact-replay pairs, the complete policy raises task success for GPT-5.6 Sol from 45.3% to 58.5% (95% task-bootstrap CI [8.2, 18.8]), with no observed success-to-failure regressions. Together, these findings recast execution-boundary repair as a constrained comparison, making the counterfactual action itself the object of verification.

★ Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach

Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, New model 39%

Text-to-speech systems increasingly process user-generated text (UGT) such as ppl and imo, whose pronunciation must be inferred from the canonical rather than surface form. We introduce UGTPhon, the first grapheme-to-phoneme (G2P) benchmark for UGT in English, Vietnamese, and Korean, together with an inference-grounded taxonomy for fine-grained diagnosis. Existing G2P models and frontier LLMs exhibit a systematic canonical-to-non-canonical performance gap, reaching up to 66.8 PER points. As a benchmark baseline, we propose a simple compositional G2P approach that incorporates canonical-form evidence through exact-match lookup and staged decoding. Across matched ByT5 and Qwen2.5-0.5B backbones, explicit canonical-form modeling consistently reduces non-canonical G2P errors. The 0.5B variant also performs competitively with much larger few-shot frontier LLMs, highlighting the benefit of explicitly modeling canonical-form inference for UGT phonemization.

★ Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models

To Duy Hinh, Nguyen Le Quoc Anh, Phan Van Tri, Khuong Nguyen-An — Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Hallucination 47%

Vietnam's Personal Data Protection Law (Law No. 91/2025/QH15) and Decree No. 356/2025/ND-CP, effective January 1, 2026, require organizations to establish and maintain Records of Processing Activities (RoPA). Manual RoPA preparation is labor-intensive, while cloud-hosted large language models (LLMs) may conflict with data-sovereignty requirements. We propose RoPA Manager, a system for automated RoPA information extraction using hybrid retrieval that combines lexical ranking over tsvector, dense-vector search, Reciprocal Rank Fusion (RRF), and locally deployed LLMs. We introduce a Vietnamese RoPA benchmark with 32 organizations, 77 processing activities, 12 field groups, and 4,338 reference values. Evaluation is reported at three distinct levels. The automated scorer, tested on perturbed data without invoking an LLM, achieved F1 = 0.9493 [0.9436, 0.9548]; this measures scorer robustness rather than end-to-end extraction accuracy. End-to-end extraction achieved token coverage of 50.04-55.25% against the reference labels. Two independent experts reviewed 1,558 reference values (35.9% of the benchmark), found no incorrect values, and achieved 99.68% agreement with PABAK = 0.9936. Value-level precision was not measured. Across 32 paired scenarios on a 24 GB GPU, locally deployed Qwen3.5-27B-GPTQ-Int4 showed no statistically significant difference from cloud-based DeepSeek-V4-Flash (difference 0.20 percentage points in favor of DeepSeek, 95% CI [-0.93, 1.32], p = 0.72), while Gemma-4-31B performed significantly worse (p < 0.01).

★ Can Jev Judge Radiology Reports? Evaluating a System One Model for Clinical Factuality

Jiaju Huang, Hao Yang, Xinyu Ma, Xinglong Liang, Kunyan Cai, Junqiang Ma — Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Hallucination 89%, Efficiency 71%

An AI-generated radiology report can resemble a physician's report while omitting an abnormality, adding an unsupported finding, or reversing its presence. Measuring these factual differences is essential for evaluating report generators. We study Jev, a System One decision model, as a simple, low-cost judge of agreement with physician-written reference reports. Our evaluator checks whether each statement is supported by the other report and combines these judgments in both directions to capture unsupported claims and omissions. A single-question configuration reaches Kendall correlations of 0.573 on RadEvalX and 0.398 on RadEvalExpert with expert error counts, outperforming an open natural language inference judge under matched decomposition and aggregation. One support question per statement retains similar expert agreement to seven while using 43-45% fewer judgment input tokens. At the documented API price, judgments cost under three cents per hundred report pairs, excluding local decomposition. In a separate controlled-error test, Jev detects false negation with an AUROC of 0.977. Local RadMatch achieves stronger agreement on clinically significant errors in both expert datasets and on total errors in the shared RadEvalExpert subset. Finding-count and error-scope analyses show that benchmark agreement reflects report size and error definitions as well as medical error detection. These results support Jev as a practical judgment component for measuring factual differences in generated radiology reports and identify where more elaborate evaluation remains valuable.

★ Improving LLM-based Autonomous Web Agents with Filtering

Zhitong Guo, Jing Yu Koh, Ruiyu Li — Thu, 24 Sep 2026 00:00:00 +0000

97% · Agents 97%, Evaluation 36%, New model 33%

Autonomous web agents, powered by Large Language Models (LLMs), have garnered significant attention for automating various web-based tasks with multi-step reasoning and decision-making capabilities. An open research question in the development of these agents lies in the format of the webpage input. Raw HTML source code, with its extensive and often irrelevant details, poses difficulties for LLMs with limited context windows. To address this challenge, we first reproduce baseline models such as GPT-3.5 and LLaMA-2-70B on the WebArena (Zhou et al., 2023) benchmark, identifying common failure modes. We then propose two retrieval strategies to filter out irrelevant context for LLM agents. We develop DeBERTa-based and T5-based models that rank HTML elements by their relevance to the task. We fine-tune them on Mind2Web trajectory data and transfer them to WebArena. Experiments show that our DeBERTa-based model improves the success rate of the LLaMA-2-70B LLM agent on WebArena from 1.97% to 2.96%. Moreover, we develop a zero-shot ColBERT-based retriever that is able to retrieve the ground-truth element with a recall of 0.52 on Mind2Web and 0.47 on WebArena.

★ Evaluating Open-Weight LLMs for Turkish Domain Documents Under Retrieval and Hardware Constraints

Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%

Most Turkish-capable large language models (LLMs) are evaluated using general-purpose benchmarks rather than long, structurally complex domain documents. This paper evaluates five open-weight 7B-8B models for Turkish document question answering under a resource-constrained local deployment setting. The primary benchmark contains 100 systematically validated questions derived from a 109-page industrial R&D report, and the evaluation protocol is replicated using a second 112-page public-sector report and an independently constructed 100-question set. All models are evaluated locally on an NVIDIA RTX 3050 laptop GPU with 6 GB VRAM using controlled prompting, decoding, and 4-bit quantisation. The principal methodological contribution is an evidence-annotated evaluation protocol that separates retrieval failure from downstream model reasoning failure without requiring additional model calls. On the primary benchmark, end-to-end accuracy ranges from 49% to 75%. Seven lexical, dense, and hybrid retrieval configurations are additionally compared using 95% Wilson intervals and exact paired McNemar tests; none significantly outperforms the character TF-IDF baseline on either document. Evidence recall saturates differently across the two reports, showing that retrieval and effective context capacity can be binding constraints for some documents but not others. These results demonstrate that model selection, retrieval behaviour, and hardware limits must be evaluated separately when deploying open-weight LLMs for Turkish domain documents.

★ Evaluating Feedback Focus and Pedagogical Adaptivity in LLM-Generated Feedback on Student Writing

Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%

We investigate whether state-of-the-art large language models (LLMs) generate feedback that reflects the pedagogical practices of expert teachers in terms of feedback focus and adaptivity. Previous evaluation efforts have examined feedback characteristics, its impact on learning, and its target, yet the focus of feedback and its adaptivity remains largely overlooked. To bridge this gap, we adopt and refine Narciss's taxonomy into seven feedback focus types to annotate teacher and LLM-generated feedback across three university writing courses. We release FeedType, a benchmark containing annotated teacher and LLM feedback from six LLMs under three prompting strategies. We assess the coverage and distribution of feedback focus types, and examine whether LLMs adapt their feedback across draft stages and student performance levels as an expert instructor does. Our findings show that while most LLMs cover most feedback focus types, they fail to reflect teacher feedback distributions and show varying levels of adaptivity, with none matching the teachers' adaptive behavior. We believe FeedType will support future research on pedagogical alignment in LLM feedback generation.

★ An Open Pipeline and Dashboard for Systemic-Risk Evidence under the EU AI Act's Code of Practice

Thu, 24 Sep 2026 00:00:00 +0000

97% · Evaluation 97%

Claims about AI safety reach audiences well beyond the AI community, yet many rely on opaque evidence or static assessments, when supporting evidence is accessible at all. We present the Systemic Risk Index, an open evaluation pipeline and dashboard built to make empirical evidence more transparent and traceable to the public. Our work organizes 19 public benchmarks into four systemic-risk categories defined by the EU GPAI Code of Practice---CBRN, cyber offense, harmful manipulation, and loss of control---and evaluates models using harm-preserving perturbations and simulated deployment contexts. The interactive dashboard lets users alternate between average and worst-case aggregation, vary how model capability affects the aggregate score, and trace each risk rating to its benchmark evidence. Across 18 models, scores fall by 14 to 37 points under worst-case aggregation, highlighting information that can be hidden by an average assessment of model risk. LLM judges show agreement with human graders comparable to human--human agreement ($\kappa = 0.78\text{--}0.82$), and a blind audit finds that $83\%$ of sampled transformations preserve the original harm. In a survey ($N = 21$), most participants report that scores are easy to understand and that the dashboard encouraged them to view model evaluations under different settings

★ Beyond Overlap: Estimating the Causal Effect of Benchmark Exposure

Divyansh Singh — Thu, 24 Sep 2026 00:00:00 +0000

96% · Evaluation 96%

Evidence that evaluation material entered training does not reveal how much it affected evaluation. This distinction leaves a contaminated benchmark score difficult to interpret: provenance can establish contact, but only a counterfactual can quantify the performance attributable to that contact. We present LeakScale, an interventional framework for estimating this missing quantity. LeakScale creates fresh executable tasks that require private, family-specific information absent from and non-derivable from the public task, controls access to that information, and estimates the resulting control-adjusted change in executable accuracy. Across 2,048 unique families, two model families, two executable domains, and 262,144 generations, exposure improves accuracy in every model-by-domain combination, with gains ranging from +7.17 to +27.31 percentage points. These findings separate two empirical questions that are often conflated: whether benchmark contact occurred and how strongly a reported score depends on it. LeakScale makes the latter directly measurable.

★ Hunyuan-A13B Technical Report

Tencent Hunyuan Team, Ao Liu, Botong Zhou, Can Xu, Chayse Zhou, ChenChen Zhang — Thu, 24 Sep 2026 00:00:00 +0000

96% · New model 96%, Efficiency 90%, Hallucination 45%

We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.

★ StateComp: Learning When to Compress History in Long Horizon Agents

Mingxuan Wang, Hongyue Chen, Yinglong Guo, Fei Luo, Chao Ning, Bo Wang — Thu, 24 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, New model 74%, Agents 52%

Long-horizon agents continuously accumulate interaction history during task execution, yet the importance of past interactions changes as the agent state evolves. Existing context management methods largely compress history based on fixed windows, periodic schedules, or current relevance, overlooking a more fundamental question: when has a past interaction become safe to replace? Premature compression may remove information still needed for future actions, while overly conservative retention leads to substantial context overhead. To address this, we propose State Conditioned Compression (StateComp), a framework that determines when historical interactions can be safely compressed according to the current agent state. StateComp constructs KEEP and READY supervision through a two-stage annotation procedure and trains an imbalance-aware router on hidden representations from a frozen language model. A bounded state representation further reduces the cost of evaluating long histories, while adjacent READY interactions are grouped into continuous spans and replaced with compact summaries during execution. Experiments on WorkBuddyBench show that StateComp reduces total agent and summarization tokens by 52.27% while maintaining task performance, and achieves a 12.67-fold speedup in representation extraction.

★ Live Assistant: Learning Whether, When, and Whom to Assist in Real-World Live Social Streams

Shujian Gao, Jiamei Yan, Yuchen Yang, Penghao Zhou, Qinglei Wang, Tiehan Fan — Thu, 24 Sep 2026 00:00:00 +0000

96% · Evaluation 96%, New model 80%

Livestreams are long-lasting interactive environments where audiovisual content, viewer activity, host behavior, and platform signals evolve together, creating assistance needs that emerge from the stream itself. We introduce \liveassistant, a framework for mixed-initiative, role-conditioned assistance that formulates livestream interaction as four coupled decisions: \textbf{whether to act, when to act, whom to address, and what to communicate}. At each 10-second interval, one autoregressive policy consumes native audio and video with synchronized comments, gifts, viewer dynamics, and room metadata, then selects \textsc{OBS}, \textsc{MEM}, or \textsc{ANS}. \textsc{OBS} remains silent, \textsc{MEM} records a private semantic update, and \textsc{ANS} specifies a recipient, task, and grounded message. To support this task, we build a trajectory engine that reconstructs real livestream sessions into structured causal supervision, yielding over 320 hours of optimization trajectories and a human-reviewed benchmark of 275 clips and 13,812 decision intervals. We train the policy with Marker-Aware Multiturn Supervised Fine-Tuning (MA-MSFT), which strengthens sparse structured decisions, followed by Streaming Multiturn GSPO (SM-GSPO), which optimizes self-generated trajectories with turn- and trajectory-level credit. On the held-out benchmark, \liveassistant reaches 71.14 state accuracy, 72.67 recipient accuracy, and 58.41 task accuracy, with consistent gains over representative streaming and general multimodal baselines. Together, the formulation, benchmark, and training framework establish livestream assistance as selective participation in a shared social stream.

★ Attention Routing Stabilizes Early: Working-Set Inference for Recurrent Language Models

Ke Wan, Chen Chen — Thu, 24 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, Evaluation 39%

Recurrent language models repeatedly apply shared network blocks to refine latent representations, but standard inference recomputes global attention at every recurrent step. We study attention dynamics across recurrent depth and find that attention support and distributions stabilize substantially earlier than hidden states and attention outputs. This suggests a two-stage structure: early steps discover a sparse working set of relevant context, while later steps refine representations over largely the same routing support. Motivated by this structure, we introduce WISE (Working-set Inference with Support Exploitation), a training-free method that uses unrestricted global attention during early recurrence and later reuses directly discovered block-structured support while keeping recurrent depth and within-support attention computation dynamic. Controlled interventions show that recurrent discovery is important and that support-only reuse better preserves model behavior than more restrictive attention-reuse alternatives. Across multi-hop QA benchmarks, WISE largely preserves full-attention performance, while context scaling reveals increasingly sparse working sets and greater efficiency gains. Quality is largely preserved through 2K context, with a measurable loss at 4K. An optimized sparse-attention implementation achieves up to a 1.76x attention speedup over native FlashAttention at 4K and a 1.36x speedup for the full 32-step attention trajectory. Our code is available at https://github.com/tbn5pj/WISE_code.

★ PRISM-VLM: A Multi-Axis Discriminative Benchmark for Compact Vision-Language Models

Sanghee Park, Kee-Eung Kim — Thu, 24 Sep 2026 00:00:00 +0000

96% · Evaluation 96%

Compact vision-language models (VLMs) now power a growing share of multimodal applications. The benchmarks used to compare them, however, inherit a frontier-centric design: each model is reduced to a single accuracy number, narrowing the inter-model gap on saturated suites and pressing models into low-score bands on harder ones. We introduce PRISM-VLM, a multi-axis discriminative benchmark that scores every item along seven axes covering the recurring failure modes (task quality, behavioral robustness, and capability bottlenecks) and combines them into a single PScore, with items recycled from fifteen public benchmarks. Across compact VLMs from the past two years, PScore separates model pairs more reliably than prior single-axis benchmarks under an item-level paired bootstrap, and surfaces behavioral differences these benchmarks average away. Even models with statistically indistinguishable PScores diverge sharply along the per-axis profile, particularly on sycophancy, which is nearly orthogonal to single-prompt accuracy. We will release the full pipeline, prompts, and per-item annotations.

★ When Parallel Drafter Meets Parallel Speculative Decoding

Fuliang Liu, Xue Li, Kun Qian, Zhibin Wang, Wanchun Dou, Wenyuan Yu — Thu, 24 Sep 2026 00:00:00 +0000

96% · Efficiency 96%

DSpark-style parallel drafters have made speculative decoding highly effective, yet their draft phase remains serialized on the critical path of every round. Parallel speculative decoding (PSD) overlaps drafting with verification, yet existing methods must guess the accepted prefix and bonus token in advance: a wrong guess reverts the whole batch to serial drafting. We present DPara, a PSD framework that reuses effective parallel drafters yet guarantees backbone--verification overlap in every round, thereby eliminating this probabilistic fallback altogether. While the target verifies, DPara's diffusion backbone precomputes draft representations for every acceptance boundary with the bonus left unspecified; a lightweight autoregressive head then combines the revealed verification outcome with the matching precomputed representation to emit the next round's draft tokens almost instantly---fully parallelizing the dominant backbone forward with verification and leaving only the negligible head cost serial. Experiments on Qwen3-8B and Qwen3-14B across seven math, coding, and chat benchmarks show that DPara achieves average speedups of $3.21\times$ and $3.52\times$ over autoregressive decoding, surpassing the strongest serial and parallel speculative decoding baselines alike.

★ ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning

Ming Ma, Yi Zhu, Yiran Zhong, Feida Zhu, Chonghan Liu, Pengkun Jiao — Thu, 24 Sep 2026 00:00:00 +0000

96% · Agents 96%, New model 91%, Evaluation 58%

Long-horizon agentic tasks require an agent to modify an environment through a sequence of tool calls, with success determined by the final state. The standard recipe assigns a single outcome reward at the end and compares trajectories sampled for the same task. As a result, a group with no successful trajectory yields no training signal, failed attempts cannot be told apart by how close they came to completion, and turns that advance the task receive the same credit as turns that only query the environment. Prior work refines the unit of comparison from the trajectory to the step, or trains a reward model to supply intermediate signal: the former still derives its signal from final success alone, and the latter estimates it with a model. We observe that the acceptance checks that decide success can also be run on intermediate states, so progress is as verifiable as the outcome. We propose ProCredit, which turns this verified progress into credit: it reruns the acceptance checks after each turn, rewards the turn by its change in progress, and uses these rewards to assign credit both across attempts at the same task and across the turns within a trajectory. Starting from Qwen3.5 base models at three scales on AppWorld, ProCredit outperforms outcome-reward baselines and progress-based baselines in task completion rate at every scale on both test sets, exceeding the strongest outcome-reward baseline by 4.1 percentage points at 4B, and results in a second environment show the same direction of improvement. Ablations show that adding the final progress to the trajectory score alone does not improve performance: the gain comes from crediting progress to the turn where it occurs.

★ Learning When Not to Listen: Selective Anti-Interference Pretraining for Language Models

Jinchang Zhu, Haowei He, Yi Ding, Rong Fu, Nie Xiaojian, Shuangyong Song — Thu, 24 Sep 2026 00:00:00 +0000

96% · New model 96%, Evaluation 42%

Language models can over-condition on irrelevant preceding text: predictions already supported by local context may still change when distant, unrelated prefix tokens are perturbed. This interference is especially consequential in long, packed, or distractor-heavy contexts, where useful evidence and irrelevant spans coexist. We propose Selective Prefix Anti-Interference Regularization (SPAR), a pretraining objective for selective anti-interference. SPAR runs the original sequence and a corrupt-prefix input in which only the far prefix is changed, then uses a short-context sufficiency gate and a gated KL objective to stabilize locally supported suffix predictions. The gate operationalizes a model-based estimate of whether the far prefix supplies additional information about the target token. Mechanism analyses show that the gate identifies locally sufficient tokens and sharply reduces prefix sensitivity on gate-selected suffix tokens. In continued training on pretrained base models, SPAR improves RULER across Qwen2.5-0.5B, Qwen2.5-3B, Llama-3.2-1B, Llama-3.1-8B, and GPT2-XL under equal counted training compute; pretraining experiments further show gains on both RULER and NoLiMa. These results show that selective anti-interference is an effective objective-level signal for robust context use.

★ Are Stated Reasoning Steps Causally Load-Bearing?

Abhiram Bhupatiraju, Rayan Nyaupane — Thu, 24 Sep 2026 00:00:00 +0000

95% · Evaluation 95%, Hallucination 32%

Chain-of-thought (CoT) monitoring assumes that the reasoning a model writes reflects the computation that directly produces its answer. Previous faithfulness metrics have been predominantly behavioral, as they simply edit the reasoning text and observe the resulting answer. However, our methodology aims to measure faithfulness causally at the activation level, specifically on self-generated reasoning. Unlike previous causal audits, which measure degradation, our interventions carry a known predicted target. In this way, each patch should switch the answer to a specific counterfactual entity derivable by construction. Specifically, we use synthetic multi-hop lookup tasks (2-6 hops). We patch the residual stream at the token span where the model states each intermediate step with the corresponding activations from a counterfactual run. For Qwen3-4B, 76.9% +/- 2.8% of stated steps are causally load-bearing (CLB) at the most responsive mid-network layer (random-position null: 11.3%; patching the underlying prompt fact: 83%, so stated steps carry approximately 96% of the achievable effect). Moreover, the standard behavioral test on the same items yields 88.2%, which overstates causal faithfulness by 11.4 percentage points (item-matched; 111:14 discordant pairs, p < 1e-15) and, for the easiest items, by up to 20 percentage points. This gap also has a clear capability dimension. Qwen3-1.7B is far less causally faithful overall (54.8%), with its faithfulness collapsing as reasoning depth increases (68% at 2 hops to 30% at 6), while Qwen3-4B remains relatively flat. Although stated reasoning can be causally meaningful, standard behavioral tests tend to overestimate its causal faithfulness, particularly on easier examples where model reasoning appears most fluent.

★ TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

Jie Yang, Yan Zheng, Jiarui Sun, Xiran Fan, Junpeng Wang, Liang Wang — Thu, 24 Sep 2026 00:00:00 +0000

95% · Agents 95%, Evaluation 57%, New model 37%

Time series agents answer analytical questions by calling external tools, and which tools they carry is decided by people before the agent runs. However, we identify two failures in this setup. Human-Agent Tool Misalignment: a library of 21 expert-curated tools helps on some tasks and hurts on others, dropping anomaly accuracy under every backbone we test. Silent Harm: one round of generic self-revision changes 147 answers and breaks 56 of them, while the final score moves by less than a point. Both follow from the same gap: whether a tool helps is decided question by question at runtime, while tools are supplied in advance and judged by a single average. To address this, we propose TimeEvo, which clusters an agent's diagnosed failures into capability gaps, plans a measurement for each, synthesizes evidence-only tools that fill them, and admits the candidate library only through a paired admission gate. Experiments on ten time series QA tasks and three backbones show that TimeEvo, starting from an empty library, improves accuracy on every task and every backbone, and that a library grown on a cheap model still gains when it is installed into stronger ones. Code is available at https://github.com/Muyiiiii/TimeEvo.

★ CART: Closed-Loop Adaptive Red Teaming for Large Language Models

Dongdong Zhang, Tengchao Lv, Yilin Jia, Yuzhong Zhao, Yupan Huang, Wenshan Wu — Thu, 24 Sep 2026 00:00:00 +0000

95% · Evaluation 95%, Agents 41%

Automated red teaming often replays a fixed set of prompts, which measures known risks but cannot learn from failures found during testing. We present CART (Closed-Loop Adaptive Red Teaming), a framework that uses each result to guide what it tests next. CART begins with broad risk coverage, follows weaknesses that emerge, keeps new probes diverse, and records the evidence and source of every finding. It separates the Challenger that creates tests, the Target being tested, which may be a text-only model or a bounded tool-using agent, and the Judge that evaluates the results, allowing these roles to be studied independently. Across three evaluation families (Frontier, JAH, and Agentic), CART discovers more failures and higher average risk than static seed replay for every Target with an available baseline. The gains extend to tool-mediated agent tests, suggesting that contextual adaptation can reveal weaknesses that direct prompt replay does not exercise. These results describe what the test policies discover, not how often failures occur in real deployments. We also find that Challenger-Judge choices affect the evidence uncovered, highlighting the need for role separation and independent review. Overall, CART turns red teaming from a one-time checklist into a continuous, adaptive, and auditable search for model and agent weaknesses.

★ The Path Matters: Evaluating Small Language Models Beyond Answer Accuracy in KGQA

Eduin E. Hernandez, Sergio A. Diaz, Luis F. Garcia, Nurassyl Askar, Stefano Rini — Thu, 24 Sep 2026 00:00:00 +0000

95% · Evaluation 95%, Hallucination 33%

Small language models (SLMs) are increasingly paired with knowledge graphs (KGs), yet end-to-end KG question answering conflates graph access, search, navigation, reasoning, and answer generation. This coupling makes it difficult both to determine whether an SLM can faithfully execute the reasoning path implied by a question and to attribute failures to navigation rather than to other stages of the pipeline. We isolate this capability by employing the THESEUS navigation and traceability framework and using frozen, off-the-shelf SLMs as local action policies. At each hop, the environment exposes the legal outgoing graph actions, and the model selects one executable graph action and decides whether to stop, without task-specific parameter updates, model-controlled beam search, or free-form answer generation. This controlled setting allows us to evaluate terminal-answer accuracy with Hits@1 together with path fidelity, using Path Edit Distance (PED) as the primary trajectory metric. Across the Kinship and MQuAKE-ST KGQAs, similarly sized local models differ substantially in answer accuracy and path fidelity, with the two metrics sometimes favoring different models. This model-dependent behavior also extends to prompting, as a single demonstrated trajectory can improve or degrade navigation depending on the model. These results motivate evaluating SLM graph reasoning beyond endpoint accuracy alone.

★ How Much Were You Told? Measuring External Information in Peer Reviews

Matthieu Dubois, Pablo Piantanida, Fran\c{c}ois Yvon — Thu, 24 Sep 2026 00:00:00 +0000

95% · Evaluation 95%

Conference policies distinguish using Large Language Models (LLMs) to polish one's own review from delegating the critique, but current Artificial Text Detection (ATD) methods largely measure surface form rather than the origin of its content. We instead measure the external information carried by a review: information not explained by the reviewed paper and a generic reviewing instruction. We propose Self-Conditioning, an unsupervised information-theoretic estimator that compares the likelihood of a review under its production context with its likelihood when that context is augmented with hints extracted from the review itself. On the IntelLabs peer-review benchmark, Self-Conditioning separates fully-delegated from machine-polished reviews with AUC up to $1.0$ while remaining largely insensitive to surface rewriting. Moreover, as generators receive increasing amounts of externally-provided information, their scores move monotonically towards the human regime, unlike standard ATD baselines. High-temperature sampling can evade the estimator, but at the cost of output quality.

★ Exact Feedback Is Not Control: Evaluating Text-based Closed-Loop Revision in LLMs

Haitong Jiang, Chunlin Liu, Yile Wang, Yuhong Feng — Thu, 24 Sep 2026 00:00:00 +0000

95% · Evaluation 95%

Closed-loop revision is increasingly used in large language model (LLM) applications, but failures may reflect incomplete feedback or ineffective responses to correct feedback. We introduce a fixed-budget revision protocol with deterministic verifiers that report all remaining violations across exact-length, lexical, and compositional constraints. Fixing feedback correctness and completeness isolates model-side revision behavior. Across 19 open- and closed-source models, controller-level mean final joint success ranges from 17.4% to 99.8%, with substantial cross-model gaps persisting under identical initial drafts. Controlled experiments reveal reproducible model-specific responses to exact feedback. Post-training and scale reshape these responses without consistently bringing them closer to exact correction. Across all constraint families, failed trajectories often repeat earlier outputs, and prior recurrence is associated with lower subsequent recoverability. Matched-state interventions show that removing earlier dialogue while holding the current draft and feedback fixed changes recurrence escape without reliably improving final success; effects depend on the model, task, and trigger-state composition. Exact feedback makes revision errors observable, but does not make the closed loop reliable. Code and reproduction instructions: https://github.com/kevinjiang0121-cyber/exact-feedback-code.

★ Towards Efficient Reasoning: Learning Causal Shortcuts for Diffusion Language Models

Dian Jin, Kairong Han, Baohong Li, Xinpeng Dong, Zijing Hu, Nuanqiao Shan — Thu, 24 Sep 2026 00:00:00 +0000

95% · New model 95%, Efficiency 53%

Diffusion Language Models (DLMs) have attracted significant attention for their strong reasoning ability. However, under a bidirectional attention mechanism, DLMs operate over an exponentially large exploration space compared to autoregressive models (ARMs), making it challenging to focus on reasoning-guiding tokens under random masking. We define causal shortcuts as token chains that cover the full sequence and provide explicit guidance towards correct reasoning trajectories. We analyze the effects of causal shortcuts on the reasoning accuracy and convergence speed of DLMs, and find that they largely improve answer convergence efficiency and generation accuracy. Motivated by this, we propose a Causal Shortcut Learning (CSL) Framework for DLMs. Specifically, we introduce a step-by-step token extraction procedure to extract causal shortcuts from data, and apply parallel prioritized masking on these tokens during training to enable efficient and accurate convergence to correct answers via causal shortcuts. Extensive experiments across multiple reasoning benchmarks and two base models demonstrate that CSL consistently outperforms existing SFT-variant baselines, achieving an average improvement of $1.92\%$ over SFT-only models, and up to $4.20\%$ on MATH-500. The code is available at the \href{https://github.com/ZJUDianJin/Causal-Shortcuts-Learning}{https://github.com/ZJUDianJin/Causal-Shortcuts-Learning

Reinforcement Learning with Decomposed Subtasks

Mattie Terzolo, Mikolaj Sacha, Ayan Sinha, Andrew Rabinovich — Thu, 24 Sep 2026 00:00:00 +0000

94% · New model 94%, Agents 84%, Evaluation 68%

Group Relative Policy Optimization (GRPO) and related policy-gradient methods for training language model agents collapse an entire multi-turn rollout into a single scalar trajectory reward before it enters the policy update. When the task composes distinct skills, especially under sparse and delayed environmental feedback, this collapsing is lossy: the optimizer must implicitly infer which competency drove the outcome and how that should change behavior. We argue the right primitive is not a better scalar but a decomposition: trajectory reward should be split along subtasks before it enters the policy update. We introduce Reinforcement Learning with Decomposed Subtasks (RLDS), whose core is Subtask-Decomposed Advantage Estimation (SDAE): a replacement for the scalar GRPO advantage that splits trajectory reward into per-subtask shares on a fixed taxonomy, computes a group-relative advantage per subtask, and distributes per-token credit by weighting each subtask's advantage by its importance, concentrating it around the step where a reflection marks that subtask's execution as consequential. We evaluate on four agentic benchmarks: FrozenLake (sparse grid navigation), HotpotQA (multi-hop QA, one retrieval tool), ScienceWorld (long-horizon embodied science), and DeepResearch (long-form research, four tools, composite rubric reward). Heterogeneity diagnostics emitted during training show where decomposition pays off - gains scale with subtask heterogeneity, largest on the high-heterogeneity tasks ScienceWorld (+11.5 points, paired-bootstrap 95% CI [+9.8, +13.3]) and FrozenLake (+9.8 points, [+7.0, +12.8]), and within noise on HotpotQA and DeepResearch, where the diagnostics predicted little to recover. ScienceWorld is also more compute-efficient under RLDS than scalar GRPO (-10.9% wall-clock per step), as long rollouts amortize the fixed reflect-and-grade overhead.

DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents

Mingxuan Wang, Bo Wang, Fei Luo, Guorun Yao, Chao Ning, Yinglong Guo — Thu, 24 Sep 2026 00:00:00 +0000

94% · Efficiency 94%, Agents 88%, New model 69%

Long-horizon language-model agents accumulate reasoning traces, tool exchanges, and observations whose relevance changes with the current decision. Existing compression strategies often score historical units independently, but the safety of deleting several units is generally not determined by their singleton scores: redundant evidence, accumulated small effects, and the information that remains after deletion all matter. We introduce Direct Relational Set-Risk Pruning (DRSR), which formulates agent-history compression as risk-constrained selection over deletion sets. Offline, DRSR constructs exact counterfactual supervision by jointly deleting protocol-valid history Blocks and measuring the change in teacher-forced likelihood of the same recorded next output. A lightweight scorer then predicts set-level harm from online-visible relations between candidate history and the current pre-action state, together with deleted-retained and pairwise set structure. At deployment, DRSR evaluates a small set of structurally valid deletion candidates with the lightweight scorer and removes the largest feasible set under recency, protocol, budget, and learned-risk constraints, abstaining when no set is sufficiently safe. On WorkBuddyBench Full260, DRSR increases mean reward from 0.699 to 0.802 while reducing total model tokens by 20.820%. On the fixed Eval40 comparison, it obtains 0.794 reward at 1.211M tokens per task, using 35.850% fewer tokens than the uncompressed agent. Mechanistic analyses and ablations further show that decision-conditioned relations, retained-context information, pair interactions, and abstention each contribute to reliable pruning.

PotARCin: Multi-Dimensional Evaluation of Skill Acquisition in Abstract Reasoning Tasks

Claas Beger, Ryan Yi, Melanie Mitchell — Thu, 24 Sep 2026 00:00:00 +0000

94% · Evaluation 94%

The Abstraction and Reasoning Corpus (ARC) has become a prominent benchmark for evaluating general abstract reasoning and fluid intelligence in AI models. Yet standard ARC evaluation considers only a single capability: producing the correct output grid for a test input. We argue that this narrow format fails to evaluate the diversity of abilities that genuine abstract skill acquisition should enable. We introduce PotARCin, a benchmark that extends ARC by assessing understanding of a task's underlying abstract rule across five dimensions: Definition, Classification, Constrained Generation, Editing, and Inversion. PotARCin employs programmatic methods to generate new task instances and transform given inputs for a given ARC task, enabling dynamic generative sampling beyond fixed input-output pairs. Across five state-of-the-art models evaluated on the ARC-AGI-1 training set, we observe a 25-52 percentage-point performance gap between standard ARC evaluation and evaluation on PotARCin, and find that multi-dimensional evaluation reorders models that standard accuracy ranks alike. We further investigate effects of generative sampling, difficulty of corruption types, and questions of self-consistency, showing that models frequently contradict their own formalized rule even where they have stated it correctly. We also introduce P-ARC, a held-out hand-crafted test set, on which models achieve 1-8% accuracy across all five dimensions, underscoring the importance of more holistic evaluations of abstract reasoning capabilities.

KITE: KV-Invariant Transformer Expansion for Efficient Agentic LLM Scaling

Zhiheng Hu, Yixun Wei, Jian Zhou, Yizhuang Zhou, Ji Li, Xing Chen — Thu, 24 Sep 2026 00:00:00 +0000

94% · New model 94%, Efficiency 93%

Scaling a language model is not only a question of final quality: the architectural choice determines how much computation is spent during training, prompt processing, and autoregressive decoding to achieve certain model quality. An ideal model architecture should lower all above computation costs to facilitate scaling to a larger model, while ensure the larger model indeed outperforms smaller baselines. We introduce KV-Invariant Transformer Expansion (KITE), a scaling paradigm that achieves this goal. It trains the model from a smaller size to a larger size (i.e., saving training costs via upcycling), while places newly added parameters in regions that do not affect attention KV. Consequently, during inference, prefilling KV only relies on the smaller part of the model, so the inference costs are saved. As a concrete instantiation, we present Step Scale Transformer (SST), a two-tower decoder in which one tower produces KV and the other reads them. At comparable cumulative training compute, SST, a 67B MoE model with 2.15B active body parameters per decode token, achieves lower training loss than 47B and 63B MoE Transformers with 1.48B and 2.02B active body parameters, respectively, while reducing estimated inference cost by 6.7% and 31.6%.

Alignment Inertia: Auditing the Durability of Training Data Influence Through Policy Override Resistance

Renata Barreto, Markelle Roesti, Mohammad Tahaei — Thu, 24 Sep 2026 00:00:00 +0000

94% · Evaluation 94%

Platform operators increasingly rely on system prompts and fine-tuning to govern model behavior, yet it remains unclear how reliably these interventions override behavior inherited from prior training. We propose Override Success Rate (OSR) and alignment inertia to measure when operator interventions succeed or fail to change prior behavior. We evaluate zero-shot prompting and LoRA fine-tuning across Llama and Mistral in medical misinformation and hate speech. Alignment inertia persists across both models but varies by model, domain, and policy direction. Notably, in Mistral's restrictive hate-speech condition, LoRA increased inertia by 46.5 percentage points, showing that fine-tuning can reinforce rather than override prior behavior. We also use TRAK to test whether inertia is associated with weaker adaptation signals. TRAK achieves AUC of at least 0.85 in 7 of 8 conditions and outperforms model confidence, TF-IDF similarity, and embedding similarity as a predictor of inertia. These results provide an operator-facing audit of where prior training constrains downstream model governance.

AraGenre 2026: A Hierarchical Definition-Guided Arabic Genre Classification Shared Task

Thu, 24 Sep 2026 00:00:00 +0000

94% · Evaluation 94%

AraGenre is a shared task on hierarchical, definition-guided Arabic genre classification, motivated by the limited availability of annotated data in Arabic and other low-resource languages. Systems assign each Arabic text segment both a broad communicative genre and a fine-grained specific genre. The released training and development sets contain limited, primarily synthetic and controlled examples, whereas the hidden final benchmark contains noisier naturally occurring text spanning Modern Standard Arabic, Classical Arabic, and multiple dialects. Participants received natural-language definitions for 74 previously unseen specific genres, creating a zero-shot label generalisation setting in which systems had to infer class semantics rather than memorise fixed label-feature associations. The task attracted 46 registrations and 373 submissions, with 17 teams completing the final evaluation. Thakaa ranked first with a Hierarchical Macro F1 of 0.7352, followed by HoangPhong (HP) with 0.7169 and NAMAA with 0.7013. The results show strong broad-genre recognition but a substantial gap in fine-grained classification under linguistic and domain variation.

Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following

Yanzhao Zheng, Yuanqiang Yu, Tianze Xu, Chao Ma, Zhentao Zhang, Jihuai Zhu — Thu, 24 Sep 2026 00:00:00 +0000

94% · New model 94%, Evaluation 48%

Multi-constraint instruction following requires a model to respond to a query under many simultaneously active constraints. Even strong instruction-tuned models still routinely violate some of them. Existing approaches either augment supervision with sequence- or token-level RL rewards from external verifiers or learned graders, or use on-policy distillation (OPD) against a single full-context teacher whose probability mass becomes diluted as more constraints become simultaneously active. We propose CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation), which inverts the standard supervision-generation direction in distillation. Rather than enriching the teacher with information beyond what the student sees, CC-OPD ablates each constraint from the teacher's conditioning in turn, and constructs the per-constraint signal from the resulting per-token probability differentials. The resulting per-token leave-one-out log-likelihood shifts are summed, clipped, and added to the vanilla OPD reward as a token-level shaping term. All shaping terms are obtained from the frozen teacher, without an external verifier during distillation, and the reward equals vanilla OPD wherever the aggregate shift is zero. Across two Qwen model pairs and seven benchmarks, CC-OPD achieves the highest average among all evaluated student-training methods. A 1.5B student trained with CC-OPD surpasses its own 7B RL-trained teacher on the MulDimIF benchmark.

State-Grounded Conditioning: Wrapping User-Facing LLM Agents Where Direction Depends on Live State

Qi Liu, Xiaoyang Yuan, Yubin Ruan, Zhuomeng Zhang, Wenjin Wang, Di Wu — Thu, 24 Sep 2026 00:00:00 +0000

94% · Evaluation 94%, Hallucination 80%, Efficiency 79%

We introduce State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents that must condition on live user state (game state, session history, live inventory), and a distinct failure class we call direction drift: task-complete responses whose chosen direction misaligns with the current state. SGC externalises state-dependent control into rule kernels over structured inputs and three primary state slices, via Perception, Grounding, and Interaction wrappers with explicit conditioning dependencies. We evaluate SGC on a 200-session anonymised benchmark ($\approx$1,000 assistant model turns) from an in-game conversational coaching agent that guides players through consecutive competitive matches, reporting mean first-token latency and five human-annotated dialogue-quality metrics that jointly cover factual grounding and coach-like guidance progression. The Perception wrapper holds mean first-token latency at 1.5s (vs. 6.1s for PE-Agent inside a production tool-use harness); enabling all three wrappers lifts turn-level grounded accuracy from 61.1%/69.8% (Prompting / PE-Agent) to 96.7% and session-level grounded accuracy from 20.0%/26.5% to 83.5%; session-level grounding-failure incidents drop by $\approx$78% relative to the strongest baseline. A cumulative ablation shows complementary incremental gains as the wrappers are added. These results inform approximate state-slice orthogonality, without establishing independent per-wrapper effects.

SkillGym: Internalizing Human Skills into LLMs for Real-World Problem Solving

Zhilong Ge, Yuting Shao, Yutao Yang, Yuxuan Cai, Jie Zhou, Kai Chen — Thu, 24 Sep 2026 00:00:00 +0000

94% · Agents 94%, Evaluation 92%, New model 84%

Human-written agent skills encode rich workflows for real-world problem solving, but are typically used as external inference-time instructions rather than internalized as reusable model capabilities. We introduce \texttt{SkillGym}, a framework that transforms these skills into executable, verifiable training environments for large language model agents. Its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions. We construct and release 2,756 environments across 12 categories and collect 8,364 successful trajectories from multiple models and harnesses, averaging 49 tool calls and over 60k logged text tokens. These resources support supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards. Under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.1 with and without skills, respectively. Our 35B \texttt{SkillGym-Agent} reaches 51.47\% on skill-assisted SkillsBench, exceeding reported scores for Claude Sonnet 4.6, GPT-5.4 Mini, and DeepSeek V4 Pro. Without skills, it also surpasses skill-assisted bases under Codex and Claude Code, suggesting reusable procedural competence.

PCQC: Privileged Counterfactual Question Credit for Multi-Turn Medical Dialogue

Chenxuan Li, Jiayi Wan, Xinrong Chen, Zhongyu Zhao, Xuecheng Shang, Peixing Wan — Thu, 24 Sep 2026 00:00:00 +0000

94% · New model 94%, Evaluation 54%, Efficiency 50%

Large language models (LLMs) have made substantial progress on medical question-answering, yet effective medical dialogue also requires learning to ask questions that uncover relevant patient information. To train such dialogue policies, a common pipeline combines supervised fine-tuning with reinforcement learning (RL) based on final diagnostic correctness. However, this outcome-based supervision does not directly distinguish the contributions of individual questions and provides no question-level feedback for unexecuted alternatives. To address this gap, we introduce PCQC (Privileged Counterfactual Question Credit), which uses privileged patient information during training to learn from questions never asked. During training, PCQC makes alternative questions directly comparable at the same dialogue state by using privileged patient facts to construct their answers. A frozen diagnostic scorer evaluates the diagnostic utility of each resulting question-answer pair by how strongly it supports the correct diagnosis. PCQC turns these comparisons into relative question credit that teaches the policy which questions to favor, directly supervising both executed and unexecuted questions alongside outcome-based RL without requiring complete rollouts for the unexecuted alternatives. Extensive experiments across four medical benchmarks demonstrate that PCQC achieves 63.10% mean diagnostic accuracy, outperforming GRPO and ATPO by 4.38 and 4.21 percentage points, respectively. These gains are achieved with 33.1% fewer inquiry turns than GRPO.

Silent Failures in Agent-Tool Interaction: An Audit of ToolUniverse

Shreya Gopalan, Devansh Singh, Sundaraparipurnan Narayanan — Thu, 24 Sep 2026 00:00:00 +0000

93% · Agents 93%, Evaluation 63%

Agentic AI systems are increasingly adopting automated pipelines that integrate multiple tools. While prior research and benchmarks have studied about task success and task completion of these agentic systems, the research about agent to tool interaction, specifically in biology agentic workflow is limited. This study investigates specific failures in agent to tool interaction where a tool invocation appears successful, some or all of the information or functionality from the tool via API/ wrapper is incomplete or missing and there are no communications / notifications to the user or the agent about such missing information. We call this a silent failures as the user or the agents are not aware that such failure has occurred. For the purposes of this study we developed an audit mechanism to identify such silent failures in Agent to tool interaction, by examining 15 scientific tools (and their associated API documentation and tool documentations) integrated within ToolUniverse environment (ToolUniverse serves as our experimental environment rather than the object of the study itself). We structure our study around 7 failure locus characterising where the failure occurs in the chain. We observed 91 failures (manually validated post LLM based candidate discovery and automated testing), most frequent of them being missing data or fields and inconsistencies in search, filtering or ranking criteria. Most of the 91 failures occurred in API layer (51) or wrapper layer (25), with a potential of silent failure amplification downstream. The results show that silent failures originate upstream of the event and propagate downstream into apparently valid scientific outputs. We propose a concept of contextual reliability to handle such failures and suggest mechanisms for testing, disclosing, monitoring, and measuring such failures across the agent-tool interaction pipeline.

ChipMEM: Verification-Grounded Memory for EDA Agents

Thu, 24 Sep 2026 00:00:00 +0000

93% · Agents 93%, Evaluation 49%, New model 45%

Large language model (LLM)-based agents use Electronic Design Automation (EDA) tools to generate and revise register-transfer-level (RTL) designs under synthesis and verification feedback. Recent methods learn from this feedback by distilling reusable skills from execution traces or by training on rewards derived from EDA-tools. Both methods are typically evaluated on the tasks that produced the experience. Repeated access to benchmark feedback on the same task can reward task-specific revision rather than creating reusable knowledge that transfers. We introduce ChipMEM, a verification-grounded memory layer for EDA agents. It combines cross-task procedural memory with within-trajectory statistical guidance. Its procedural component distills and stores a skill only after it passes synthesis, simulation, or formal checks, rather than relying on model self-assessments. A Bayesian component maintains hierarchical Beta estimates over tool-call outcomes and ranks recovery strategies that succeeded under comparable errors. A common adapter applies the same memory interface to RTL optimization and testbench-generation agents while preserving each domain's tools and acceptance criteria. We measure performance on training tasks and evaluate whether learned skills transfer to unseen tasks. On RTLRewriter-Bench, under matched model and tool settings, ChipMEM produces equivalence-passing outputs on 39/54 scored designs versus 35/54 without memory; on the 49-design short suite, mean area improvement is 8.69% versus 5.66%. On held-out CVDP tasks, ChipMEM with a frozen procedural library achieves 20/20 accepted outcomes versus 18/20 without memory in a single evaluation per setting.

Distilling Sequential Computation in Transformer Language Models

Zixuan Lan, Jessica Yang, Yanhong Li, Karen Livescu, Jiawei Zhou — Thu, 24 Sep 2026 00:00:00 +0000

93% · Efficiency 93%

Transformer language models process sequences token by token in an autoregressive manner, making growing contexts increasingly expensive. Yet many adjacent token spans are highly predictable or frequently occur as stable units, suggesting that their representations may be compressible. We introduce a method for distilling sequential computation by replacing spans of input tokens with collapsed representations, computed on the fly by a lightweight merge module. This module generates a single surrogate embedding from a sequence of static token embeddings that captures the functional role of the multiple tokens, allowing pretrained models to operate on compressed inputs without architectural changes or re-training. We apply this approach during inference to compress both prompts and intermediate decoding steps, using a rollback mechanism to substitute stored multi-token KV cache entries with their single-step surrogates. Experiments across diverse models show that the merge module can be used to reduce effective sequence length by up to 40% with minimal accuracy degradation across language modeling evaluations and downstream tasks, including question answering, summarization, commonsense reasoning, and long-form mathematical reasoning. Additional lightweight adaptation of the merge module further improves the accuracy-compression trade-off in selected settings. These results demonstrate that sequential token computation in Transformers can be effectively approximated through condensed surrogate representations that approximate the original behavior without model updating.

Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression

Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning — Thu, 24 Sep 2026 00:00:00 +0000

93% · Efficiency 93%, Agents 51%, Evaluation 43%

Long horizon agents accumulate growing interaction histories that increase context and inference costs. We find that geometric redundancy alone is an insufficient criterion for safe compression. Although agent histories exhibit strong low dimensional structure, similar global geometry can preserve very different amounts of task evidence. At identical retained block counts, evidence aware selection raises next action Top 3 retention from 0.31 to 0.69, while centroid similarity remains 0.98. Controlled replacement further shows that action related information can be substantially altered while global geometric measures remain nearly unchanged. Motivated by this gap between geometry and evidence, we introduce Geometry Guided Evidence Preserving Memory (GEM), a training free compressor that protects task and execution evidence before using geometric residuals to complete coverage. GEM reduces mean combined token usage from 2.69M to 2.11M per task, a 21.4% reduction, while maintaining comparable task reward. Our results show that efficient agent history compression should optimize for preserved task evidence rather than geometric coverage alone.

DCRL: Decoupling and Coupling Reinforcement Learning via Policy-Reward Manifold Alignment

Henan Sun, Zehua Li, Haitao Hu, Qifan Zhang, Jianfeng Zhang, Nuo Chen — Thu, 24 Sep 2026 00:00:00 +0000

93% · New model 93%

Reinforcement learning (RL) has emerged as a key paradigm for improving the reasoning capabilities of large language models (LLMs). However, existing reward systems, such as rule-based and reward-model-based, often exhibit issues such as unstable optimization and reward hacking. In this work, we revisit the general reasoning of LLMs from a geometric perspective, conceptualizing it as a coupled manifold composed of three interdependent sub-manifolds: logical deduction, evaluation, and representation. Based on this perspective, response generation in RL can be interpreted as a decoupling process from the evaluation manifold, while reward estimation corresponds to a decoupling process from the logical deduction manifold. The limitations of rule-based and reward-model RL systems can be geometrically interpreted as the mismatch of policy-reward manifolds during RL process. To address the aforementioned misalignment, we propose Decoupling and Coupling Reinforcement Learning (DCRL) framework, which incorporates two key components: (1) a syllogistic logic-based prompt evolution mechanism that dynamically refines reward rubrics to enhance the expressiveness of the reward manifold; and (2) a policy-reward re-coupling mechanism that jointly updates the reward and policy models, ensuring consistent evaluation and mitigating manifold mismatch during training. Theoretical analysis and extensive experiments across multiple reasoning domains demonstrate that DCRL consistently outperforms both rule-based and reward-model baselines. Notably, a Qwen3-4B model trained under DCRL surpasses a Qwen3-32B baseline and approaches the performance of a Qwen3-235B model, highlighting superior effectiveness and generalization in RL.

Beyond Unsafe Detection: Counterfactually Anchored Evidence Attribution for Multi-Turn LLM Safety Failures

Srinivasan Subramanian, Kazi Aminul Islam, Md. Abdullah Al Hafiz Khan — Thu, 24 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, New model 51%

As Large Language Models (LLMs) move from conversational assistants to advanced agentic systems, guardrail failures can convert adversarial intents into harmful executions. However, most guardrail evaluation frameworks focus only on the result and assess whether a user request is safe or unsafe. This approach is insufficient for multi-turn failures, where adversarial intent is distributed across multiple turns. This motivates us to go beyond detection to identify the turns and tokens that push the conversation toward unsafe trajectories. To support this, we construct a multi-turn dataset with behavioral validation and tiered evidence supervision. The dataset contains 1,762 conversations, including adversarial conversations, benign twins, and benign variants with high-risk vocabulary. We train a lightweight hierarchical attribution model that predicts safety violations and attributes them to contributing user turns and token spans. The model achieves strong detection performance (F1=0.988), and removing the top 15% of attributed tokens reduces the adversarial classification confidence by 51.1%. The model preserves low false positive rates on benign conversations with high-risk vocabulary, with false positives below 1% on both borderline benign and benign high-risk vocabulary conversations, compared to 37.3% and 94.7% for a keyword-based surface-risk baseline. Independent human annotation supports the model's attribution performance, with the top-five attributed turns containing a human-identified evidence-bearing turn in 84.5% of adversarial cases.

Controlled Attribute-Specific Summarization of Interrogative Dialogues

A Aditya Bhardwaj, Arjit Singh Arora, Md Shad Akhtar — Thu, 24 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, Hallucination 73%, New model 42%

Effective summarization of interrogative dialogues is a critical task in forensic and investigative settings, requiring high factual accuracy, coherence, and attribute-specific relevance. In this work, we introduce CASPER, a novel Chain-of-Thought Attribute-Specific Prompting for Evaluative Summarization framework that leverages structured prompting and iterative refinement to generate high-quality summaries of interrogator-witness interactions. We construct MINDSum, a dataset extending the MIND corpus, comprising 6,000 utterance pairs annotated with event details, factual statements, character descriptions, and fillers. CASPER employs RoleEval, a hierarchical evaluation mechanism where multiple roles (officer, inspector, senior inspector) iteratively assess summaries based on predefined criteria. By integrating entity extraction and structured feedback loops, CASPER significantly improves factual consistency and contextual completeness compared to existing baselines. Experimental results demonstrate that our framework outperforms standard summarization models on both lexical (ROUGE) and semantic (BERTScore) metrics, while human evaluation confirms its alignment with expert reasoning. Our findings underscore the potential of controlled summarization in high-stakes domains, paving the way for AI-driven forensic intelligence.

When and Where to Trust the Teacher: Unifying On-Policy Distillation and GRPO through Entropy-Calibrated Credit Assignment

Jie Zhang, Jingxiao Yang, Zhehao Huang, Yuhang Liu, Xiaolin Huang — Thu, 24 Sep 2026 00:00:00 +0000

93% · New model 93%

Reinforcement learning with verifiable rewards (RLVR) supervises mathematical reasoning through final-answer correctness, but provides little guidance on individual tokens. On-policy distillation (OPD) supplies dense feedback on student-generated responses, yet teacher preference need not reflect correctness. Recent hybrids combine OPD and verifier-derived advantages or reweight task credit using teacher ratios. However, teacher guidance enters after verifier-based group normalization, and token reweighting need not preserve the total task credit assigned to each response. We introduce Unified Entropy-Calibrated Credit Redistribution for GRPO (UECR-GRPO), which integrates verifier and teacher signals within a single GRPO-style update at both the response and token levels. \emph{Path-Utility Unification} (PUU) combines verifier reward and a teacher-to-anchor path log-ratio in a single KL-regularized objective. Its on-policy implementation uses a length-normalized teacher score and combines both rewards before group normalization and PPO clipping, allowing teacher evidence to influence the response ranking. \emph{Entropy-Calibrated Redistribution} (ECR) then uses the signed teacher--old-policy token gap to redistribute the verifier-derived component. Full-vocabulary teacher entropy attenuates uncertain guidance, while a response-wise zero-sum projection preserves the total task credit and its token-wise sign before clipping. Across five mathematical reasoning benchmarks, UECR-GRPO achieves average $\mathrm{Avg@12}$ accuracies of 17.21\% and 65.09\% with Qwen3-1.7B and Qwen3-4B students, respectively, exceeding the strongest baseline at each scale by 0.89 and 0.56 percentage points.

Agent-Editing World Model: Rethinking World Modeling for LLM Agents

Shuang Sun, Guoxin Chen, Fanzhe Meng, Jia Deng, Huatong Song, Jinhao Jiang — Thu, 24 Sep 2026 00:00:00 +0000

93% · Agents 93%, New model 91%, Evaluation 90%

Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from \emph{task-state contamination}, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the \textbf{Agent-Editing World Model (AEWM)}, which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines \textbf{Action Judge} to distinguish \textsc{Critical}, \textsc{Exploratory}, and \textsc{Noisy} decisions with \textbf{State Revision} to edit noisy reasoning--action continuations from the same observed history. \textbf{EditAct} integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5\% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed \textbf{AEWM-RFT}, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.

Giving Credit Where It's Due: Redundancy-Aware Learning for Efficient Reasoning

Yuqing Zhou, Hong Wang, Manqing Mao, Zhuoer Wang, Samson Koelle, Jie Yuan — Thu, 24 Sep 2026 00:00:00 +0000

92% · New model 92%, Efficiency 90%, Evaluation 37%

Large reasoning models can produce correct yet unnecessarily long reasoning traces. Existing methods improve reasoning efficiency with trajectory-level objectives or local token- and step-level signals, but rarely model inter-step semantic dependencies. This limits their ability to distinguish redundant steps from those that support later deductions, making it harder to shorten reasoning without sacrificing accuracy. We introduce RECAP (REdundancy-aware Credit Assignment via Propagation), which addresses this limitation by assigning credit where it is due based on both a step's downstream role in the reasoning structure and its contribution to solving the problem correctly. We define structural responsibility to capture the step's downstream role by measuring how strongly later reasoning depends on it, using credit propagated backward from the final-answer node through an outcome-independent, LLM-annotated semantic dependency graph. However, a step can have high structural responsibility yet steer the reasoning away from the correct solution. RECAP therefore introduces step efficacy to measure answer-directed progress through changes in gold-answer log-likelihood as each step is added. Together, these signals reshape rollout-level GRPO advantages into step-specific updates. RECAP requires neither a separately trained process reward model nor preconstructed concise trajectories. Across two 7B models and four mathematical reasoning benchmarks, RECAP improves the accuracy-efficiency trade-off. On Qwen2.5-Math-7B, it improves pass@1 by 2.0-3.7 percentage points while reducing reasoning tokens by 8%-31% relative to GRPO across all four benchmarks. Analysis suggests these savings reflect fewer reasoning operations and less dead-end reasoning, rather than more compact expression.

Learn How to Act from Your Own Interactions: On-Policy Self-Distillation for GUI Agents

Yan Zhang, Daiqing Wu, Huawen Shen, Liang Li, Gang Cao, Zhi Gong — Thu, 24 Sep 2026 00:00:00 +0000

92% · New model 92%, Agents 55%

Graphical User Interface (GUI) agents enable the fulfillment of complex user instructions through multi-turn interactions with software environments, requiring step-wise reasoning and long-horizon memory to guide actions and retain task-relevant information, respectively. Recent on-policy self-distillation (OPSD) methods have achieved strong performance on GUI grounding, a foundational subtask for GUI agents, owing to dense token-level supervision from privilege-conditioned self-teachers. However, extending existing OPSD methods to multi-turn GUI agents is hindered by self-teachers' limited privilege-following ability and insufficient privileged guidance. In this paper, we introduce GUI-SD-v2, the next version of GUI-SD, which extends OPSD from GUI grounding to multi-turn GUI interaction and addresses key limitations through a two-stage training framework. Specifically, GUI-SD-v2 first strengthens privilege following by jointly optimizing rollouts with and without privileged guidance from the same GUI states. Furthermore, it selectively distills step-specific reasoning and memory guidance through a privilege-conditioned self-teacher, supporting action decisions and the retention of task-relevant information for subsequent interactions. Extensive experiments on two representative GUI agent benchmarks, AndroidWorld and MobileWorld, show that GUI-SD-v2 compares favorably with existing OPSD baselines while consistently outperforming the evaluated state-of-the-art methods in both Pass@1 and Pass@3 success rates. Code and training data will be publicly released.

LabourCrew: A Multi-Agent RAG Framework for Trustworthy Adversarial Deliberation and Statutory Reasoning over Labour Law

Thu, 24 Sep 2026 00:00:00 +0000

92% · Evaluation 92%, Hallucination 92%, Agents 44%

In statutory question answering, every claim must be traceable to evidence, not merely relevant, since unverifiable labour-rights answers carry serious legal consequences. Current systems fall short: single-pass RAG cannot detect insufficient evidence, while multi-agent legal-debate systems treat grounding as a prompting convention, letting agents cite unretrieved evidence. To address this gap, we introduce LabourCrew, a multi-agent RAG framework built around three grounding mechanisms: StatuteGraph, a graph index that explicitly links chapter, section, proviso, and cross-reference structure rather than fixed-length spans; an Evidence Exchange Protocol that confines advocates and an interpreter to an evidence ledger, making citation to unretrieved text impossible, while a fault-tolerant supervisor board runs advocates in parallel so individual failures degrade rather than crash the system; and a Calibrated Trust Gate that replaces categorical accept/reject decisions with a trust score, thresholded via conformal risk control for a distribution-free bound on the false-accept rate. We evaluate on LabourActQA, a 500-item Bangla question set from the Bangladesh Labour Act, 2006, spanning seven reasoning categories and three difficulty tiers. The framework drives the empirical false-accept rate to 0.081, within the target level ($\alpha = 0.10$), achieves the highest Answer Relevancy among HyDE RAG, Graph-RAG, and Hierarchical RAG (0.862 $>$ 0.839, 0.815, 0.828), and degrades gradually rather than catastrophically as question difficulty increases. These results show that calibrated abstention, not retrieval quality alone, is what makes legal question answering auditable in low-resource statutory domains.

Beyond Poetry: Can Large Language Models Generate Classical Arabic Maqamat?

Thu, 24 Sep 2026 00:00:00 +0000

92% · Evaluation 92%

Large language models (LLMs) have shown strong performance in creative text generation, yet their ability to produce culturally grounded and stylistically constrained literary forms remains underexplored. Prior work has focused largely on modern language varieties and poetry, while classical prose traditions such as maqama remain largely unstudied. The maqama is a classical literary genre characterized by rhymed prose (saj), dense rhetorical ornamentation, and episodic narrative structure, making it a challenging testbed for evaluating whether LLMs can move beyond surface fluency toward deeper literary competence. In this paper, we present the first controlled evaluation study of maqama generation with LLMs, comparing five models under zero-shot, few-shot, and rule-based prompting, and evaluating outputs through both human annotation and an LLM-as-a-judge framework across dimensions such as rhetorical richness, saj density, structural coherence, and stylistic authenticity. Our results show that prompting strategy plays a strong role in stylistic quality: few-shot prompting most consistently improves saj density, while its effects on rhetoric and coherence vary by model, with the strongest models (GPT-4o and GPT-5.4-mini) benefiting most from rule-based prompting on these dimensions, though zero-shot prompting yields the highest aggregate scores across all five models. We further observe systematic differences between models in stylistic alignment with Arabic maqama conventions, and corroborate our findings with a second independent LLM judge, paired statistical significance testing, and non-LLM proxy measures of saj.

Math Reasoning in LLMs is Organized by Approach, Not Topic

Sajad Goudarzi, Samaneh Zamanifard, Moloud Nasiri, Hamed Rahimian — Thu, 24 Sep 2026 00:00:00 +0000

91% · Evaluation 91%

Mathematical reasoning benchmarks are typically organized by topic, but language models may organize their internal computation by reusable reasoning approach instead. In this paper, we investigate whether open math-capable LLMs organize internally by topical sub-skill or by reasoning approach, and we present evidence that the approach is the key. We introduce a generation-replay protocol: a model first generates a solution, after which we replay the exact prompt-plus-generation trajectory and extract activation-importance signatures over the reasoning tokens. We cluster these signatures without supervision across eight models and five mathematical reasoning sources, then evaluate the recovered structure with structural, semantic, and intervention tests. Across all 40 model-source cells, the recovered clusters outperform matched-size random baselines. Two independent frontier-LLM judges find approach-level coherence in 77-82% of real clusters versus 6-11% in within-source controls, and topic-pure clusters usually receive labels finer than the topic itself. In approach-controlled prompting, changing the requested reasoning approach shifts cluster assignment in seven of eight model conditions, whereas paraphrases largely preserve it. These results indicate that math-capable LLMs organize internal mathematical computation by reasoning approach rather than benchmark topic. The implication is that topic-stratified benchmarks and topic-balanced training corpora can still miss the axis that matters: even deliberately topic-balanced corpora may remain imbalanced over reasoning approaches.

NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task

Thu, 24 Sep 2026 00:00:00 +0000

91% · Evaluation 91%

NADI 2026 is the seventh edition of the Nuanced Arabic Dialect Identification (NADI) shared task series and the second dedicated to multidialectal Arabic speech processing. This edition comprises five tasks and eight subtasks spanning Automatic Speech Recognition (ASR), Spoken Dialect Identification (SDID), Text-to-Speech (TTS), Spoken Language Translation (SLT), and Spoken Language Understanding (SLU). NADI 2026 emphasizes realistic evaluation through low-bandwidth, mixed-dialect, code-switched, out-of-domain, and zero-shot settings, while introducing TTS, SLT, and SLU to the series for the first time. The shared task attracted 21 participating teams from at least 13 countries, with 48 test-phase submissions and 14 submitted system-description papers. Results show that out-of-domain generalization remains a major bottleneck and highlight the effectiveness of recent Arabic-specialized speech models, multimodal dialect identification approaches, and ensemble methods. Overall, NADI 2026 provides a broader and more challenging benchmark for robust Arabic dialect speech processing.

FLEET: From Logits Entropy to Enhanced Trajectories in Text Generation

Oleksii Streltsov, Oleksandra Vitko — Thu, 24 Sep 2026 00:00:00 +0000

91% · Efficiency 91%, Evaluation 48%

Solutions based on large language models (LLMs) often rely on temperature sampling to improve accuracy and stability by aggregating multiple samples from the completion distribution. However, this memoryless approach is inherently suboptimal: because it lacks awareness of prior generations and their evaluations, it produces an increasing proportion of semantically duplicate answers as more samples are drawn, leading to diminishing returns. To address this limitation, we introduce FLEET, a novel method that integrates a memory mechanism into the generation process. FLEET represents each generation as a sparse trajectory through states whose entropy exceeds a predefined threshold and uses these trajectories to infer per-token utility scores that adjust the logits. Benchmark evaluations demonstrate that FLEET achieves the same accuracy as the repeated sampling baseline, with a 3x speedup, and substantially improves accuracy on complex coding tasks (LiveCodeBench Pass@32 increases from 59.9% to 66.2%) under the same budget. Furthermore, in the greedy-decoding configuration evaluated here, the approach is deterministic and uses a single calibration pass to derive its principal hyperparameters, requiring only minimal modifications to existing LLM pipelines.

Same Scores, Different Decisions: Evaluating JEV and Language Models for Legal Document Understanding

Fan Zhang, Yankai Chen, Zhuohan Xie, Yixi Zhou, Sijia Peng, Lei Fan — Thu, 24 Sep 2026 00:00:00 +0000

91% · Evaluation 91%

Contract inference requires multiple judgments about a shared document, but aggregate accuracy can conceal changes in the individual decisions. Repeated agreement is also insufficient: a model may consistently return the wrong answer. In this paper, we compare Jev with nine language models on ContractNLI, evaluating inference cost, response time, average correctness, and correctness across repeated request conditions. Controlled comparisons vary hypothesis visibility, requested outputs, and output order while keeping the contract and target judgment fixed. Jev has the lowest cost and median response time among the evaluated configurations, while hosted language models achieve higher baseline accuracy. Rankings by baseline accuracy differ from rankings by correctness across every condition and repeat, although small differences in the latter do not establish a general stability advantage. Development diagnostics further reveal compensating corrections and regressions, as well as persistent errors. These findings motivate evaluating cost and response time alongside whether individual judgments remain correct as the request configuration changes. Code: https://github.com/ZF-Utokyo/Jev-Benchmark

Log-Depth Recurrent Language Modeling

Yiqin Wang, Nuri Cingillioglu, Charles Pert — Thu, 24 Sep 2026 00:00:00 +0000

91% · New model 91%, Efficiency 78%

Language modeling using Transformers has become commonplace despite their fixed computational depth and quadratic runtime with respect to input tokens. Recurrent models on the other hand offer linear depth but no parallel execution. In this work, we extend balanced-tree recursive operators from sequence encoding to autoregressive prediction, enabling all prefix representations to be computed with logarithmic depth and linear runtime. Our experiments provide an initial characterization of this model class, demonstrating robust length extrapolation and performance approaching that of ALiBi-based Transformers, highlighting its potential as an alternative architecture for language modeling.

Discover, Falsify, Revise: Auditing Input-Use Claims from Source Code to Predictive Contribution in Agent-Discovered Cell Models

Mengran Li, Bo Li, Chengyang Zhang, Yang Yan, Jinfeng Xu, Zhenchao Tang — Thu, 24 Sep 2026 00:00:00 +0000

90% · Evaluation 90%

AI virtual cells aim to predict cellular responses to specified interventions, yet held-out predictive performance alone does not establish use of the supplied perturbation information. This prediction-claim gap matters in agentic model discovery, where language-model agents generate and revise predictors using score-based feedback. We introduce CELLAUDIT, which audits input-use claims by asking whether an input can enter the cited computation, whether fitted predictions depend on it, and whether that dependence improves prediction of observed response. On a paired morphology-transcriptomics perturbation benchmark (BBBC047), an agent-selected predictor attains a mean held-out Global Pearson correlation coefficient (PCC) of 0.3153 but remains invariant to compound replacement; a control-profile-only predictor reaches 0.3142. Source inspection identifies a compound-query pathway blocked by singleton key-value attention, and the invariance persists after refitting with disjoint control wells. In a stratified audit of 48 candidates across two linked tasks, 47 change predictions under compound replacement on both held-out folds, but only 20 show target-loss gains with intervals above zero on both folds. On BBBC047, falsification-guided revisions recover positive mean compound contributions while retaining gains over the control-profile-only baseline. In matched sci-Plex searches, audit-enriched feedback yields higher held-out performance and larger mean compound and dose contributions across five trajectories, although paired intervals span zero. Refitting fixed designs on an independently acquired cohort shows predictive generalization need not imply generalization of input-use claims: dose contribution persists, whereas support for compound identity does not. CELLAUDIT adds a falsification layer to agentic model discovery, moving from generate-score-revise toward discover-falsify-revise.

From Reasoning Strings to Partial Orders: Verifier-Certified Rule Transport through Quotient Policy Optimization

Bang Xie, Hao Liu, Zhiyuan Peng, Xin Yin, Chenhao Ying, Yuan Luo — Thu, 24 Sep 2026 00:00:00 +0000

90% · New model 90%, Evaluation 77%

Many computations admit several valid execution orders because independent subgoals or disjoint state updates can commute. Reinforcement learning with verifiable rewards usually treats each successful trace as a separate token sequence, so serialization choices can be mistaken for logical dependencies. We introduce Verifier-Certified Rule Transport (VCRT), which replays adjacent operation pairs with native verifiers. Pairs whose two orders are accepted and reach the same canonical state provide commutation certificates; rejected or state-changing reversals provide anti-diamonds. VCRT uses anti-diamonds to preserve genuine prerequisites and assigns policy credit to the total probability mass of each certified orbit. It also constrains post-swap consistency, source retention, and policy drift. We evaluate leave-one-environment-out transfer across ProofWriter, CLRS, and Lean through a shared anonymized relation-graph interface. All training and checkpoint decisions are frozen before held-out evaluation, which uses one greedy trajectory per item without search or verifier feedback. VCRT obtains a 77.60% macro pass rate versus 64.53% for the strongest matched baseline, a paired gain of 13.06 points (95% bootstrap CI [12.58, 13.54]). Lean accounts for most of this gain at 33.49 points, while ProofWriter and CLRS improve by 2.85 points on average. Mechanism tests consistently favor anti-diamond supervision, whereas No-Orbit is statistically indistinguishable from full VCRT. The evidence does not establish a general benefit from exact orbit aggregation.

Exact Quantile Balancing and Load-Error Injection for Mixture-of-Experts

Pit Neitemeier, Jiaze Li, Alessio Serra, Philipp Scholl, Sohir Maskey — Thu, 24 Sep 2026 00:00:00 +0000

90% · New model 90%

Mixture-of-Experts (MoE) training requires global load balance to prevent expert under-utilization and local balance for efficient expert-parallel execution. Existing distributed Quantile Balancing (QB) uses shard-dependent or approximate global quantiles, while token-independent expert biases cannot ensure microbatch-level balance. We introduce Exact Quantile Balancing (EQB), which computes exact global-batch BF16 quantiles with negligible communication, and Load-Error Injection (LEI), which injects local load errors directly into router-score gradients. On 7.5B-parameter MoEs trained for up to 500B tokens, EQB improves global balance and downstream performance over naive QB, while LEI improves local balance and outperforms the GShard loss at comparable quality.

Reference-Based Analysis of Coherence and Diversity in Open-Ended Text Generation

Esteban Garc\'es Arias — Thu, 24 Sep 2026 00:00:00 +0000

90% · Evaluation 90%

Evaluating open-ended text generation involves understanding how different properties of a continuation relate to its perceived quality. We present a reference-based framework for examining coherence and diversity through three perspectives: aligning their evolution with human trajectories, comparing their summaries with a human continuation of the same prompt, and estimating their likelihood under a human reference distribution. Experiments with human quality ratings suggest that diversity-based alignment and mean-based comparisons capture quality-related variation, although the comparisons do not establish a predictive advantage for temporal alignment over simpler baselines. Reference likelihood also shows positive associations with ratings, with results varying across reference configurations and scoring horizons. Together, these analyses provide a structured way to examine how measured coherence and diversity relate to human judgments, while distinguishing similarity to human references from quality itself. Code and analysis resources are available at https://github.com/EstebanGarces/likely_human.

★ FinFIRST: Benchmarking Search Agents for Financial Information Retrieval, Sourcing and Traceability

Wenqing Wang, Haitao Xiang, Xinyi Zhao, Mingming Yin, Ying Zhong, Zhaoxin Huan — Wed, 23 Sep 2026 00:00:00 +0000

98% · Evaluation 98%, Agents 81%, reliability 51%

Financial search is a highly demanding task for LLM agents, requiring not only a correct final answer but also temporally valid information retrieval, authoritative source selection, entity and period alignment, unit and definition consistency, and verifiable evidence for all conclusions. Existing benchmarks predominantly evaluate only the final answer, making it difficult to localize errors or assess whether an answer is well-founded. To address this gap, we introduce FinFIRST (Financial Information Retrieval, Sourcing and Traceability), the first financial benchmark to jointly evaluate answers and supporting evidence through atomic rubrics. FinFIRST comprises 123 expert-authored tasks spanning a graduated difficulty spectrum, constructed from aggregate patterns of real-world financial scenarios through an 18-field taxonomy, a six-axis coverage blueprint, a registry of 138 financial sources, contributions from over 50 finance experts, and a six-stage quality-control pipeline. Each task is accompanied by an evidence-grounded reference package decomposed into atomic criteria across three dimensions: raw-information acquisition, source verification, and computation and answer formation. We evaluate 15 model configurations under a unified tool setting. Claude-Opus-5 achieves the highest atomic score of 87.59%, while GPT-5.6-Sol attains the highest strict pass rate of 71.54%. Computation and answer formation consistently lag behind raw-information acquisition across systems. FinFIRST retains final-answer correctness as the primary objective while making the supporting research process measurable, verifiable, and diagnosable.

★ Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

Quan Nguyen-Tri, Mukul Ranjan, Zhiqiang Shen — Wed, 23 Sep 2026 00:00:00 +0000

98% · Efficiency 98%, capability 85%

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value (KV) caching and scalable parallel decoding mechanisms. Existing acceleration methods typically study KV caching and parallel decoding in isolation, overlooking the I/O bottlenecks that arise when cache reuse and parallel token verification are jointly applied. In this work, we introduce $\textbf{Flash-dLLM}$, a training-free inference acceleration framework for fast and memory-efficient dLLMs. Flash-dLLM first identifies GPU memory I/O as a dominant bottleneck in KV-cache-enabled dLLM inference and addresses it with an I/O-aware fused KV-cache kernel that reduces redundant memory movement. Building on this optimized cache mechanism, Flash-dLLM further proposes an efficient KV-cache-driven draft-and-verify decoding strategy, where the dLLM itself serves as both drafter and verifier without requiring an auxiliary model. This unified design enables faster decoding while preserving generation quality and improving scalability to longer sequences and larger batch size. Extensive experiments on mathematical reasoning and code-generation benchmarks demonstrate that Flash-dLLM consistently outperforms existing state-of-the-art dLLM acceleration methods in both inference speed and memory efficiency. In particular, it achieves $5.1\times$ and $11.0\times$ speedups over prior strongest baseline Elastic-Cache on GSM8K and HumanEval, respectively.

★ ShowTellArena: Evaluating Business Workflow Understanding from Demonstrations

David Garg, Ritobrata Sarkar, Ehsan Azarnasab, Siddhartha Borah — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Agents 73%, reliability 37%

We often teach a colleague by showing the work and explaining the decisions as we go. How can we check what an agent understood from the same lesson? We introduce ShowTellArena, a benchmark protocol and public dataset for comprehension after narrated business demonstrations. The v1.0 release contains 50 business workflow tasks, with recordings, screenshots, narration, fixture seeds, and 502 questions. Tasks span finance, hiring, procurement, customer decisions, inventory, and logistics. The protocol holds the business scenario and quiz fixed while allowing each product to capture the lesson through its own teaching interface. Questions test operational rules, boundaries, exceptions, and errors in proposed automations. We analyze 218 selected pilot attempts across 39 workflow cases, including 28 cases attempted by all three evaluated systems. These exploratory results expose both answer errors and failures to complete the teaching experience. We describe the release's verification gaps and the pilot's uneven coverage, exclusions, and grading provenance. The contribution is an inspectable dataset and assessment workflow that others can extend; the selected pilot is not a controlled product ranking.

★ Compressing Long Context into Answer-Aligned Memory Embeddings for LLM Inference

Wed, 23 Sep 2026 00:00:00 +0000

97% · Efficiency 97%, capability 83%

Large language model (LLM) inference is constrained by the quadratic scaling of self-attention and the linear scaling of the KV cache, increasing latency, energy consumption, and GPU memory demand as context length scales. Existing soft-compression methods either lack query-guided memory selection at inference time, train without answer-targeted supervision, or couple compression tightly to a specific decoder architecture. We propose a Context-to-Answer-Aligned Memory Compression (CMC) framework, which compresses long input contexts into compact Context Memory Embeddings (CMEs) aligned to any frozen decoder's embedding space, reducing inference costs without modifying decoder weights. CMC introduces a two-tier KV cache that combines question-guided CME selection with a local context window, and trains the compressor with answer-targeted distillation from a frozen LLM. Experiments across nine encoder-decoder combinations and four QA benchmarks show that CMC consistently outperforms the baseline, achieving up to 7.3 EM and 4.0 F1 point gains on SQuAD, while reducing inference time and energy consumption by up to 20% and peak reserved GPU memory by up to 50% at 3,000 generation tokens. Ablation studies confirm that each architectural component and training objective contributes to the performance.

★ The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Wenbo Pan, Zhichao Liu, Shujie Liu, Jingying Zeng, Chin-Yew Lin, Xianfeng Tang — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Agents 89%, capability 30%

LLM agents increasingly work on long-horizon tasks, and the decisions they make along the way, such as which hypothesis to test or which implementation to build on, determine the outcome of the whole run. Making these decisions well is becoming a key capability for both engineering and research agents. We refer to the ability to make good long-horizon decisions as the taste of an agent. While existing benchmarks measure the end-to-end success of agents on long-horizon tasks, none of them measures the taste of an agent. To address this problem, we build Taste-Bench, a benchmark of taste questions constructed automatically from trajectories that agents produced in engineering and research tasks. Each question presents a decision fork, a point in a trajectory where multiple directions are available and one of them leads to a better outcome, and the evaluated model chooses among these directions without seeing what happens after the fork. We mine these forks automatically from parallel attempts at the same task and from detours inside a single trajectory, without needing human annotation. We evaluate frontier models on Taste-Bench and find that the best model answers only 59.7% of the questions correctly. We further find that forks whose deciding evidence appears later in the trajectory are much harder for every model, and that a larger reasoning budget does not improve the accuracy. Finally, we show that taste can be trained. We distill the judgment of a teacher that has seen the outcome into a student model, and the student makes better decisions on unseen tasks and improves end-to-end success on held-out SWE-bench Pro tasks.

★ CQ4OE: A benchmark for assessing LLM-assisted ontology generation from competency questions

Jiayi Li, Ziyuan Wang, Daniel Garijo, Mar\'ia Poveda-Villal\'on — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Agents 36%

Ontology generation from Competency Questions (CQs) is a central yet labor-intensive phase of Ontology Engineering. While large language models (LLMs) offer promising automation capabilities, current evaluations remain fragmented. Task formulations are heterogeneous, gold standards often lack fine-grained CQ provenance, metrics conflate lexical overlap with structural and logical adequacy, and reference ontologies are not always explicitly designed around the evaluation CQs. Here, we address these limitations with CQ4OE, a benchmark for the systematic and reproducible evaluation of LLM-based ontology generation from CQs. For each ontology in the benchmark, we build a CQ-driven gold OWL ontology with explicit provenance linking each CQ to the classes, properties, and axioms required to answer it. From this resource, we define two complementary evaluation tasks. CQ2Term supports term-level evaluation of CQ-specific class and property prediction over 99 CQs, and CQ2Onto supports ontology-level evaluation over 118 CQs, including hierarchy, property modeling, and axiom-level structure. We demonstrate CQ4OE with experiments using nine LLMs under zero-shot, iterative, and multi-agent generation strategies, showing that LLMs recover explicit vocabulary terms more reliably than creating ontologies, particularly in property modeling, hierarchy construction, and axiom generation.

★ SpecialEduBench: Benchmarking Vision-Language Models on Knowledge, Skill, and Attitude in Language Intervention for Autistic Children

Jihoi Na, Taeyeong Kim, Sungjune Kong, Jaemin Jung, Min Joung Park, Kyungtae Joo — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, reliability 65%

Language is the target of most early intervention for autistic children. Because the goal and the method change from child to child, the work falls to a teacher who takes one child at a time and judges each scene as it unfolds. Artificial intelligence is now being brought to that work, yet the benchmarks that reach special education ask what a model knows rather than what it does in front of a child. Building one is not straightforward, since whether a response is good teaching depends on what the child has just done, so no answer key applies. The evidence that settles it is visual as much as verbal, since the length of a wait, a shift of gaze, and the child's uptake leave no trace in a transcript. We introduce \emph{SpecialEduBench}, which measures pedagogical competence along knowledge, skill, and attitude, with 4,537 knowledge items and with 200 skill items and 68 attitude items built on recorded intervention, the attitude items crossing pressure with monitoring into 192 response cells. Seven special-education experts wrote, scored, and reviewed the items, and we revised the judge model's instruction against the reference scores they set. Across eight frontier vision-language models no axis is saturated, since the strongest still fails about a tenth of the honesty cells. The models converge where the knowledge is factual and separate where the task is situated, and the failures gather where pressure is applied. We intend the benchmark as an audit to run before deployment and as a starting point for models built for this domain.

★ TSS: Target-Side Sparsification for Speculative Decoding in Domain-Specific Large Language Models

Haibo Hu, Lianming Huang, Qiao Li, Nan Guan, Chun Jason Xue — Wed, 23 Sep 2026 00:00:00 +0000

97% · Efficiency 97%, capability 72%

Speculative decoding accelerates large language model inference through collaboration between a lightweight draft model and a target verifier. Existing methods mainly improve the draft side, while the target model is typically kept dense and unchanged. We show that, under domain-specific inference, full-depth target verification is not always the optimal choice. Counter-intuitively, skipping selected target layers can reduce verification cost while simultaneously increasing draft acceptance and preserving, or even improving, downstream task performance. Based on this observation, we propose TSS, a target-side sparsification framework for speculative decoding. TSS employs an acceptance- and metric-aware breadth search to explore multi-layer skip configurations without imposing a fixed priority between the two objectives. The selected configurations are stored in a domain-to-configuration mapping and applied by a lightweight skip controller, allowing one complete target model to support multiple sparse verification paths without retraining or permanent parameter pruning. Experiments on Spec-Bench across multiple domains, model scales, and speculative decoding methods show consistent improvements in draft acceptance and downstream task performance. In Translation setting, TSS increases the average accept length from 2.70 to 4.53 (+67.8%), improves BLEU from 0.131 to 0.237 (+80.9%), and raises end-to-end throughput from 75.6 to 127.3 tokens/s, corresponding to a 1.68X speedup.

★ EADC: Evaluation of Advanced and Deep-level Compliance in Large Language Models

Yan Zhang, Ruien Li, Yaoyao Peng, Wanxin Ren, Yijia Zhang, Wusheng Zhang — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, reliability 77%, Agents 46%

Large Language Models (LLMs) have been used in various industries. However, ensuring their compliance with complex laws and regulatory frameworks remains a great challenge. Existing evaluation paradigms mainly rely on static benchmarks that suffer from three severe limitations: First, the compliance rules being used do not comply with the requirements of Artificial Intelligence (AI) laws and regulations; Second, they only handle apparent, explicit compliance risks, leaving implicit and covert compliance risks undetected; Third, they fail to track the systematic propagation of risks along logical dependency chains or evaluate compliance within nuanced, context-based real-world scenarios. To bridge this critical gap, we introduce EADC, a novel advanced evaluation benchmark of LLMs based on an AI compliance knowledge graph and AI compliance legal experts. By mapping abstract legal rules into structured logical multi-relational graphs, our framework enables automated, evolving agents to distill and synthesize highly sophisticated adversarial scenarios. This compliance benchmark is reviewed and corrected by human AI legal experts throughout the whole process. The resulting dataset (4,435+ QA pairs) provides an extensive, multi-dimensional taxonomy covering critical regulatory frontiers, including bias and discrimination, fairness, personal privacy protection, and values. Crucially, our compliance dataset moves beyond shallow string-matching by incorporating contextual long-horizon interactions and logic-driven hazard chains, capturing deeply embedded compliance anomalies that bypass traditional filters. Experiment evaluations demonstrate that our framework exposes critical regulatory blind spots in state-of-the-art LLMs, offering a rigorous, AI laws and regulations-aligned benchmark to safeguard high-level and deep compliance in the application of LLMs.

★ Same Chart, Different Story: Bias in Vision-Language Chart Interpretation

Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, reliability 86%

Vision-language models (VLMs) are increasingly used to interpret charts and generate natural-language explanations for socially consequential data. However, they may produce different narratives for the same chart when only the referenced social group changes, reinforcing stereotypes and misleading decisions. Despite these risks, no benchmark exists for systematically evaluating bias in chart interpretation across social dimensions. We introduce ChartBias, the first benchmark for auditing bias in VLM-based chart interpretation. ChartBias contains 820 manually curated real-world charts spanning six attributes: race, income, age, religion, immigration status, and gender, yielding 4,319 valid chart, attribute instances and 8,638 paired generations where the chart is fixed and only the group term is swapped. Across 12 proprietary and open-source VLMs, totaling 155,484 model responses, we find three widespread failure modes: narrative shift (same chart, different narratives), group hallucination (assigning a chart to a group without evidence), and preference polarity (favourable trends often linked to one group). We further propose a multi-agent mitigation framework that serves as a strong baseline by separating chart-grounded evidence extraction from group-conditioned generation and using a counterfactual judge to verify that group-driven differences are supported by the chart. The framework substantially reduces narrative shift while preserving chart-grounded reasoning. Our findings show that evaluating chart understanding requires measuring not only accuracy, but also fairness and consistency across social groups. We release ChartBias at https://github.com/vis-nlp/ChartBiasBench.

★ CompKV: Compensation-Aware KV Selection for Long-Context LLM Inference

Zhen Huang, Ruizhe Yao, Danyi Liu, Xinrui Chen, Shuwei Li, Siru Zhong — Wed, 23 Sep 2026 00:00:00 +0000

97% · Efficiency 97%, capability 72%

Despite their strong performance, large language models (LLMs) are bottlenecked by KV cache memory traffic during long-context inference. Sparse attention is widely used to accelerate LLM inference by computing exact attention over a selected subset of tokens. To recover the contribution of tokens excluded from exact attention, recent methods apply coarse-grained compensation to the omitted attention tail. However, existing methods typically select tokens based on attention mass and only then compensate for the unselected tokens. This decoupled design overlooks their interaction: selection should prioritize tokens that would leave the largest compensation error if omitted. To address this limitation, we introduce CompKV, the first compensation-aware sparse attention framework that divides tokens into blocks and explicitly optimizes selection for the downstream compensation mechanism. Our theoretical analysis shows that the residual left by block-level mean compensation is governed by both block attention mass and within-block logit variation. We approximate this residual using compact block-level statistics, yielding a deployable selection criterion. We further develop an efficient asynchronous implementation. Experiments on RULER and LongBench-Pro show that CompKV performs best among the evaluated sparse baselines while delivering up to a $6.85\times$ self-attention speedup over full attention.

★ SWE-Serve: Benchmarking Agentic Engineering For Production Inference Serving

Jennifer Williams, Dave Farris, Jeff Farris, Jiantao Jiao — Wed, 23 Sep 2026 00:00:00 +0000

97% · Evaluation 97%, Agents 84%, Efficiency 47%

We introduce SWE-Serve, a benchmark for evaluating agents on production inference engineering tasks. Implementing an inference feature can require coordinating multiple changes across the serving stack, including model support, runtime execution, and public APIs. Existing benchmarks provide limited coverage of production inference engineering: repository-level software engineering benchmarks do not target inference, while general terminal-agent benchmarks include only a few inference tasks. Dedicated inference benchmarks, meanwhile, focus primarily on isolated kernel generation or performance optimization rather than repository-scale production feature implementation. SWE-Serve provides 53 repository-grounded tasks derived from recent production changes to SGLang, spanning six inference engineering families. Each task executes on either CPU or a single GPU (H100) and is evaluated with hidden functional and regression tests, including, where applicable, end-to-end (E2E) serving tests and calibrated performance gates. Executable no-op and oracle controls, adversarial verifier review, and closed-book execution support task validity and evaluation integrity. Across 11 models and 31 model-effort configurations, the best-performing configuration achieves 75% mean pass@1. SWE-Serve exposes a substantial gap between completing tasks locally and achieving production correctness. On 19 tasks with end-to-end coverage, model-serving E2E tests reject roughly one-third of patches that pass every other test (45.9% under the verifier versus 69.4% with E2E tests excluded from scoring), with pass rate increasing for each model's best-performing configuration. By making the production correctness gap directly measurable, SWE-Serve enables the field to track whether future agents move beyond completing tasks locally to achieving production correctness.

★ Peerify: Benchmarking Peer-Review Claim Verification

Wed, 23 Sep 2026 00:00:00 +0000

96% · Evaluation 96%, reliability 46%

Peer review plays a central role in scholarly publishing, yet verifying whether reviewer claims are supported by manuscript evidence remains a largely manual and time-consuming process. We present Peerify, a pipeline for manuscript-grounded verification of peer-review claims. Given a manuscript and a review comment, the Peerify pipeline decomposes reviews into atomic claims, retrieves relevant manuscript evidence, and determines whether each claim is supported by the paper. To support the development and evaluation of the pipeline, we construct a benchmark of 800 claims derived from authentic peer-review interactions collected from NeurIPS 2024 and ICLR 2024, including a 300-claim hand-labeled subset used to audit the automated supervision. We evaluate state-of-the-art language models and retrieval strategies within the Peerify pipeline, together with entailment baselines. Our results demonstrate the importance of retrieval-centered verification and claim decomposition, while highlighting the challenges posed by ambiguous and interpretive reviewer claims. Automated labels agree with human consensus on 90.3% of audited claims ($\kappa = 0.87$), while off-the-shelf entailment models stay below 0.24 macro-F1.

★ Trains but Doesn't Learn: A Post-Training Delivery Benchmark for LLM Agents as Forward-Deployed Engineers

Weihang Ding, Junfei Zhan — Wed, 23 Sep 2026 00:00:00 +0000

96% · Evaluation 96%, Agents 83%, reliability 68%

Post-training is becoming a service (PTaaS): a customer hands an operator data and a goal, and a forward-deployed engineer (FDE) returns a fine-tuned, evaluated, and deployed model under a budget, a human-approval gate, and reproducibility requirements. Seating an LLM agent in the FDE seat raises a question existing benchmarks cannot answer: not whether an agent can raise a metric, but whether it can be trusted to deliver. We answer it on a governed delivery plane, where an agent drives ten stages and an oracle scores each stage from platform-recorded facts. The central silent failure is the run that trains but does not learn (TBDL): loss falls, every signal stays green, and the delivered model is no better than the base. An operator-run acceptance gate catches every such run before payment, and a detector calibrated on known-corrupted runs flags severe corruption mid-run. We ran four frontier agents (Claude Opus 5, GPT-5.6-luna, Gemini 3.7 Flash, DeepSeek V4-Pro) end to end on metered L40S, A100, and H200 GPUs across 8B to 70B open bases, certifying every scenario before scoring. We also ran a human FDE arm under the same oracle and compare every agent against it.

★ Efficient Iterative Retrieval with Heterogeneous Batching

Wed, 23 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, capability 67%

Modern information retrieval increasingly employs both embedding and generative models to handle complex queries. However, current serving systems suffer from low throughput and poor GPU utilization because they execute these models in isolation. Coarse-grained partitioning, such as dedicating GPUs to specific tasks, fails to adapt to dynamic workloads and creates computational "bubbles". To address these, we present Orthrus, a serving system that performs heterogeneous batching within a unified inference loop. The primary challenge lies in unifying embedding and generation workloads with conflicting computational patterns while optimizing batch composition for high performance. Orthrus addresses these challenges through chunked embedding with incremental pooling and by adjusting batch composition in a workload-aware manner. Evaluation on four A100 GPUs shows that, relative to baseline deployments, Orthrus achieves 1.28$\times$--4.52$\times$ higher throughput on controlled workloads and up to 55.8% lower end-to-end p99 latency on an iterative-RAG benchmark. We release our code at https://github.com/illinoisdata/Orthrus .

★ Evaluating Coding Agents on Kernel Exploit Generation

Junyoung Jang, Gwanhyun Lee, Hwiwon Lee, Kyuheon Kim, Jongseong Kim, Jinho Jung — Wed, 23 Sep 2026 00:00:00 +0000

96% · Evaluation 96%, Agents 84%, reliability 48%

Coding agents now find real vulnerabilities in production software. However, bug discovery results do not measure whether agents can construct exploit primitives. We introduce KEX-bench, a benchmark for evaluating coding agents on exploit primitive generation against real operating-system kernels. KEX-bench contains 45 task instances across 40 Linux and Windows CVEs, covering kernel address leak, instruction-pointer control, heap read, heap write, and arbitrary address write. Each task runs in an isolated virtual machine, exposes controlled tools, and uses a deterministic verifier to check primitive-specific success. We evaluate state-of-the-art coding agents paired with frontier and open-weight models under fixed tool-call budgets. Without a reference proof of concept (PoC), the strongest configuration solves 1 of 20 Windows tasks (5.0%) and 14 of 25 Linux tasks (56.0%). With a reference PoC, the strongest configuration solves 31 of 45 tasks (68.9%). This highlights the gap where agents reach kernel crashes but fail to shape kernel state into exploit primitives. We release KEX-bench for reproducible research on AI-assisted exploitation at https://kex-bench.github.io.

★ You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs

Yuanteng Chen, Qiwei Lai, Chen Tianqi, Peisong Wang, Yuantian Shao, Nanxin Zeng — Wed, 23 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, Evaluation 83%

Fine-grained mixture-of-experts (MoE) architectures have become a mainstream design for open-weight LLMs, with hundreds of experts and increasingly many selected per token. This shift makes dynamic expert pruning an attractive route to cheaper inference. Yet existing evidence comes largely from coarser architectures and likelihood-scored multiple-choice benchmarks, leaving three central questions open in the fine-grained regime: how redundant per-token expert selection is, how effectively existing pruning methods exploit that redundancy, and what governs a model's sensitivity to pruning. We fill this gap with a systematic empirical study of twelve fine-grained MoE checkpoints spanning nine architecture families, with a core suite of eleven benchmarks covering knowledge QA, mathematics, code generation, and general reasoning. We find that expert selection is far more redundant than the field's operating points assume: uniformly retaining about two thirds of the selected experts preserves 98.8% of unpruned performance on average, requiring only a one-integer change and delivering 1.2-1.7x measured speedup across two serving backends. This simple baseline leaves little room for dynamic allocation at conservative budgets: even the best published rules differ from it by under 1% at matched expert budgets. Their value emerges under aggressive pruning, where the best rules recover up to 3.0% over uniform truncation, with gains concentrated in the generative tasks that suffer the sharpest degradation. Sensitivity to aggressive pruning also depends on the model: larger and thinking models are more resilient, whereas multimodal models are more vulnerable. Together, these findings reveal how much expert computation fine-grained MoEs can dispense with, and establish when dynamic allocation earns its complexity, informing both practical deployment and future pruning methods.

★ Magnitude Profile Pruning: Calibration-Free Structured Attention Head Removal for Transformer Compression

Kasun Dewage, Marianna Pensky, Heranga K. Rathnasekara, Suranadi De Silva — Wed, 23 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, capability 72%

Structured pruning of attention heads provides a hardware-friendly way to compress Transformer language models. However, existing methods for measuring head-level importance require calibration data, gradient computation, or Hessian estimation. These requirements add extra overhead and make the methods depend on the data. Our work presents Magnitude Profile (MP) scoring, a training-free criterion for head importance that identifies dispensable heads through statistical outlier detection on weight row norms. Heads whose projection weights fall within the population bulk are pruned, while heads exhibiting outlier norms, which carry disproportionate representational capacity, are preserved. Our work further gives MP-G, a variant that handles Grouped Query Attention (GQA) by distributing shared key-value group scores across associated query heads. Across five models evaluated on WikiText-2 perplexity at 12.5%-50% head sparsity, MP-G achieves the best perplexity on OPT-6.7B at all sparsity levels (18.46 at 12.5%, 27.87 at 25%, 152.0 at 50%). MP-G also gives the best results on RoBERTa-large at 12.5% and 25% sparsity, with perplexity values of 7.27 and 10.28, outperforming calibration-dependent baselines including Wanda-Head, SparseGPT-Head, and Gradient-Head. It requires zero forward passes, calibration samples, or gradient computation. At 50% sparsity, head pruning yields up to 16% parameter reduction with 50% attention FLOP savings. Our results show that weight-only statistical scoring can match or outperform data-dependent methods for structured head pruning, providing a practical, zero-cost criterion for Transformer compression.

★ Beyond Static Charts: Can Language and Vision Language Models Generate Interactive Data Visualization Interfaces?

Mizanur Rahman, Aaryaman Kartha, Enamul Hoque Prince — Wed, 23 Sep 2026 00:00:00 +0000

96% · Evaluation 96%, Agents 44%, capability 38%

Data visualization is central to analytical reasoning, but real-world analysis increasingly requires language-driven interactive interfaces rather than static charts. Although recent large language and vision language models (LLMs/VLMs) have shown promise in generating static charts from natural language, their ability to generate interactive data visualization interfaces remains largely unexplored due to the lack of benchmarks. We introduce VIS-GEN, a benchmark for evaluating how well LLMs/VLMs can generate interactive visualization interfaces from natural language queries. VIS-GEN comprises 3,042 samples covering diverse analytical intents, including data filtering, temporal analysis, and visualization editing, each paired with dataset metadata and natural language queries that are designed to reflect realistic, goal driven data exploration scenarios. We benchmark 14 state-of-the-art open-source and closed-source LLMs/VLMs, revealing large performance gaps and frequent failures on queries involving implicit intent, multiple interaction alternatives, and complex editing operations, highlighting interactive interface generation as a key open challenge beyond static chart synthesis. To address this, we propose a structured multi stage interface generation framework that decomposes the task into visualization design representation, generation of multiple interface candidates, constraint-aware critique, and self-refinement. This approach improves the best models pass rate by 15.9 percentage points, demonstrating a practical path toward more reliable language-driven interactive visualization systems. We release VIS-GEN at https://github.com/vis-nlp/VIS-GEN.

★ PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models

Xiaocheng Lu, Shuhan Guo, Ziyue Ma, Jie Zhang, Jian Liu, Jingcai Guo — Wed, 23 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, capability 63%, reliability 37%

Diffusion language models (dLLMs), such as LLaDA and Dream, have become competitive with autoregressive (AR) LLMs in generation quality while supporting native parallel decoding. A standard acceleration strategy is block-wise decoding, where each forward pass predicts a block of length B and commits high-confidence tokens. However, B couples two distinct decisions: the look-ahead horizon and the number of tokens to commit. Existing accelerators address this limitation through indirect heuristics, such as volatility tracking, delimiter detection, and learned scoring. In contrast, we show that the required information is already encoded in the model's own per-step confidence: in-window confidence typically follows a context-dependent cliff, whose saturation point directly identifies the appropriate look-ahead horizon. We propose PACE-dLLM, which fits this parametric cliff in closed form at each step, sets the next horizon by its saturation point, and uses an independent confidence threshold for token commitment. Under a saturated-yield abstraction, we show that the cliff-anchored horizon is the smallest horizon attaining maximal useful per-pass yield: fixed horizons that undershoot it incur a worse asymptotic NFE rate, while overshooting adds no useful yield. On four reasoning and code benchmarks, PACE-dLLM achieves the best average accuracy on both open-source dLLM backbones, with average wall-clock speedups of 5.23x on LLaDA and 3.06x on Dream (up to 8.52x on math) over the unaccelerated semi-AR baseline, advancing the quality-throughput Pareto frontier.

★ Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding

Dohyun Kim, Sungjun Han, Hyungguk Kim, Yusik Kim, Jamin Shin, Paul Hongsuck Seo — Wed, 23 Sep 2026 00:00:00 +0000

96% · Efficiency 96%, reliability 37%

Autoregressive OCR vision-language models accurately convert document images into text and structured markup, but require one sequential decoding step per output token, limiting inference speed. Unlike open-ended text generation, OCR outputs are strongly grounded in the input image, making diffusion-based parallel generation promising. However, when several tokens are predicted in one diffusion step, each is predicted before the others are known. Committing them directly can therefore introduce errors. We therefore introduce GravityOCR, a parameter-shared AR-block-diffusion model jointly trained for parallel drafting and causal AR verification. Verifying drafts before commitment lets the model commit multiple output tokens per round without a separate drafting network. The causal AR path also enables GRPO with sequence- and structure-level OCR rewards, avoiding diffusion-trajectory likelihood estimation while updating the shared drafter parameters. On OmniDocBench v1.6, AR-path GRPO improves the Overall score from 94.92 to 95.16 without reducing diffusion drafting efficiency, while the final model remains close to the original GLM-OCR score of 95.48. In an SGLang serving deployment, GravityOCR commits an average of 9.7 output tokens per forward pass and achieves a $3.94\times$ decode-only speedup on region crops and a $1.32\times$ end-to-end page-processing speedup over AR decoding.

★ Same Quantity, Different Answer: Numerical Representation Invariance in Language Models

Ephraim Atta-Duncan — Wed, 23 Sep 2026 00:00:00 +0000

95% · Evaluation 95%, reliability 82%

Numerically equivalent word problems should yield the same canonical answer whether a quantity is written as a decimal, fraction, percentage, number word, scientific notation, or an exactly converted unit. We generate 3,600 exact-rational problems and 8,600 prompts spanning five identity-preserving transformation families, and evaluate five open-weight systems. After a fixed syntax audit that normalizes common answer forms without an LLM judge, canonical accuracy is 0.969-0.996, but orbit correctness falls to 0.848-0.981 and orbit invariance to 0.851-0.981; invariant-but-wrong orbits account for at most 0.003. Most of the broad strict-parser collapse arises because multiplication-form scientific notation lies outside the implemented number grammar, illustrating how evaluator interfaces can masquerade as reasoning failures. A distinct semantic pathology remains: Mistral Small 4 scores 0.699 on unit-converted inputs and produces 265 errors differing from the label by exact powers of ten. In a separate 9,000-call experiment that allocates equal calls to the compared arms, representation consensus does not outperform paraphrase consensus on a low-error subset and produces substantially more false alarms. The accompanying ancillary archive contains the frozen benchmark, evaluation and audit records, consensus raw responses, manifests, analysis code, and a one-command paper build.

★ FrontierMath Erd\H{o}s

Tom Adamczewski (Epoch AI), Thomas F. Bloom (University of Manchester) — Wed, 23 Sep 2026 00:00:00 +0000

95% · Evaluation 95%, Agents 68%, reliability 31%

We introduce FrontierMath Erd\H{o}s (FME), a benchmark of 68 Erd\H{o}s problems that are open as of August 2026. To solve a task in FME, AI systems must resolve (prove or disprove) one of the 68 conjectures in the proof assistant Lean. Our 68 problems were selected by the second author among 652 open problems on erdosproblems.com for their mathematical interest and difficulty. AIs have recently resolved several open problems in mathematics, but these demonstrations fall short of a systematic study of AI capabilities. FME evaluates every AI model on the same fixed problems, autonomously and under the same budget. We evaluated five AIs with a budget of \$300 per problem. One (GPT-6 Astra) scored 3%, and all others scored 0%.

★ Rollout Efficiency in Reinforcement Learning for Reasoning Large Language Models: A Taxonomy and Future Directions

Wed, 23 Sep 2026 00:00:00 +0000

95% · Efficiency 95%, Agents 62%

Reasoning-oriented reinforcement learning enables large language models to solve mathematical, coding, and other multi-step tasks, but shifts a substantial portion of the training cost to rollout, where trajectories are generated for policy updates. Efficient rollout mechanisms are therefore essential to reduce this cost while maintaining the freshness, consistency, and statistical validity of training data. This survey provides a systematic taxonomy of recent research on rollout efficiency for reasoning-oriented reinforcement learning, classifying existing approaches from both mechanism and bottleneck perspectives. Based on this taxonomy, we analyze how different technique families address distinct sources of rollout inefficiency, examine opportunities and potential conflicts for combining them, identify gaps in the evaluation and reporting of efficiency gains, and discuss open challenges and future research directions.

★ GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression

Baher Mohammad, Ammar Ali, Stamatios Lefkimmiatis — Wed, 23 Sep 2026 00:00:00 +0000

95% · Efficiency 95%, capability 67%

Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better preserves each layer's distinct calibration geometry. Coupled with structured sparsity, this yields highly efficient weight decompositions without sacrificing functional fidelity. Across diverse architectures, scales, and modalities, our method achieves state-of-the-art results, consistently outperforming independent structured weight decompositions and alternative pairwise weight factorizations, which operate under heuristic grouping strategies. By replacing heuristic engineering strategies with a convergent, optimization-driven pipeline, we establish a theoretically grounded foundation for scalable, transformer compression across different modalities.

★ Disaggregated Quantization: Specializing LLM Prefill and Decode

Wed, 23 Sep 2026 00:00:00 +0000

95% · Efficiency 95%, capability 75%

Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce memory traffic during generation. We propose "disaggregated quantization" (DQ), which specializes computation formats, weights and storage placement to both of these phases. On Qwen 3 and Gemma 3, removing activation quantization specifically on decode improves accuracy on decode-heavy tasks without increasing inference cost. Training separate compute-native prefill weights accelerates prompt processing relative to weight-only inference while matching or exceeding its accuracy at 2-3-bit decode on both decode-heavy and prefill-heavy tasks. With released Qwen3.8-27B GGUF decoders, training an NVFP4 prefiller improves 1-bit accuracy by 32.5 points on MMLU-Pro and 35.3 on MMMU-Pro without modifying the decode checkpoint. To accommodate the additional checkpoint on a single device, offloaded disaggregated prefill (ODP) streams its weights from SSD, amortizing loading over prompt length. On the same 27B model, ODP delivers a 1.78x time-to-first-token speedup over the weight-only baseline at 8K prompt length in llama.cpp. We evaluate accuracy under disaggregated serving in vLLM and further validate shared-weight format disaggregation through post-training quantization on models up to 2.8T parameters.

★ HySparse2: Hybrid Sparse Attention with Two-Level KV Sharing

Jianyu Wei, Yizhao Gao, Qihao Zhang, Shimao Chen, Zhengju Tang, Yu Cheng — Wed, 23 Sep 2026 00:00:00 +0000

95% · Efficiency 95%, Agents 82%, capability 58%

Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context demands efficient prefill, compact KV-cache storage, and accurate long-context retrieval. To meet these demands, we introduce HySparse2, a hybrid sparse attention architecture with two-level KV sharing. At the outer level, KV Bridging adopts a YOCO-style self-decoder and cross-decoder structure, but bridges only full-attention layers. The self-decoder uses hybrid sliding-window attention (SWA), while the cross-decoder uses hybrid sparse attention. The KV caches for full-attention layers in the cross-decoder are generated from the hidden states of full-attention layers in the self-decoder. At the inner level, HySparse2 retains HySparse's core KV Reuse design with two refinements. First, it replaces block-level sparsity with token-level sparsity for finer long-context retrieval. Second, it removes the separate SWA branch from sparse layers and instead forces a sliding window of recent tokens into the sparse selection. This two-level KV sharing allows all cross-decoder KV caches to be constructed from self-decoder hidden states. Prefill can therefore exit after the self-decoder, skipping all cross-decoder layers. On an 80B-A3B MoE model, HySparse2 outperforms HySparse and Hybrid SWA on long-context retrieval and multi-turn agentic tasks, while substantially reducing prefill computation and KV-cache storage.

★ CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents

Trang Nguyen, Eulrang Cho, Bingqing Chen, Tim Dettmers — Wed, 23 Sep 2026 00:00:00 +0000

95% · Efficiency 95%, Agents 89%, capability 77%

Agents often work on complex problems that require millions of tokens of context, which necessitates compacting across sessions due to limited context windows. We develop CliffCompaction, an autocompaction technique that reduces cost by up to 50% under a bounded context while maintaining or improving performance on Terminal-Bench and achieving new levels of efficiency for test-time scaling and state-of-the-art results on KernelBench. The per-rollout savings of CliffCompaction make the performance--cost trade-off of test-time scaling more efficient, adding over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs. Under parallel test-time scaling, CliffCompaction lets Kimi K2.6 match Opus 4.7, and exceed Opus 4.6 and GPT-5.3 Codex at lower cost. The key to CliffCompaction's effectiveness is that it keeps compacted information faithful by only truncating or dropping content, never rephrasing or rewriting it. We never compact a compaction---each pass operates only on original content, and prior compacted output is discarded, preventing context drift from accumulating. These properties sustain continual learning over sessions exceeding a million tokens: on KernelBench, CliffCompaction reaches CUDA kernel speedups of $2.23\times$ after 200 steps and $3.58\times$ after 400 steps, surpassing specialized search algorithms and trained agents despite being a general-purpose compaction technique. We open-source a scaffold-agnostic API-proxy implementation of CliffCompaction usable with Claude Code, Codex and other harnesses.

★ Passes Alone, Fails Together: Benchmarking Semantic Coordination in Parallel LLM-Agent Development

Haocheng Xia, Eugene Wu, Yongjoo Park — Wed, 23 Sep 2026 00:00:00 +0000

94% · Evaluation 94%, Agents 66%, reliability 47%

Parallel coding agents can produce patches that work alone but fail when merged. This happens when one agent changes an interface or rule that another agent still relies on. We study these failures with stale, a benchmark for semantic coordination. Our evaluation runs the same tests on each patch alone and on their combination, counting only failures introduced by combining the patches. We use three tiers: synthetic tasks with controlled interface changes, pairs of merged pull requests, and constructed tasks that use real Django helpers. Among 834 runs on 417 mined Django pairs, only one showed interference after correcting the grading procedure. On constructed tasks using 12 Django helpers, interference occurred in 97% of runs. A message describing the completed concurrent change recovered 82% of runs. Reviewed pull requests may contain few unresolved parallel changes, even when agents fail on controlled tasks using real code. The constructed failure rates do not estimate how often these problems occur in practice.

★ Conduct Under Pressure: What Sixty Language Models Do When a User Pushes

Tapan Parikh — Wed, 23 Sep 2026 00:00:00 +0000

94% · Evaluation 94%, reliability 88%

We study what LLMs do when a user applies pressure in an uncomfortable situation: a user insists, begs, flatters or grieves, and the model gives up a correct fact, writes a document it should refuse, or cheers a plan that will cost the user money. We send frozen multi-turn scenes, identical for every model regardless of the reply, to 60 models from 13 vendors, and label each transcript with a codebook built by open coding and then frozen: a trajectory (the model held its position or folded) and a manner (how it held or folded). Two findings separate. Whether a model holds tracks its generation, meaning how recent it is: fold rate correlates with a public capability index at Spearman -0.64, with little vendor effect. How it holds tracks the vendor: six of the 17 manner codes sort by vendor at permutation p <= 0.001, corrected across the codebook. We report four vendor profiles on the codes that cleared reliability. We also ask which parts of the labeling need a person. Six LLM coders from three vendors apply the codebook more consistently than three human coders do (Krippendorff's alpha 0.66 against 0.46), agree with the codebook's author on trajectory at kappa 0.84 to 0.91 on transcripts the codebook's examples never touched, and match an adjudicated human reference at 0.83. Blind machine readings recover the codebook's categories but cannot tell which of them a second reader would apply the same way. We conclude that for behavior a non-specialist can judge, the human contribution is authoring and bounding the codes and owning a small reference, not producing labels at volume.

★ CoVeR: Coverage-Based Routing of Verifier Calls in Agentic Retrieval

Daeyoung Roh, Donghee Han — Wed, 23 Sep 2026 00:00:00 +0000

94% · Efficiency 94%, Agents 76%, reliability 52%

An agentic retrieval system issues a sequence of search queries and must decide, at each step, whether the evidence collected so far is enough to stop. Delegating that decision to an LLM verifier or a prompt judge makes stopping reliable, but the verifier then reprocesses the growing evidence after every retrieval step, a substantial repeated cost. We show that most of these calls can be skipped without materially changing answer accuracy: a single threshold on a frozen sentence-embedding coverage margin detects the states in which the evidence is still plainly incomplete, and the verifier is called only on the ambiguous remainder, a gate we call CoVeR (Coverage-based Verifier Routing). Across three multi-hop QA benchmarks, with the evaluation protocol fixed before the full-scale run, the CoVeR-gated agent matches the answer accuracy of both the full-budget agent and the always-verify baseline within a fraction of an EM point. It cuts 62-68% of verifier calls, and 93% in a saturated regime. Routers built on evidence counts, lexical overlap, or BM25 relevance, alone or learned in combination, give weaker overall trade-offs, the gate transfers without re-tuning across deciders and agent scales, and its drafter distills into a 921k-parameter head atop the frozen encoder, leaving no LLM in the routing loop. The same signal cannot replace verification: matching a claim is far easier than deciding the claim is supported.

★ Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibratione

Zixuan Wang, Bingjie Zhang, He Zhao, Dandan Guo — Wed, 23 Sep 2026 00:00:00 +0000

93% · reliability 93%, Efficiency 35%, capability 33%

Large language models (LLMs) aligned for safety often suffer from over-refusal, incorrectly rejecting benign yet safety-related instructions. Prior studies primarily attribute this to static representation overlap, largely overlooking the underlying dynamic mechanisms. In this paper, we present the mechanistic analysis of over-refusal through the lens of internal routing conflicts within transformer attention. We discover that a sparse subset of Hypersensitive Safety Heads misfires on Hard-Safe prompts, exhibiting abnormal attention entanglement that forcefully binds harmless target entities to refusal semantics. This triggers a severe, high-entropy routing conflict that deprives target entities of necessary attention. To counteract this, we propose Semantic Routing Calibration (SRC), a lightweight, training-free inference framework. SRC precisely localizes and dynamically suppresses these hypersensitive safety heads at the inference stage. Coupled with a dual-branch logits fusion that acts as a safety regularizer during subsequent decoding, SRC seamlessly restores trustworthy reasoning. Extensive experiments demonstrate that SRC alleviates over-refusal, with intrinsic safety performance preserved as much as feasible.

★ From Offline Proxies to Online Decisions: A Layered Engagement Evaluation Framework for Conversational AI

Xuanyi Li, Vaskar Nath, Hossein Amirkhani, Jay Li, Alex Deng — Wed, 23 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, reliability 51%

Online A/B experiments are the decision standard for user engagement, but traffic and readout time limit how many conversational-AI changes can be tested. We ask whether an offline signal designed to be computable without treatment-arm user exposure agrees with the outcomes of those experiments. We contribute a reusable construction and diagnosis checklist that treats an offline proxy as a chain of three alignments: behavioral label to product outcome, learned classifier to candidate-assistant behavior, and aggregated offline signal to experiment effect. A companion evaluation protocol audits the whole composite by interval-aware decision agreement, which compares offline and online confidence intervals instead of point estimates, and by within-experiment ranking. The instantiation we evaluate comprises a fixed evaluation suite on which candidate behavior is scored, an engagement classifier trained to predict session/prompt level engagements, and a calibration layer mapping sample-level score differences to online model-level engagement deltas. We then report the audit: 489 paired offline-online contrasts (one candidate arm against its control) from 27 experiments on a deployed multi-turn assistant, spanning model checkpoints to system-prompt tuning. Our primary test uses the 113 contrasts from eight experiments that ran after the map was frozen: on these the composite reaches 81.1% F1, against 34.3% for the raw classifier score it is built on, and makes no wrong-direction calls where that raw score makes 31. Every offline prediction was computed before its experiment ran to prevent overfitting. The evidence supports using the composite to prioritize candidates before scarce experiment traffic is allocated---in our deployment of the experiment, selecting among training checkpoints and tuning system prompts.

★ A Semiotics-Aware Framework for Evaluating Fidelity and Coverage in Natural Language Generation

Lorenzo Zangari, Davide Picca — Wed, 23 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, reliability 49%

When two texts describe the same expression, standard metrics based on lexical overlap or whole-text similarity may fail to detect meaningful differences in how that expression is framed. We propose a framework to evaluate semiotic alignment between texts, where a semiotic profile encompasses both the contextual meaning and the discourse references made salient by a text. Our approach yields two scores, Semiotic Fidelity and Semiotic Coverage, estimating how much of one text's profile is supported by the other and how much of the other's profile it recovers. Experiments show that coverage is typically lower than fidelity, and that alignment between LLMs and human-curated data is highest at low sampling temperatures, while higher temperatures reduce this alignment.

★ The Delegation Blind Spot: Auditing Product Decisions from Agent Choices

Shivam Gupta — Wed, 23 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, reliability 72%, Agents 52%

Successful agent execution need not identify which future product improvement its user would value. We present a decision-specific audit that maps a declared observation channel and product-value contrast to compatible intervals and witness populations. Its foundations are established identification and decision theory; the contribution is an executable measurement workflow and a controlled study of its limits. A frozen experiment makes 4,800 requests to two pinned model snapshots on shared synthetic tasks. All 36 conservative primary intervals remain unresolved despite different execution accuracy. An exploratory 2,400-call follow-up records supplied preferences and resolves three of nine comparisons per model. A deterministic extractor resolves seven of nine without model calls or calibration observations, exposing unnecessary uncertainty introduced by model-generated reports. A further 14,400 controlled multinomial simulations distinguish structural ambiguity from weak identification and finite calibration precision. We propose a source-labeled decision receipt and provide an offline viewer for inspecting the audit. These results motivate preserving decision-relevant structured input and diagnosing why a decision is unresolved before collecting more telemetry. The study contains no human participants or real customer outcomes. Full proofs, raw model provenance, controlled experiments, and reproducible analyses accompany the report.

★ Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation

Lijuan Tang, Yuemeng Zheng — Wed, 23 Sep 2026 00:00:00 +0000

93% · Evaluation 93%, Agents 78%, reliability 45%

A coding agent must emit a valid tool call--a parseable invocation of a tool in the provided schema--before the harness can execute its chosen action. We study how local serving stacks affect this protocol step and show that measured outcomes can depend on the serving layer rather than model behavior alone. In Ollama, the default tools= request is gated per model by a static template flag: some models are accepted and return calls as text, some return native tool_calls, while Phi-3 and Gemma-3 are rejected before inference. In our harness, rejection and retry exhaustion are not preserved as structured failure metadata, so downstream analysis can misclassify them as model non-calls and naively report 0% fidelity. Adding a text tool list while retaining the native channel recovers much of the measured fidelity for accepted models, whereas a uniform text protocol reduces fidelity for Llama-3.2, which has native tool-call support. Cross-stack probes on Ollama, llama.cpp, vLLM, and SGLang show different handling of the same request. Constrained decoding removes parse failures but can induce non-termination, and turn-pooled versus per-instance estimates differ by up to about 55 points. We conclude with a checklist for treating serving behavior as part of the evaluation protocol.

★ Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents

Laizhen Li, Jiarui Li, Juanjuan Zhao, Kejiang Ye, Ye Li, Cheng-zhong Xu — Wed, 23 Sep 2026 00:00:00 +0000

93% · Efficiency 93%, Agents 90%, capability 66%

Large language model (LLM) agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. We study whether task feedback can instead turn recurring control into reusable executable code, while reserving LLM calls for task-specific semantic reasoning. We introduce Growing Harness, a failure-guided training paradigm that learns the agent harness itself from a strategy-free scaffold that exposes fixed model and tool interfaces but encodes no task-solving controller. Function-level execution traces localize each failure to a bounded code surface, an optimizer repairs a window of failures jointly, and a success-first held-out gate rolls back repair sequences that harm prior capability. Accepted edits accumulate in one shared harness, allowing its control structure to emerge from task feedback. Across BrowseComp-Plus and WebArena-Verified with three deployment models from 4B to 120B parameters, Growing Harness achieves the highest mean success in five of six benchmark-model settings and trails the best mean by 0.7 pp. in the sixth. Relative to a Tool-Calling agent, it reduces LLM calls by 76.0-91.8% and deployed-agent inference cost by 74.4-98.6%. On WebArena-Verified, its success remains 44.7-45.3% across model scales, whereas Tool-Calling falls to 6.7% with the 4B model. Ablations show that trace-local edits, joint repair, and gate-based rollback each improve final success. These results show that persistent program growth can move recurring control out of model context and into low-cost code, yielding reusable specialist agents that remain effective with smaller deployment models.

★ What Does 99% Accuracy Measure? A Reproducible Audit of Shortcut Learning in a Widely Used Fake News Corpus

Yuvraj Verma — Wed, 23 Sep 2026 00:00:00 +0000

92% · Evaluation 92%, reliability 70%

Text classifiers trained on the ISOT/Kaggle "Fake and Real News" corpus routinely report accuracy and F1 above 0.98, a level of performance that sits uneasily beside the difficulty of assessing veracity. Using a transparent TF-IDF and linear-classifier pipeline as a measurement instrument, we audit the corpus along three leakage channels and two distribution-shift protocols, releasing all code and derived numbers. First, the benchmark is partly degenerate: a classifier given only the subject metadata field, with the article text discarded, attains F1 = 1.000, since the two classes have disjoint subjects. Second, removing all three leakage channels, metadata, a newswire source tag present in 99.2% of real articles, and 6,251 duplicate documents contaminating 19.4% of a naive test split, lowers F1 by only 1.21 points (0.9935 to 0.9814); the residual signal is diffuse editorial style rather than a few giveaway tokens, since deleting the 1,000 highest-weight unigrams still leaves F1 = 0.926. Third, this style signal does not transfer: under a topic-disjoint protocol, average precision falls from 0.9995 to 0.9475 and deployed F1 from 0.9905 to 0.8067, with a prior-matched analysis confirming a genuine 5.2-point loss of discrimination, while temporal transfer is nearly lossless. A fine-tuned DistilBERT is stronger in-distribution (F1 = 0.9993) but degrades far more under topic shift, losing 12.9 average-precision points against the linear model's 5.2. Transferred to the independent LIAR benchmark, all three models fall to near-chance ranking (ROC-AUC 0.54-0.57), none beating a majority-class baseline. We conclude that within-corpus scores here quantify source and topic separability rather than veracity, that added capacity exploits the shortcut rather than avoiding it, and we recommend metadata-only, small-sample, and topic-disjoint baselines as inexpensive diagnostics for future work.

★ Understanding Reliability in LLM-based Human Behavior Simulation

Pei Wang, Lei Wang, Yuanzi Li, Xu Chen — Wed, 23 Sep 2026 00:00:00 +0000

92% · Evaluation 92%, reliability 69%

Large language models (LLMs) are increasingly used to simulate human survey responses and behavioral reactions, yet unreliable simulations can mislead social science conclusions. However, existing evaluations focus on end-to-end scores, leaving it unclear how different aspects of the simulation process interact to determine reliability. We propose ReliMap, which decomposes LLM-based human behavior simulation into three structured layers and evaluates reliability at both the individual level (R1) and population level (R2) across three configuration dimensions: model capacity, profile completeness, and population coverage. Through experiments across four simulation tasks and eleven LLMs, we find that all models exhibit substantial distributional bias without profile conditioning. Profile conditioning reduces this bias with diminishing returns. Larger models benefit more, and attribute informativeness matters more than quantity. Critically, R1 gains do not reliably transfer to R2--individual and population-level reliability can move in opposite directions. At the population layer, increasing coverage reduces variance but not systematic bias, with R2 stabilizing at around 50-100 individuals. These findings highlight that reliable simulation cannot be achieved by optimizing any single layer in isolation, but requires coordinated improvement across all three.

★ Impact Is Not Invalidation: Ask About the Claim, Not the Diff

Atul Anand — Wed, 23 Sep 2026 00:00:00 +0000

92% · Evaluation 92%, reliability 69%, Agents 68%

Memory systems for coding agents must decide, when a repository changes, which of their stored claims have become false. Content anchoring invalidates a claim whenever the artifact it came from changes, which fires constantly. Semantic-equivalence classification asks whether a diff preserves behavior, a question about the diff rather than about any stored claim. We show the second signal fails for a reason unrelated to model capability: asked whether a commit preserves behavior, five models spanning a 40x price range fire on 59-72% of real commits and reach precisions of only 0.291 to 0.329 against a 0.25 base rate. Asked instead whether one specific claim still holds, the same models on the same diffs reach 0.705 to 0.974. A control that hands the behavior-preservation judge the claim text, changing only the question, moves precision by 0.010 and 0.016; changing the question moves it by 0.49 and 0.65. We also compare against pytest-testmon, a deployed regression-test selector with coverage-derived dependency data: it reaches 0.868 recall at 0.415 precision, so near-complete knowledge of what a change can reach does not identify what it falsifies. Ground truth is execution, not annotation: a claim is a test function passing at commit t, and it has flipped if that same assertion text fails at t+1. Building this required an observation we did not find in prior work. On a CI-gated mainline a commit that leaves a pre-existing test failing cannot merge, so the naive construction has an empty positive class by design. We report 10,369 claims with 184 execution-verified flips mined from 23 Python libraries, splits held out by repository, a post-knowledge-cutoff split, a shuffled-diff null, a paraphrase control, and a leave-one-repository-out analysis over 17 repositories.

★ When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning

Jianzhe Lin, Xiaolin Li, Yunda Liu, Fei Wang, Jubin Chheda — Wed, 23 Sep 2026 00:00:00 +0000

92% · Evaluation 92%, Agents 82%, reliability 34%

A social agent's most basic decisions (should I react to this post? who should I reach out to?) are not purely content problems. The right action often hinges on the latent relationship between people -- tie strength, reciprocity, mutual connections -- rather than on which content is most salient. Standard LLM agent loops do not explicitly represent how new relational evidence should revise the agent's current social hypothesis, leaving them prone to surface-obvious choices when relational and content cues diverge. We formalize this failure mode with a relationship-reasoning benchmark: 500 synthetic social worlds with friendships, follows, reaction histories, and feeds, yielding 1,000 queries over two tasks, reaction selection and warm introduction (finding the best bridge to a target person). By construction, the surface-obvious candidate differs from the relationship-grounded oracle in about 53% of queries, forming an overturn subset where the agent must use relational evidence to revise an initially plausible choice. We propose ReAdapt (Relationship-Adaptive Agent with Policy-driven sTate), which augments the ReAct loop with an explicit structured social state z = (G, B, R, N, D) capturing goal, belief, relationship, norm, and disclosure. After each tool observation, ReAdapt runs a typed Adapt step that updates this state and emits a policy operation (continue, switch, abandon, or clarify) before choosing the next action. With Gemini-3-Flash on a stratified subset of n = 150 queries per task, ReAdapt improves warm-introduction accuracy from 37% to 51% (+14 points) and reaction-selection accuracy from 69% to 77% (+8 points). Oracle regret drops from 0.260 to 0.152 and from 0.095 to 0.053, respectively. Holding the model, tools, and environments fixed, these results suggest that explicit relational-state adaptation helps LLM agents turn retrieved social evidence into revised decisions.

★ Testing-Driven Reliability Audit of Trajectory-Based Early Outcome Prediction for LLM Agents: Target-Specific Calibration Transfer Persists Within a Single Benchmark

YanZe Cao — Wed, 23 Sep 2026 00:00:00 +0000

92% · reliability 92%, Agents 87%, Efficiency 78%

Predicting early outcomes based on trajectory can decrease the expenses associated with agent evaluation by terminating a run once the outcome becomes sufficiently predictable, assuming that the predictor's confidence is properly calibrated. Calibration is at risk when a predictor is applied to an agent on which it was never trained, but it is not known whether such transfer failures are broad across agent systems or concentrated in specific target agent/head combinations. Using public SWE-bench Verified trajectories and a frozen dual-head early-outcome prediction pipeline, we ran a leave-one-agent-out calibration audit, a shared-predictor leave-two-agents-out control, oracle prior correction, and a robustness battery over training cohorts, task resampling, task halves, jackknife, and thresholds. Fixed-scaffold TerminalBench analysis served as a pre-registered boundary test. Broad same-predictor pairwise heterogeneity was not supported; the median pairwise corrected-gap differences were 0.0180 (SUCCESS head, 45 pairs) and 0.0385 (FAILURE head, 35 pairs), and the pre-registered heterogeneity criterion was not met on either head. Two specific combinations, gpt-5-mini/SUCCESS and claude-opus-4.6/FAILURE, showed persistent calibration-transfer errors (median corrected gaps 0.1377 and 0.1107) without a sign reversal under any frozen control. TerminalBench did not establish cross-benchmark replication: the success target produced zero decisions (INDETERMINATE), and the failure target did not satisfy the pre-registered persistence criterion. Therefore, a strong target-specific calibration-transfer error can exist within one frozen environment, but the evidence does not establish that the error is intrinsic to the model or general across benchmarks.

★ From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs

Zhentao Tan, Chang Liu, Yao Liu, Yue Wu, Jieping Ye — Wed, 23 Sep 2026 00:00:00 +0000

92% · Efficiency 92%, capability 52%

As large language models (LLMs) scale rapidly, dense full-parameter adaptation becomes increasingly expensive, motivating sparse and modular architectures such as Mixture-of-Experts (MoE) models. This shift raises a key question for parameter-efficient fine-tuning (PEFT): at what granularity should parameters be selected and updated? Existing PEFT methods such as LoRA operate on predefined weight matrices, while expert-level sparse tuning methods update entire selected experts. However, we observe that activated experts are internally sparse, with only a small fraction of intermediate channels strongly responding to downstream tasks, indicating that expert-level adaptation is still too coarse. We propose NSFT (Neural Sub-expert Fine-Tuning), a fine-grained PEFT framework that refines MoE adaptation from experts to sub-experts. NSFT decomposes each expert along the intermediate dimension into structured channel groups and selects task-relevant sub-experts by combining routing importance with intra-expert activation saliency. To optimize sparse partial updates, NSFT further introduces learning-rate scaling and dynamic gradient scaling to compensate for the reduced effective update magnitude. Experiments on OLMoE and Ling-mini-2.0 across challenging domain-specific tasks and general benchmarks show that NSFT consistently outperforms representative PEFT and expert-level sparse tuning baselines, while using substantially fewer trainable parameters and preserving competitive general capability. These results suggest that sub-expert-level adaptation is a more precise and efficient PEFT paradigm for MoE LLMs.

★ Certified Against Which Oracle? Execution Labels Set the Reported Risk of Conformal Abstention for Text-to-SQL

Jiamiao Liu, Dewen Qiao, Yu Zhang, Xuetao Chen — Wed, 23 Sep 2026 00:00:00 +0000

92% · reliability 92%, Evaluation 87%, Agents 37%

A conformal abstention certificate for text-to-SQL is only as truthful as the correctness labels it is calibrated on. The uncertainty pipelines that read confidence off execution consistency take those labels from the single database a benchmark ships, an oracle known to be lenient. We run a preregistered intervention on Spider-Realistic, swapping that database for the benchmark's distilled multi-instance test suite. Across four SQL-specialist checkpoints and two split schemes, the swap raises the certificate's held-out risk 2.73 to 10.23 points above the risk its own labels report. Neither oracle reports the risk experts assign. Under blinded labels from two SQL experts, a certificate calibrated at a nominal 0.10 carries 20.0 and 17.2 points of risk on two checkpoints. The stricter oracle errs in both directions: most of the answers it rejects are not judged wrong, and some of those it accepts are. An AI-assigned census of what it rejects finds a semantic error in a quarter to a third of them, depending on the population. It attributes most of the rest to underspecified questions, synthetic instances or suspected reference-query defects, a flag supported by a preregistered blinded expert audit. The oracle also decides how a confidence score is judged. Every execution-consistency score looks better under the labels of the oracle that built its clusters, in 16 of 16 combinations. Under expert labels, building such a score on suite clusters instead of shipped-database clusters raises its area under the ROC curve (AUROC) by 6.96 points on one checkpoint and 1.53 on the other. On the second, the expert interval excludes the 8.3 points the suite labels report. A certificate should be reported with both oracles, and an oracle-relative difference read as semantic risk only after the benchmark is audited. A consistency score should be evaluated under an oracle that did not build it.

★ ChainUQ: Reasoning Consistency-Aware Uncertainty Quantification for Large Language Models

Dahai Yu, Rongchao Xu, Lin Jiang, Ximiao Li, Guang Wang — Wed, 23 Sep 2026 00:00:00 +0000

92% · reliability 92%, capability 58%, Evaluation 38%

While large language models (LLMs) exhibit impressive reasoning capabilities, response-level confidence may remain unreliable when intermediate claims conflict with the final conclusion. Therefore, effective uncertainty quantification (UQ) is required to capture logical inconsistencies within the reasoning chain, not just the correctness of the final output. Current approaches have two major limitations: (1) their reliance on token-level probabilities fails to capture reasoning consistency, and (2) they lack mechanisms to dynamically calibrate confidence using the structural logic of the generated chain. To advance existing research, we introduce ChainUQ, a reasoning consistency-aware uncertainty quantification framework for LLMs. ChainUQ consists of two key technical components: an alignment-aware lightweight UQ module that estimates a raw intrinsic model confidence score from frozen features aligned to the final conclusion, and a reasoning consistency-aware calibrator that refines this score using reasoning-chain consistency evidence. Evaluations across diverse in-distribution and out-of-distribution benchmarks show that ChainUQ consistently improves response-level uncertainty estimation, achieving an average 3.1% relative gain in AUROC and up to 45.0% relative reduction in ECE, and can be directly transferred to new settings without additional fine-tuning.

★ DTOC: Dynamic Tool Output Compression for Adaptive Context Management in AI Agents

Wed, 23 Sep 2026 00:00:00 +0000

92% · Efficiency 92%, Agents 91%, reliability 45%

As agent capabilities have grown, practical limitations increasingly stem from constrained context windows rather than model capacity. Common strategies, such as truncation, heuristic aging, and lossy summarization, may discard useful information or introduce hallucination risk. To address these challenges, we propose Dynamic Tool Output Compression (DTOC), a framework for scalable context management in LLM-based agents that models context updates as explicit and reversible operations within the agent reasoning loop. DTOC retains full tool outputs in external memory while inserting compact placeholders into the active context, enabling selective reconstruction when needed. We formalize the DTOC mechanism, integrate it into a ReAct-style agent architecture, and provide a production-oriented implementation supporting on-demand restoration of compressed outputs. Experiments on DeepSWE reveal model-dependent effects: for responsive models (Sonnet 4.6, GPT-5.4), DTOC reduces input tokens (10.3 and 12.7%) and agent steps (2.4 and 32.3%), while increasing solve rates (2.5 and 1.5 times higher) and lowering cost per solved task (3 and 3.5 times lower cost per solved task). For the other models results are more mixed, with GPT-5.5 doubling solve rate and halving cost, but no impact on solve rate and negative impact on cost for the other models. Ablation results show reversibility is critical: disable-only compression variants degraded performance, while full DTOC recovered baseline accuracy at substantially lower context cost. These findings indicate that explicit, reversible context management can improve the efficiency of long-horizon agent reasoning without degrading task performance.

★ REFLEX with Jev for Efficient Selective Control in LLM Agents

Tiantong Wu, Wei Yang Bryan Lim — Wed, 23 Sep 2026 00:00:00 +0000

92% · Efficiency 92%, reliability 77%, Agents 66%

LLM agents often use generative models for bounded decisions, raising the question of when these decisions can be handled more efficiently without reducing task success. We study REFLEX, an agent architecture that uses Jev as a fast, typed decision layer and calls a strong LLM when confidence is low, or generation is required. On a frozen 100-task benchmark, REFLEX achieves 95% success with 72.7% fewer strong-model calls than a strong-only agent, with reductions persisting across three fallback families. Controlled interventions show that reliability depends on action-set size and near-valid alternatives near authorization boundaries. External BFCL and $\tau$-style evaluations reveal limited advantages over a cheap generative cascade when ordinary routing is already highly accurate. These findings identify when selective control with Jev can reduce computation and where its benefits are limited.

★ JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

Yubo Li, Yidi Miao, Ramayya Krishnan, Rema Padman — Wed, 23 Sep 2026 00:00:00 +0000

92% · Efficiency 92%, reliability 81%, Evaluation 78%

LLM-as-a-judge enables evaluation across diverse tasks, but inference cost and confidence reliability become critical at scale. We study whether a decision-only judge can provide an economical first pass and identify when stronger evaluation is needed. Comparing jev-as-a-judge with sixteen generative and reward-model judges, with blinded human adjudication, we find it within three percentage points of a state-of-the-art LLM judge, our strongest comparator, on ordinary preference and evidence-grounded factuality at 0.36% of the comparator's fee. Larger gaps arise when judgments require checking a derivation or resisting an elaborately written wrong answer. On several benchmarks, JEV's gap to this comparator is concentrated in low-confidence decisions. A frozen cascade that accepts confident verdicts and escalates uncertain ones retains 99% of the comparator's accuracy at lower cost.

★ The Limits of Simulated Societies: How Post-Training and Survey Fine-Tuning Erase Cross-Cultural Variance

Rojin Ziaei — Wed, 23 Sep 2026 00:00:00 +0000

91% · Evaluation 91%, reliability 63%

Using large language models (LLMs) to simulate diverse human populations has the potential to transform many aspects of computational social science, yet many evaluations score the average response rather than the spread of opinion within real groups. Here, we develop a diagnostic framework that measures point accuracy alongside dispersion retention, the ratio of predicted to human standard deviation ($\dr$), on 10{,}000 respondent--question pairs from the World Values Survey (WVS) spanning twelve countries and six continents. We evaluate eleven zero-shot language models and five variants fine-tuned on WVS data with SFT, DPO, and GRPO. We identify a failure mode we term \textit{consensus collapse}, where alignment training compresses outputs toward one stereotype per group. Along the post-training trajectory from the Llama~3.1 70B base to the Tulu~3 checkpoints, the first stage, supervised instruction tuning, removes half of the spread with minimal accuracy gain ($\dr$ 1.22 to 0.59; accuracy $+0.9$ points), the later stages do not restore it, and a gap opens between WEIRD and non-WEIRD countries that survey fine-tuning then deepens while pursuing higher point accuracy. The most accurate model (Tulu~3 70B-DPO fine-tuned on WVS, 57.9\%) keeps half the human spread overall ($\dr = 0.50$) and 11\% of it for Nigeria, against 0.70--0.87 for WEIRD countries. Raising the sampling temperature to 1.0 leaves the Wasserstein-1 distance ($\wone$) to human distributions unchanged for both fine-tuned DPO models, and GRPO on Qwen~3.5 9B does not restore the spread under either an accuracy reward or a distribution-shaped reward. Mixing the aligned model with an unaligned prior raises $\dr$ from 0.51 to 0.62 on a held-out split but leaves Nigeria at 0.36. Point accuracy alone therefore misjudges these simulators, and current post-training trades diversity for consensus.

★ Block-Level Weight-Space Structure Persists Under Post-Training: An Empirical Study Across LLM Families

Zhaohui Wang — Wed, 23 Sep 2026 00:00:00 +0000

91% · Efficiency 91%, Evaluation 75%

Modern LLMs are deployed as families of post-trained variants (base, instruct, chat, code) derived from a shared set of pre-trained weights. We present an empirical study of how post-training transforms weight-space geometry, covering eight configurations across four architecture families (Qwen2.5, Llama-3.1/3.2, Mistral, Gemma-2). We identify a granularity gap: post-training modifies every tensor (zero of 291-339 tensors remain byte-identical, so hash-based deduplication achieves 0% savings), yet preserves block-level structure (mean cosine similarity exceeds 0.99 and relative Frobenius distance stays below 0.13). Post-training therefore acts as a structured perturbation that shifts every parameter while leaving block-level geometry intact. The property is not universal: independently trained specializations (for example, Qwen2.5-Coder) attain cosine similarity around 0.64 with the general base, indicating a disconnected region of weight space. Perturbation magnitude varies systematically with model scale, architecture family, and post-training recipe. As a practical application, we build LinkerLLM, a lazy loader that aliases shareable blocks across co-resident variants, achieving 18-48% GPU memory savings and enabling up to five 7B-parameter variants on a single 24 GB consumer GPU. Five of eight configurations retain at least 94% of the unshared variant's quality on MMLU, ARC-Challenge, HellaSwag, and WinoGrande; the remaining three (Mistral-7B, Gemma-2-2B, Llama-3.2-1B) have one below-threshold benchmark each (87-91%), which we report transparently rather than gate the block-sharing decision on a single threshold.

★ How to Estimate Whether You Have Found Several Needles in a Haystack: Measuring Calibration in Multi-Label Text Classification

Wed, 23 Sep 2026 00:00:00 +0000

91% · Evaluation 91%, reliability 88%, capability 52%

A key factor in deciding whether to trust an automatic prediction is its confidence score, which should be calibrated to match the actual probability of the prediction being correct. Most confidence calibration metrics target binary or multi-class tasks, while multi-label calibration remains largely underexplored. Multi-label classification tasks, such as assigning medical codes to clinical notes or determining news topics, are usually dominated by a large number of negatives, i.e., labels that do not apply. We show that existing binning schemes to compute label-wise expected calibration error either underestimate the error, simply reflect label frequency, or suffer from many bins with very few instances. To achieve trustworthy label-wise calibration errors, we propose a new binning scheme that gives equal weight to positive and negative label assignments. Our empirical study demonstrates that in contrast to existing binning schemes, our new scheme results in meaningful estimates of calibration error in hierarchical and in extreme multi-label classification. We also show that calibrating confidence scores of large language models for multi-label predictions is an open challenge. Our detailed analysis lays the foundation for further research by providing a solid evaluation metric for measuring calibration in multi-label classification.

★ The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence

Xiaoyu Yang, Jie Lu, Wei Duan, En Yu — Wed, 23 Sep 2026 00:00:00 +0000

91% · Evaluation 91%, capability 64%, reliability 30%

Long-context LLMs focus on retrieving distant evidence from extensive context, yet existing work has largely focused on overcoming distance alone. In this work, we identify the Proximity Trap, insufficient attention to distant evidence often arises less from distance itself than from cumulative competition with abundant, task-irrelevant proximal background. To address the Proximity Trap, we introduce LYRA (Long-context heavY-tailed Relevance Alignment), a t-distributed directional matching mechanism that reshapes the context retrieval distribution, directing more attention mass toward task-relevant evidence, while preserving the relative positional information encoded. Extensive experiments on LongBench-v2, RULER, and LongBench demonstrate consistent improvements across context lengths and task categories. We further introduce ProxBench, a multi-level fine-grained benchmark for evaluating distant evidence utilization under increasing proximal background interference. Project page: https://xiaoyuyoung.github.io/LYRA/

★ EquivSVA: A Formally Verified Dataset of Behavioral Assertions Across Equivalent RTL Implementations

FNU Aditi — Wed, 23 Sep 2026 00:00:00 +0000

91% · Evaluation 91%, reliability 84%

Large language models are increasingly used to generate SystemVerilog Assertions from natural-language specifica- tions and register-transfer-level designs. Existing datasets and benchmarks support important goals such as large- scale training, formal evaluation, specification-to-assertion generation, and mutation-based testing. A complemen- tary need is to study whether a generated assertion cap- tures externally observable behavior or depends on inci- dental details of one RTL implementation. We present EquivSVA, a formally verified dataset organized around behavior families. Each family contains four structurally distinct RTL implementations of the same externally ob- servable behavior, shared interface-level gold properties, three controlled mutants, and formal-validation evidence. EquivSVA contains 120 behavior families across 12 cat- egories, 480 reference RTL implementations, 914 gold properties, and 360 mutants. Every final family passes a fixed 17-job validation suite covering RTL equivalence, gold-property proofs, property reachability, mutant dis- tinguishability, and gold-property checks on mutants. We also provide fixed family-safe train, development, and test splits. As a small demonstration of the analyses en- abled by the dataset, we evaluate the publicly released, Apache-2.0-licensed Qwen2.5-Coder-7B-Instruct model on the held-out test split. Of 293 interface-only generated properties, 93 are formally sound, and the number of sound properties varies across equivalent implementations for 14 of 24 test families. These results illustrate how behavior-family organization can support controlled stud- ies of assertion-generation robustness without requiring changes in intended functionality. The dataset, generators, validation scripts, and case-study artifacts are publicly released at https://github.com/aditigupta96/EquivSVA.

★ AIBuildAI-2.5: Efficient Autonomous AI Model Development Through LLM-Guided Tree Search

Peijia Qin, Ruiyi Zhang, Qi Cao, Han Guo, Li Zhang, Pengtao Xie — Wed, 23 Sep 2026 00:00:00 +0000

90% · Efficiency 90%, Agents 88%, capability 83%

Autonomous agents that automatically build artificial intelligence (AI) models could broaden access to AI across science and engineering. A popular line of such agents frames model building as a code search problem and solves it by tree search, in which each node is a candidate program and the tree grows by generating a child program from a parent, and these agents now approach the capability of experienced AI engineers on realistic benchmarks. However, these agents have three weaknesses in efficiency that have not been fully addressed. First, only a small number of candidates can be executed within a realistic budget, so search rules that rank nodes by executed rewards, such as Monte Carlo-style tree search, rely on few and noisy scores and select the next node to explore less effectively. Second, no resource-aware strategy is used to schedule training jobs, which can lower hardware utilization and training efficiency. Third, every agent call is served by a single powerful model, which inflates inference cost. Here we introduce AIBuildAI-2.5, an agentic system that carries out the tree search with LLM agents and addresses each of the three issues. AIBuildAI-2.5 proposes a novel LLM-guided tree search, in which a judge scores each candidate on its expected improvement, grounding, and feasibility, and a selector ranks the pool of candidates from these scores and the state of the search. In addition, AIBuildAI-2.5 comprises a scheduler that launches training jobs with the current hardware resource status taken into account and a router that assigns lower-cost LLMs to less demanding tasks while reserving the most capable LLM for the most challenging sub-tasks in the AI model building workflow. AIBuildAI-2.5 ranks first on MLE-Bench with a medal rate of 73.3%, and outperforms a strong baseline on six autonomous AI research tasks from AIRS-Bench.

★ Clarification Is Not Correction: LLMs Fail to Let Go

Wed, 23 Sep 2026 00:00:00 +0000

90% · reliability 90%, Evaluation 73%, Agents 65%

Dialogue failures in language models are usually framed as memory failures: context too long, summaries lossy, a constraint forgotten. We argue this misses a deeper problem: in many conversations the model does not forget, it commits too early. An ambiguous early turn collapses into a single hidden interpretation, and later clarification is filtered through that commitment. We call this early posterior collapse: unresolved user intent collapsing into a committed task state before ambiguity is resolved. We study it with controlled dialogue tasks in writing, planning, and coding using Gemini-2.5-Pro and Gemini-2.5-Flash. Across thousands of trials, the same information in different orders yields different outcomes, even when the final dialogue contains equivalent task-relevant information. This order effect suggests later clarification is treated as extra context rather than a corrective signal: it refines a stale task state without invalidating it. Coding tasks are especially vulnerable, suggesting early assumptions get embedded in structured artifacts such as interfaces and control flow. Standard prompting and memory strategies do not reliably help: summaries can collapse ambiguity, and chain-of-thought can reduce explicit wrong commitment in reasoning traces without improving final task success. These findings motivate uncertainty-preserving state management. If assistants cannot let go of early interpretations, robustness cannot rely on post hoc correction alone; it must keep ambiguous early turns from hardening into one task state. Assistants should hold tentative hypotheses while ambiguity remains, ask before executing when high-impact ambiguity persists, and rebuild from a revised state when later evidence invalidates an earlier reading. Rather than one prompting fix, we aim to redirect research for interactive LLMs from retaining more context toward preserving uncertainty.

★ From Decorative to Load-Bearing: Task Difficulty Shapes the Causal Role of Chain-of-Thought

Renee Jia, Di Mu — Wed, 23 Sep 2026 00:00:00 +0000

90% · Evaluation 90%, reliability 80%, Agents 41%

Chain-of-thought (CoT) monitoring is only meaningful if written reasoning causally constrains the answer. We introduce continuation-based causal testing, an ablation-patch intervention that perturbs one reasoning step, truncates the chain, and forces the model to continue from the corrupted prefix. It measures how load-bearing a CoT is for the final answer, a behavioral notion distinct from mechanistic faithfulness. Across Gemma-2-9B-IT, Llama-3.1-8B-Instruct, and DeepSeek-R1-Distill-Qwen-7B on GSM8K, MMLU, and BIG-Bench Hard, CoT load-bearingness tracks model-relative task difficulty: on easy tasks models silently bypass their own reasoning; on hard tasks they follow corrupted steps and propagate errors. A matched 2x2 analysis shows task difficulty dominates perturbation type: error propagation rises 16x from GSM8K to BBH multistep arithmetic, and a variance partition over 28,584 continuations attributes 98.8% of explained deviance to task difficulty versus 0.8% to perturbation type. Reasoning-specific RL suppresses error propagation and compresses the gradient. A four-variant judge-sensitivity analysis and blind two-annotator study (n=500) show the error-propagation vs. non-propagation label is invariant to judge prompt, with perfect inter-annotator agreement (Cohen's kappa = 1.00). This gradient creates a structural problem for CoT-based oversight and AI safety monitoring: where the trace is easy to read it carries little signal, and where it matters errors propagate before a monitor can intervene. Linear probes on hidden states separate silent bypass, self-correction, and error propagation, but additive activation steering provides limited causal control, flipping only about 25% of error-propagation cases at best. Behavioral mode is readable but not reliably controllable.

★ Calibration as a First-Class Criterion in LLM Evaluation

Mario Sanz-Guerrero, Katharina von der Wense — Wed, 23 Sep 2026 00:00:00 +0000

90% · reliability 90%, Evaluation 76%, Agents 36%

Calibration of language models -- the alignment between expressed or implicit confidence and empirical correctness -- is a well-studied subfield within NLP. Methods to measure it already exist. The problem is adoption: outside this subfield, NLP research regularly introduces new models, datasets, and benchmarks without checking whether the model's confidence scores are meaningful. We argue that this adoption gap is a major obstacle to trustworthy LLM evaluation. Miscalibration causes problems in two distinct areas: at deployment, where overconfident mistakes cause real harm, and inside the research pipeline, where methods like LLM-as-a-judge, synthetic data generation, and active learning rely on calibrated confidence without verifying it. Standard calibration metrics only require two inputs per example: a confidence score and a correctness judgment. Most benchmarks in use today already provide both, meaning calibration can be reported immediately. For open-ended generation, however, defining these two inputs is still an open challenge. We argue that each NLP subfield should pair its main performance metric with a calibration score and call for treating calibration as an essential property of every model rather than a niche topic.

★ Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation

Alexandre Cristov\~ao Maiorano — Wed, 23 Sep 2026 00:00:00 +0000

89% · Evaluation 89%, reliability 79%, capability 30%

Marketers increasingly use large language models (LLMs) as "synthetic personas" to predict how an audience will react to a piece of copy before it ships, encouraged by evidence that profile-conditioned LLMs mimic human samples. But is that prediction actually valid against real behaviour - and does the persona machinery help? We present a sim-to-real validity study using the Upworthy Research Archive - thousands of headline A/B tests on shared real traffic, with measured click-through - as held-out ground truth. We compare a ten-persona panel, grounded in the real audience's demographics, against a no-persona zero-shot baseline that simply asks the model how likely a typical reader is to click. Two findings stand out. First, ground-truth reliability is the binding constraint: most A/B tests have no statistically distinguishable winner, so validity can only be measured on the reliable subset (n = 399). Second, and counter to the persona-simulation premise, persona conditioning degrades predictive validity: the no-persona baseline ranks variants markedly better (Kendall {\tau} = 0.361, a medium effect; top-1 accuracy 49.2%) than the persona panel ({\tau} = 0.084; top-1 34.6%), with non-overlapping confidence intervals. Asking the model directly taps an accurate population-level prior; forcing it to role-play specific personas injects bias and noise. The result replicates across three independent Upworthy splits, holds in direction on a different-domain news dataset, and is robust to seed, prompt phrasing, and model choice - across three Gemini tiers and a different model family (OpenAI gpt-4.1, significant paired gap). The takeaway: for predicting aggregate engagement, a plain LLM ranker beats persona simulation - synthetic personas are not merely a weak predictor, they are worse than not using them. All numbers regenerate from a public, artifact-first replication package.

★ Retrieved-Span Training for Efficient Query-Focused Meeting Summarization on QMSum

Edward Xi Yang (Ertas AI) — Wed, 23 Sep 2026 00:00:00 +0000

89% · Efficiency 89%, Evaluation 68%

QMSum provides no scorer, making query-focused meeting summarization results difficult to compare. We rescore or generate 15 systems under one implementation. Through a common inference port, a released 406M Fusion-in-Decoder specialist loses 6.30 ROUGE-1 when moved from capped long input to 2,000-word retrieved spans. Fine-tuning it on this span regime recovers the loss. On test it scores 36.33 ROUGE-1 versus 35.41 for our 1.2B system; the meeting-cluster 95% interval for the difference is [-0.27, +2.22], so QMSum does not statistically separate them. The smaller system uses about one-third as many total parameters and less than half the peak inference memory. Within the fixed 1.2B base, span-regime fine-tuning adds 5.29 [+4.02, +6.56], while replacing the first 4,500 transcript words with 2,000 retrieved words adds 1.55 on test and 0.29 on validation. Separately, under one concise prompt and reference-overlap scorer, a released 406M specialist exceeds five proprietary hosted models by at least 6.2 ROUGE-1, but output length and absent human or factuality evaluation limit this ordering. Conclusions are limited to QMSum and automatic metrics.

★ ChainDoRA: Tensor-Train Factorized Weight-Decomposed Low-Rank Adaptation for Parameter-Efficient LLM Fine-Tuning

Ashfak Yeafi, Mehedi Hasan, Md Khairul Islam — Wed, 23 Sep 2026 00:00:00 +0000

89% · Efficiency 89%, capability 52%

Parameter-efficient fine-tuning (PEFT) adapts large language models (LLMs) to downstream tasks while updating only a small fraction of their pretrained parameters. Low-Rank Adaptation (LoRA) uses two trainable low-rank matrices, while Weight-Decomposed Low-Rank Adaptation (DoRA) further separates weight magnitude and direction but retains the dense LoRA-style factorization in its directional branch. We propose ChainDoRA, a weight-decomposed adaptation framework that constructs the directional low-rank factors from a connected Tensor-Train (TT) chain, where the adapter rank forms the boundary rank between input- and output-side TT contractions and an independent TT rank controls representation capacity and parameter cost. Under a controlled 15,119-example response-only adaptation setting with LLaMA-7B, ChainDoRA is evaluated against matched LoRA and DoRA baselines on seven commonsense reasoning benchmarks. ChainDoRA with TT rank 16 achieves a seven-task average accuracy of 72.30%, compared with 69.88% for LoRA and 69.39% for DoRA, while requiring only 5.35M trainable parameters versus 56.10M for LoRA and 56.98M for DoRA, corresponding to a 90.62% reduction relative to DoRA. Ablations over TT rank and adapter placement show controllable parameter-accuracy trade-offs, indicating that connected TT parameterization can substantially reduce the parameter cost of magnitude-direction adaptation while preserving, and in this setting improving, downstream reasoning performance.

★ Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs

Wed, 23 Sep 2026 00:00:00 +0000

89% · reliability 89%, Evaluation 54%

In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction tuning changes how models behave under conflict, but whether it rewires the underlying circuit or merely gates/reweights already present components, remains unknown. We provide the first mechanistic base-vs-instruct comparison of conflict-resolution circuits, across three families (Llama-3.2-3B, Qwen-2.5-3B, Gemma-3-4B). Five independent methods, node and edge attribution, superposition role analysis, causal ablation, and path patching, converge on gating, with the same heads, in the same late-layers, are found to be reweighted rather than replaced with a high node overlap (0.60-0.82). Behaviorally, tuning shifts models toward parametric memory, making instruct models reject a terse counterfactual context far more than base ones, the opposite of a naive user-following expectation. Yet this added skepticism is a factor of framing since it disappears when the same false claim is delivered as a coherent, evidential passage. The robustness that instruction tuning buys against terse injection is therefore real but narrow. More broadly, we believe that because the conflict circuit is preserved rather than rebuilt, interpretability and control tools calibrated on base models should transfer directly to their deployed instruct siblings.

★ Modular Norm RandOpt: Population-Efficient Ensembling through Architecture-Aware Perturbations

Kirato Yoshihara, Hiroaki Hamade — Wed, 23 Sep 2026 00:00:00 +0000

89% · Efficiency 89%, capability 80%, reliability 35%

RandOpt samples weight-perturbed language models and ensembles top-ranked candidates through plurality voting, but its global perturbation scale ignores heterogeneous module geometry. We propose \mbox{\textbf{\emph{Modular Norm RandOpt}}}, an architecture-aware sampling method using module-wise natural norms and calibrated scales while preserving selection and voting. It outperforms RandOpt using $3\times$ fewer candidates on Countdown and at least $12\times$ fewer on GSM8K, with corresponding wall-clock savings. Evaluations across seven tasks and three Qwen scales ($0.5$B--$3$B) show higher mean accuracy than RandOpt on Countdown, GSM8K, and MATH-500 at every scale. The gains extend to Llama 3.2 $3$B and Gemma 3 $4$B on Countdown and GSM8K. On Qwen2.5-1.5B, our ensembles also achieve higher mean accuracy than iterative baselines on both tasks at comparable main-run evaluation budgets. On GSM8K, a tail-density diagnostic implies only a $1.2$--$1.8\times$ candidate reduction, while most ensemble improvement is associated with more favorable correct-expert support. These results highlight perturbation geometry as a key design choice for population-efficient, gradient-free search around pretrained models.

★ Beyond Scalar Sensitivity: Activation-Aware Mixed-Precision LLM Quantization with Cross-Layer Refinement

Akihiro Yoshida, Yuma Ichikawa — Wed, 23 Sep 2026 00:00:00 +0000

89% · Efficiency 89%, capability 75%

Mixed-precision weight quantization is commonly formulated as a Multiple-Choice Knapsack Problem (MCKP), yet existing solvers rely on scalar sensitivity proxies that collapse each weight matrix's Hessian into a single number and treat every module independently. We prove that even the optimal scalar proxy incurs multiplicative distortion up to $\sqrt{\kappa(\mathbf{A})\kappa(\mathbf{B})}$ relative to the full activation-aware quadratic, where $\kappa(\mathbf{A})$ and $\kappa(\mathbf{B})$ denote the condition numbers of the input- and output-side Hessian factors. This bound varies from $10^1$ to $10^{13}$ for typical LLM modules, making inter-module sensitivity ranking unreliable. To address these limitations, we propose Cross-layer Activation-aware Sensitivity Allocation (CASA), a two-phase method. In Stage 1, the scalar proxy is replaced by an activation-aware metric derived from the Kronecker-factored Hessian, reducing the MCKP to a form whose continuous relaxation admits a closed-form solution. In Stage 2, a cross-layer-aware local search evaluates bit-width updates using the end-to-end model loss. Experiments on multiple LLMs across different bit budgets show that CASA achieves lower perplexity than the latest scalar-proxy baselines, especially at ultra-low bit-widths ($<3$ bits per weight). Moreover, the performance gain in zero-shot accuracy tracks the per-model average condition-number over modules, confirming the distortion bound as a practical indicator of scalar-proxy failure.

★ Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance

Sebastian Cochinescu — Wed, 23 Sep 2026 00:00:00 +0000

89% · Evaluation 89%, reliability 85%

Conversational agents often express answers in a uniformly confident register. We test whether expressed uncertainty, provenance-aware assertion, and explicit belief revision can be implemented as a behavior layer over a fixed language model; we do not test believability or trust. The layer combines three epistemic states, per-claim confidence and typed provenance, a provenance-gated expression rule, and a persistent revision store with auditable acknowledgments and partial resistance to false corrections. We evaluate it on a constructed, mechanically scored multi-session benchmark using a synthetic model and Qwen2.5-0.5B-Instruct. The synthetic instrument passes all five checks. On the real model, acknowledgment soundness, a by-construction guarantee, holds in 100% of cases, and true corrections are accepted more often than false ones (0.44 vs. 0.15 on held beliefs; 0.875 vs. 0.420 including rule-accepted corrections of unheld facts), but the pre-specified expression-fidelity, contradiction-separation, and provenance margins fail. A disclosed post hoc analysis shows that expression gated on mean answer-token probability ranks correctness below chance end to end (AUC 0.41, conversation-clustered), whereas gating on sampling consistency discriminates (AUC 0.66). A consistency-gated configuration selected from this finding and evaluated under a separately committed protocol meets the conversation-level manipulation and capability-equivalence criteria and replicates on a redrawn conversation set. The manipulation result is selection-dependent, and both criteria remain unresolved when uncertainty is clustered over the 60 facts. The supported conclusions are limited to the by-construction audit guarantee, store-dependent partial correction discrimination, and a benchmark- and model-specific failure of token-probability gating; scaling the fact base is required before human evaluation.

★ Adversarial Course-of-Action Generation: Game-Theoretic Multi-Agent Algorithms for COA matching & COA generation

Natan Vidra, Alina Kapanova, Arun Kanhai, Spurthi Setty — Wed, 23 Sep 2026 00:00:00 +0000

89% · Evaluation 89%, Agents 51%, reliability 39%

Course-of-action (COA) generation is a distributed planning problem: a system must propose structured candidate actions, evaluate them against an adversarial response, and surface options that remain tactically coherent under changing conditions. We present COA-Bench, a small offline benchmark and reproducibility artifact for comparing COA generation policies through self-play. Following the BattleCOA terminology, we reserve COA matching for asset-effect matching and COA generation for course-of-action generation; the present artifact does not implement either DecisionFunction directly. Instead, it represents COAs as typed action chains with conditional branches, assigns a synthetic COA quality score, compares opposing COAs with a BLUE-vs-RED advantage score and Nash-gap distance, and scores doctrinal coherence with an FM 3-0-inspired heuristic rubric. Across 50 synthetic scenarios spanning five operational templates, a sampled best-response policy that draws eight RED candidates reduces BLUE advantage from .516 to .485 and BLUE wargame win rate from .920 to .820; a two-stage multi-agent council with five BLUE proposer agents, RED-team adjudication, and critique-driven revision obtains .509 BLUE advantage and .820 BLUE win rate. We also identify and fix a benchmark-design issue in which scenario framing was stored as metadata but had no effect on generated COA content. COA-Bench is not an operational battle-management system and uses no real, classified, proprietary, or human-subject data. The contribution is an inspectable evaluation harness, preliminary benchmark evidence, and lessons for building auditable agentic planning artifacts.

★ Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks

Travis Weber, Rohit Taneja — Wed, 23 Sep 2026 00:00:00 +0000

88% · Efficiency 88%, reliability 78%, Agents 77%

On repeated work, agents are inconsistent. We ran 42 tasks three times each and found that, depending on the model, 38% to 74% returned answers that did not agree. Consistency is what a buyer, an auditor, or a regulator requires, and agents do not have it. They are wasteful too: 95.3% to 97.2% of what an agent generates goes to re-deriving a plan the system already knows. We propose skill habit formation. An agent mines its own execution history for candidate skills, deterministic variants that compete against the incumbent rather than replacing it. A candidate declares the region of input space it claims, so the common case runs as a script and the rest falls through to reasoning. Four gates of ascending cost admit candidates; the central one tests a candidate's execution trace against a retained reference, within a tolerance measured from that reference's own run-to-run variability. On text-to-SQL, three of four reasoning arms reproduced their own output on 11 to 13 of 42 repeated questions and the fourth on 26 of 42, while a habit-formed variant reproduced on all 456 dispatches we repeated and was non-inferior to every arm it replaced (p<0.0001). It also used 14% to 56% fewer tokens, turning net positive after 7 to 53 reuses. We measured what this costs in accuracy. The guard admitted work it should have deferred on 2.6% of natural paraphrases and 26% of inputs near its boundary, and 11 of 13 such failures were invisible to the trace-conformance gate at any threshold. Deterministic errors repeat exactly: a bad habit is as reliable as a good one, and that is the price of the property that makes the system auditable. Separating routing from parameter extraction raised end-to-end accuracy from 0.888 to 0.952 at 43% of the cost.

★ RAG-NAROK: Retrieval-Aware Knowledge Corpus Poisoning in RAG with Source-specific Refutation

Abdullahil Kafi, Alvi Ataur Khalil — Wed, 23 Sep 2026 00:00:00 +0000

88% · reliability 88%, capability 56%

Retrieval augmented generation (RAG) systems have emerged as the dominant architecture for grounding large language model (LLM) outputs in verifiable external knowledge, yet their structural reliance on a dynamic retrieval pipeline introduces a largely unexplored class of adversarial vulnerability. Existing knowledge-base poisoning attacks are fundamentally static. Adversarial documents are pre-computed and injected without any awareness of what the victim system will actually retrieve for a given query, leaving the attack blind to the competitive documentary landscape that surrounds its payload in the generator's context window. Unlike traditional static poisoning attacks that are blind to the retrieved context, we introduce RAG-NAROK (Retrieval-Anchored Generation Negation And Response Quality Collapse), a RAG attack framework that adapts to the query text. RAG-NAROK exploits the transparency inherent in RAG pipeline to first extract the legitimate source identities, then generate Anchor-Specific Refutation documents that explicitly name and devalue retrieved sources while leveraging recency and authority biases to steer the text generation toward a target answer. Our results demonstrate that RAG-NAROK significantly outperforms static baselines across diverse domains, revealing a fundamental tension between RAG transparency and AI security.

★ Qwen3.8-Omni: Towards Native Omni-Modal Agents

Qwen Team — Wed, 23 Sep 2026 00:00:00 +0000

88% · Agents 88%, Efficiency 45%, capability 42%

We introduce Qwen3.8-Omni-Flash, a natively multimodal agentic model for real-world multimodal productivity. Compared with previous omni models, which primarily emphasized perception and interaction, Qwen3.8-Omni-Flash substantially improves multimodal understanding and reasoning, as well as performance on long-horizon agentic tasks. These capabilities are supported by a native multimodal co-training strategy that preserves strong text-domain capabilities while facilitating the transfer of agentic capabilities from text to audio and video tasks. The model inherits the sparse mixture-of-experts (MoE) architecture of Qwen3.8-Next and extends the context window to one million tokens, supporting long-context multimodal reasoning and long-horizon planning. These advances enable integration into production workflows as a primary agent or a specialized sub-agent, supporting video editing, long-form audio and video translation, music-conditioned music video or movie generation, and video-based note or omni-skill creation. To address the lack of native audio and video support in existing agent harnesses, we release Qwen-MM-Plugins, a lightweight open-source plugin framework for multimodal productivity. We further frame real-time multimodal interaction as a system-level challenge requiring orchestration of context and memory management, tool use, and sub-agent delegation. Accordingly, we release Qwen-Live-Harness, an open-source framework for building responsive, real-time multimodal agents based on Qwen3.8-Omni-Flash. Extensive evaluations demonstrate that Qwen3.8-Omni-Flash achieves strong performance across multimodal understanding, reasoning, long-horizon agentic execution, and video productivity tasks. These results and the accompanying open-source tools support Qwen3.8-Omni-Flash as a practical foundation for deploying natively multimodal agents in research and production.

★ Fast Matrix Multiplication in fp8: Certified Coefficient Optimization and Measured Error

Shuxiao Xie, Shuyang Xie, Yuan Cao, Dezhi Ran, Wei Yang, Tao Xie — Wed, 23 Sep 2026 00:00:00 +0000

88% · Efficiency 88%, capability 65%, Evaluation 37%

A Strassen-type algorithm has many realizations with the same exact product and multiplication count yet different fp8 error because basis changes reshape coefficient geometry, posing the question of which to run. No current account settles this: classical stability controls worst-case $\ell_1$ growth, not the expected-error magnitude, and the Dumas--Pernet--Sedoglavic optimizer could only be called probably optimal, its global optimality unproved. To settle this, we attach to each realization a coefficient functional $\Phi$, a scalar summary of its coefficient geometry, which we minimize over the change-of-basis orbit. This Kempf--Ness problem on a Hadamard manifold lets us certify the global $\Phi$ optimum rather than merely search for it: an exact moment-map zero fixes $\Phi_{\min} = 200/9$, and de Groote's classification extends that optimality to every exact real rank-7 $2\times2$ decomposition. Every exact real rank-7 realization therefore has a $\Phi$-predicted RMS constant at least $5/3$ times that of the cubic algorithm, at fixed noise coefficient. We then introduce an explicit block-scaled e4m3 model in which $\Phi$ is the leading-order coefficient of relative expected mean-squared error, and we test the resulting $\Phi$-predicted ordering against realized fp8 error. Ordering and re-basing experiments support that prediction within tested fused block-scaled regimes, and on real matmul tiles from four architecture families the $\Phi$-optimal realization falls in the fp8 low-error region. Across two $\sim$70B models on real deep_gemm kernels, the same realization removes 10 to 55% of classic Strassen's excess NLL over the clean model. Algorithm realization thus becomes a mathematically certified design problem rather than a tuning choice: an independent low-precision axis with a global $\Phi$ optimum and measured fp8 relevance.

★ Beyond Imitation: Auditing the Recoverability of Reasoning in Distilled Models

Ruitong Li, Binjie Guo, Aisheng Mo, Guowei Su, Han Wang, Jie Li — Wed, 23 Sep 2026 00:00:00 +0000

88% · Evaluation 88%, reliability 69%, Efficiency 41%

A correct teacher solution becomes useful supervision when the receiving student can continue its reasoning. We measure this compatibility with prefix recovery: after revealing 25%, 50%, or 75% of a verified solution, we test whether the student completes it correctly. We connect recovery to the cosine conflict between cross-entropy and reverse-KL gradients over the full vocabulary. Across adjacent Qwen3 teacher-student pairs from 0.6B to 8B parameters, reverse-KL distillation delivers its most consistent mathematical and code improvements for the two students below 2B parameters. On a fixed cohort of 1,000 trajectories, average prefix recovery rises from 71.0% to 91.9% as student size increases from 0.6B to 4B, and the robust-fragile recovery gap contracts from 46.0 to 14.4 percentage points. With the teacher fixed at 8B, conflict separation falls from 0.993 to 0.233. An independent objective intervention finds the largest reverse-KL rescue on fragile trajectories. The three measurements locate the same capacity-dependent transfer regime: distribution matching has the greatest headroom when correct traces remain unevenly recoverable. Prefix recovery provides a practical diagnostic for selecting costly distribution-level distillation.

★ Coding Agents are Strong Prompt Optimizers

Wed, 23 Sep 2026 00:00:00 +0000

88% · Efficiency 88%, Agents 86%, capability 59%

Search-based prompt optimizers improve prompts through iterative search: they propose edits, execute fresh rollouts, score the resulting trajectories, and retain only edits that improve a validation metric. We show that this optimization loop is unnecessary. Given only a static corpus of agent trajectories, an off-the-shelf coding agent can directly synthesize an optimized prompt, requiring neither environment access nor validation data. We call this approach \textit{Coding-Agent Skill Distillation} (CASD). The key insight is reflection scope. Rather than reasoning over a small batch of trajectories at each optimization step, the coding agent writes and executes analysis code to compute corpus-wide statistics, identifies systematic failure modes, inspects representative episodes, and distills the resulting insights into behavioral rules. Across four agentic benchmarks (ALFWorld, $\tau^2$-bench retail and telecom, and SpreadsheetBench-Verified), under matched data access, a single CASD pass outperforms GEPA, a state-of-the-art reflective prompt optimizer, on three of four benchmarks and outperforms validation-gated reflective search (SkillOpt) on all four, improving the unoptimized baseline by 16.6 percentage points on average versus 10.9 for GEPA and 5.3 for SkillOpt. Because CASD performs a single offline analysis pass rather than iterative search, producing an optimized prompt costs approximately \$1.60---over $22\times$ cheaper than validation-gated search. Even when competing methods are granted additional validation data and unrestricted environment access, CASD remains ahead on two of four benchmarks. These results suggest that corpus-scale statistical reflection is a viable alternative to iterative search for prompt optimization.

★ Decoupling Is Not Identification: Supervised Evidential Learning in Next-Token Prediction

Ge Wang — Wed, 23 Sep 2026 00:00:00 +0000

88% · Evaluation 88%, reliability 86%

A next-token probability says what a model predicts, not how much training support lies behind it. A Dirichlet head can represent this distinction by separating mean $m$ from concentration $S$, but decoupling does not identify what $S$ means. Here we propose an Evidential Next-Token Prediction (ENTOP) framework to audit this gap on character-level Moby-Dick, using exact 8-gram count as a reproducible lexical-support label and withholding count regression from 20% of context types. Standard implicit evidential training carries essentially no count signal beyond confidence on held-out-label types (partial Spearman $\rho = 0.001 \pm 0.014$), whereas explicit supervision generalizes ($\rho = 0.201 \pm 0.010$; matched-pair win $= 0.822 \pm 0.021$). CE predictive entropy is at chance for unseen 8-grams (AUROC $= 0.490 \pm 0.004$), while supervised vacuity reaches $0.772 \pm 0.003$, comparable with an indexed CE-representation baseline ($0.769$) but below the tautological corpus oracle ($1.000$). Neither longest-suffix nor representation-distance strata explain where amortization succeeds. Increasing count weight under the digamma objective improves support fit only by sacrificing prediction. A constant predictor wins natural log-RMSE, and vacuity does not improve error deferral. These results motivate a minimum evidence protocol---confidence control, matched pairs, held-out labels, a constant baseline, and a decision test---and show that concentration can pass identification while failing calibration and utility.

★ Layout-Guided Masking for GROBID: Lightweight Structural Gains in Large-Scale Scientific PDF Ingestion

Luca Foppiano, Sana Khamassi, Vipul Gupta — Wed, 23 Sep 2026 00:00:00 +0000

88% · Efficiency 88%, capability 57%, Evaluation 45%

Transforming scholarly PDFs into machine-readable fulltext remains a bottleneck for large-scale information systems. Recent vision-based parsers improve accuracy, but need GPUs and may introduce noise into the extracted text. GROBID, a modular font-stream parser running on CPU, is the de-facto standard for structuring scientific articles and underpins several of the largest open scholarly corpora. We pair it with a lightweight CPU detector localising figure, table, and paratext (header, footer, page number) regions, encoded as typed-area masks whose tokens are routed to GROBID's specialised models or discarded. On two PMC corpora, Bioinformatics (1,926 articles) and Materials Science (2,595), scored against JATS with a section-aware structural protocol, our extension improves over plain GROBID on most metrics (NS $+0.025$/$+0.013$; $+0.086$ paragraph recall on Materials Science, $d_z{=}1.08$), and caption-linked figure recovery improves on both corpora. On the external Table-BRGM benchmark, table detection recovers F1 $0.16 \to 0.94$ and table structure follows (GriTS-Top $0.27 \to 0.78$, below the strongest GPU system). On body text, against four vision-based systems (Docling, MinerU, olmOCR, dots.ocr), it has the best paragraph precision on both corpora, the best section detection on Materials Science, and a character error rate within 0.004 of the best GPU parser. End-to-end on CPU, it costs $2.7$--$3.2\times$ less than the cheapest GPU system (Docling) and $10$--$14\times$ less than generative parsers.

Robust Failure, Conservative Repair: Textual Knowledge Distillation from Cross-Model Failures

Andrew Ren, Haokun Liu, Chenhao Tan — Wed, 23 Sep 2026 00:00:00 +0000

87% · reliability 87%, capability 60%

Failure-based textual knowledge distillation aims to discover gaps in a model's knowledge by examining its task errors. The distilled knowledge can be useful for the reasoning of both this model ("source model") and other models. However, this transfer of knowledge may not be stable. We define a rule atom to be a standalone rule injected into a model's textual input at inference time. A rule atom can encode transferable task knowledge or model-specific reasoning patches that can confuse other models. Also, the injected rule atoms can be misapplied to unrelated cases, causing the model to incorrectly flip its answer based on irrelevant information. Building on a pipeline that distills training examples into task-specific cheat sheets that aid model reasoning, we examine when failure-derived rules can improve these cheat sheets. Our early experiment shows rule distillation from a single model's failures underperforms the baseline cheat sheet on non-source model families. This motivates Robust Failure, Conservative Repair (RFCR), a textual distillation procedure that derives rules from failures shared across models, sharpens their application boundaries using boundary cases, and abstains when no useful rule is found. On a 400-item BIG-Bench Hard task set, RFCR improves the baseline cheat sheets from 68.50% to 71.25% (+2.75 pp; 95% CI [+1.25,+4.50]) without performance degradation on previously correct cases. Ablations and cross-model diagnostics support that accuracy gains come from both new knowledge injection and strict rule-application control.

FIRE: Failure-Informed Runtime Engineering for Reliable Language-Model Agents

Nikita Agarwal, Nivedit Jain — Wed, 23 Sep 2026 00:00:00 +0000

87% · Agents 87%, reliability 86%, capability 60%

Language-model agents often reach a working solution and then fail to consistently deliver it. We study runtime policies: targeted natural-language instructions and action denials applied by the agent harness at states that preceded observed failures, without changing model weights or the user prompt. With this, keeping capability constant, we observe a meaningful unlock in delivered reliability. Across the complete 87-task Terminal-Bench 2.1 suite, with two attempts per task, policies increase repeated success (pass^2) in all three GPT-5.6 tiers: 50.6% to 54.0% for Luna, 55.2% to 60.9% for Terra, and 64.4% to 73.6% for Sol. Sol's best-of-two success changes by 1.2 points while repeated success rises by 9.2, showing that policies chiefly convert reachable solutions into dependable delivery. We further cover 14 tasks under Terra's frozen portfolio. Policy-guided Terra reaches 71.4%, compared with 64.3% for unassisted Sol, at about half the cost, demonstrating how engineering around models could unlock dependability for a use case. To isolate the mechanism we run a randomized five-arm experiment: real policies reach 61% on eligible tasks, versus 39% without a policy, 36% with a timing-matched sham, and 39 to 43% with generic verification or reconsideration. The intended corrective behavior appears in 22 of 24 coded policy attempts, against at most 14 in any other arm. Runtime policies are therefore a practical reliability layer: they make capabilities an agent already possesses substantially more repeatable.

Self-Cleaning and Captured Anyway: One Measured Primitive for Error in a Store an Agent Writes to Itself, and What a Falling Score Actually Measures

Wenhui Chen, Jianlin Chen, Ziyao Lin, Chi Man Vong — Wed, 23 Sep 2026 00:00:00 +0000

86% · Evaluation 86%, reliability 82%, Agents 54%

"An agent that writes its conclusions into a store it later retrieves from closes a loop usually reported as one-way contamination. Taking the loop to the infinite-tenure limit against an append-only store gives a different picture: because writing never deletes, the reachable state space has a hard upper edge at (n-1)/n, so the outcome is a choice between two edges rather than a decay. At f_0 = 0.9 the interval between the two modes holds 3.6% of 220 runs where a uniform spread would put 20.6%, and is strictly empty on the first 15; the pooled mean describes 8.2% of the runs it summarises, the median 68.2%. Everything the model contributes is carried by one measured primitive with no fitted parameter, the copy function \gamma(\phi): on 36 Wikidata facts, sign(\hat{\gamma} - \gamma_{crit}), with \gamma_{crit} = 1/k at r = 0, w = 1, predicts the direction of drift on 353 of 360 real-fact runs (39 of 40 synthetic in the same batch). Scale does not rescue the store: pooled frontier capture is 0.850, with claude-sonnet-4.5 captured on 20 of 20 seeds against our registered prediction of <0.5. What the interval tests is distinguishability rather than count: on the real facts, multi-valued runs have 6.4x its occupancy of the rest. It survives at f_0 in {0.1, 0.3, 0.5}, capture peaks at f_0 = 0.5, and of four interventions with criteria frozen first, timing dominates fraction at matched budget while a consistency gate drives every model to 0.993. The resampling unit is the seed, at a design effect of 3.75 on a pooled level: under a 44-seed control the ordering supporting claim 4 collapses from Spearman +0.98 at three seeds to +0.31-0.80 at forty-four, while claim 2's ordering is exact there (+1.00, p = 0.017). All 87 graded rows are in Appendix W, 37 of them graded withdrawn, failed, self-correcting, undecidable or an acknowledged limit, against 50 that are not."

VACS: Value-Aligned Compositional Shielding for Multi-Agent Reasoning

Yiyao Zhang, Diksha Goel, Hussain Ahmad, Shixun Huang, Jun Shen — Wed, 23 Sep 2026 00:00:00 +0000

86% · reliability 86%, capability 44%, Agents 44%

Multi-agent reasoning systems in high-stakes domains must be both accurate and safe, yet agents often follow heterogeneous value priorities (e.g., rigor, conciseness, safety), causing conflicting recommendations. Existing methods do not jointly provide: (i) principled inference of each agent's implicit values from behavior, (ii) compositional formal safety guarantees without full online communication, and (iii) value-aware conflict resolution with faithful explanations. We present VACS (Value-Aligned Compositional Shielding), a four-layer framework addressing all three. Layer 1 learns value-dimension rewards from pairwise preferences using Bradley-Terry modeling and infers per-agent value weights via deep MaxEnt IRL. Layer 2 encodes value constraints in a Lean-inspired DSL and synthesizes compositional assume-guarantee shields for runtime safety. Layer 3 resolves disagreement through nucleolus-based credit allocation and Hamiltonian consensus optimization under long-term value constraints. Layer 4 extracts a critical reasoning path from co-state sensitivities and generates formally grounded natural-language explanations. Our contribution is primarily a unified systems design with formalized interfaces and operational guarantees at the verifier-constrained decision level, rather than a complete end-to-end formal proof of all language-model internals. In controlled proof-of-concept evaluations with role-conditioned agent panels on NEJM-AI QA, MathInstruct-Subset, and a cybersecurity incident-response benchmark (CyberSec-Eval), VACS outperforms strong baselines in accuracy (85.4%, 95.0%, and 90.0%) while reducing logical inconsistency rates to near zero.

Damage Predicts Recovery: When Calibration Data Matters in Compressing Financial LLMs

Junyi Ye, Mengjia Yu, Debapriya Hazra, Guiling Wang — Wed, 23 Sep 2026 00:00:00 +0000

86% · Efficiency 86%, reliability 75%, Evaluation 43%

Post-training quantization and pruning rely on a small calibration corpus. Whether specialized domains such as finance require domain-matched calibration data remains unsettled. We argue that the answer depends on the task-level damage caused by compression rather than on domain mismatch. If compression preserves the target capability, changing the calibration corpus has little effect. If compression causes large losses, task-formatted calibration can recover part of the loss. We test this hypothesis across two model families, six compression configurations, three token-matched calibration corpora, and ten financial classification and numerical question-answering tasks. The results support this hypothesis. Quantization largely preserves task performance, and calibration choice has little effect in this case. Pruning reduces numerical QA accuracy by over 40 points. In these damaged settings, another generic corpus does not help, while FinMix, a mixture of financial task examples, recovers a large part of the loss. The link between damage and recovery holds across model families and scales. These findings support a practical rule. Measure task-specific compression damage first, and construct specialized calibration data only when the damage is large.

Agensh: Scaling Organizational Intelligence to 1,024 Agents

Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia — Wed, 23 Sep 2026 00:00:00 +0000

86% · Agents 86%, Efficiency 79%, Evaluation 34%

A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator's capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.

A Computational Approach to Measuring Semantic Change in Sanskrit Literature

Tanay Agrawal — Wed, 23 Sep 2026 00:00:00 +0000

85% · Evaluation 85%

Diachronic word embeddings have become the modern standard for tracking semantic change, yet they have been largely validated on modern, high-resource, and well-segmented languages. This paper tests whether the paradigm transfers to Sanskrit, an ancient, low-resource language whose phonological fusion (sandhi), morphological inflection, compounding, and polysemy pose a unique challenge. I assemble a 2.7M-token corpus spanning four canonical periods, recover word boundaries with a neural byte-level sandhi splitter and lemmatizer, and train per-period embeddings across configurations. To evaluate the system, I curate a validation set from historical scholarship and test recovery directionally with anchor displacement. Of 21 testable shifts, 19 move in the philologically attested direction (sign test, p=0.00011). I further show which configuration the language forces and comment on opportunities for improvement.

Efficient Cost-Aware LLM Evaluation via Bayesian Bandit Gittins Indices

Qian Xie, Yueli He, Nairen Cao — Wed, 23 Sep 2026 00:00:00 +0000

85% · Efficiency 85%, capability 61%, Evaluation 41%

Exhaustively evaluating every candidate LLM configuration on every benchmark item to identify a high-performing one is costly. We formulate configuration selection as a cost-aware Bayesian bandit problem and propose GittinsEval, which draws on the Bayesian-optimal Gittins policy to determine which configuration to evaluate next and when to stop. We extend the policy with an anytime recommendation rule over both fully and partially evaluated configurations, using an LCB-style score to account for posterior uncertainty. GittinsEval is computationally efficient, requiring only lightweight online updates after offline precomputation. Across GSM8K, PIQA, AlpacaEval, and MMLU response matrices, GittinsEval is consistently competitive, with particularly strong gains over configuration-level Bayesian optimization on large-example benchmarks and over cost-unaware bandit baselines on large-candidate tasks. Crucially, GittinsEval often attains near-zero simple regret using only 1% to 2% of the exhaustive-evaluation cost; it also offers an adaptive stopping rule that typically triggers at 1% to 10%.

Toolcompass: Guiding Tool Trialing, Not Suppressing It

Junlin Fang, Chong Zhang, Do Nguyen-Thanh, Xiaogang Xu, Zhen Fang, Sean Du — Wed, 23 Sep 2026 00:00:00 +0000

85% · Agents 85%, capability 64%, Efficiency 32%

Large language model (LLM) agents must generalize from tools seen during training to unseen tools at deployment. A key challenge is tool trialing, i.e., excessive trials waste the interaction budget, whereas selective trials enable exploration of unfamiliar tools. Existing outcome-based post-training leaves wasteful trials unguided, while turn-level supervision may suppress necessary exploration. We introduce ToolCompass, a post-training framework that guides tool trialing by organizing tool-call representations according to shared functions. Specifically, ToolCompass models each function class as a von Mises--Fisher distribution and jointly reduces intra-function variation across domains and increases inter-function separation. This structure transfers experience from seen tools to functionally similar unseen tools, directing exploration away from unrelated alternatives. ToolCompass requires no ground-truth call traces or unseen-tool access and incurs no inference overhead. Experiments on AppWorld and FTRL show consistent gains across GRPO, RFT, and DMPO. improves AppWorld OOD task success by up to 10.71 percentage points over vanilla post-training and performs best among competitive baselines on both benchmarks.

How Strongly Should Task State Influence an LLM Agent?

Chenyu Zhang, Wonbin Kweon, Jiawei Han — Wed, 23 Sep 2026 00:00:00 +0000

85% · Agents 85%, reliability 80%, Evaluation 78%

Long-horizon assigned work requires an LLM agent to track the state of a task: which steps are done, blocked, cancelled, or open to repetition. Agent systems either keep this state as text in the prompt and rely on the model to read that text, or move the state into a module that enforces it, and each system is evaluated as a whole, so no one knows how much reliability comes from the state being shown, told, or enforced. We fix the task rules, the model, and paired episodes and vary how strongly task state reaches the agent: a raw transcript, an exact checklist, per-turn directives from a state machine compiled from the brief and advanced only by execution receipts, or an enforcement gate on that machine that refuses state-violating actions; every episode is scored by exact payload matching against dynamic ground truth. Across three models, two reasoning regimes, and two domains, four findings hold without per-turn reasoning: displaying accurate state is unreliable, an unverified ledger the agent writes itself beats an accurate checklist it is shown, directives help in proportion to the model's obedience, and enforcement needs no obedience but is bounded by the correctness of its state and by the matcher that maps requests to steps; per-turn reasoning at a 235B agent compresses these separations without repairing the text rungs. The same gate, compiled from $\tau^2$-bench's airline policy, raises a 235B agent's pass$^1$ from 0.39 to 0.54 and changes nothing for a 35B agent that rarely violates the policy; on PM-Bench, where acting turns on recognizing a cue rather than on state, showing the record is the best rung--matching or beating both gates and reversing the ledger-over-checklist finding--and enforcing the matcher's judgement drops a 35B agent below its raw transcript. Enforcement pays when failures are state-decidable and frequent, and hurts when the gate's judgement is wrong.

Auditing Proxy-Based Validation Across Text Spans

Daein Weon, Dong Ho Kang — Wed, 23 Sep 2026 00:00:00 +0000

85% · Evaluation 85%, reliability 70%

Evaluation scores are often validated by their agreement with inexpensive proxy labels. When the score and the proxy are computed from the same text span, however, that agreement can arise from surface evidence the two share rather than from the semantic construct the proxy is meant to represent. We make the distinction explicit by declaring the score, its span, the proxy and the target construct as a validation contract, then re-evaluating that proxy rule strictly outside the scored span. In a controlled HotpotQA correctness experiment varying only the shared text boundary, the score agrees with its proxy far better than with correctness at a 50-character prefix: the gap is +0.184, collapsing to at most +0.045 from 120 characters onward. At that short prefix the score still predicts whether the answer string appears later (AUC 0.634) while an equivalence test places its agreement with correctness at chance, so the reported proxy agreement does not establish that the score ranks correctness. On OR-Bench, suppressing each model's recurring opening templates removes most of the score's association with the refusal proxy, while matched-volume deletion removes almost none and construct agreement stays at chance. Only three of eleven external contracts support the off-span control, and none of the routing studies we sampled released the generations it needs. We therefore ask that a proxy-based validation claim declare the span each label is read from, report the construct agreement beside the proxy agreement, and release the generations that let the proxy be re-read off the scored span.

ClusterFewshot: Improving Few-shot Optimization for LLMs workflow

Omri Bar Haim, Shahar Katz, Lior Wolf — Wed, 23 Sep 2026 00:00:00 +0000

85% · Efficiency 85%, capability 55%

The performance of large language model (LLM) workflows often depends on selecting a small set of in-context demonstrations to guide model behavior on new tasks. Recent methods improve this process by augmenting prompts with successful reasoning paths. However, their demonstration selection relies on random sampling or metric-based rankings, overlooking the semantic structure of the task. We propose ClusterFewshot, a strategy that combines semantic structuring with utility-aware scoring to construct representative and effective few-shot demonstration sets. Evaluated within DSPy-based pipelines, ClusterFewshot substantially reduces optimization cost across multiple benchmarks, while consistently improving accuracy relative to prior bootstrap-based methods in both standalone prompt tuning and hybrid prompt-weight optimization.

The Free-Recipe Limit: Every Recipe Effect Measures Which Premise of an Idealised Learner Broke

Wenhui Chen, Jianlin Chen, Ziyao Lin, Chi Man Vong — Wed, 23 Sep 2026 00:00:00 +0000

85% · Evaluation 85%, reliability 51%, Agents 33%

Fix a corpus and send recipe search to infinity: try every order of the skills, every arrangement from blocked to interleaved, every composition, and keep the best. Two quantities decide what that search was worth: the diameter of the reachable set it explores, and the resolution at which anyone can tell two endpoints apart. Where the diameter falls below the resolution, no amount of search converts into a decision, and the signature is not an absence of winners but winners that do not survive re-running. We measure this recipe-search wall with 761 fine-tuning runs on 12 base models (0.5B-14B, three pretraining families) over competition-mathematics skills: base checkpoints, supervised fine-tuning under AdamW, exact-match scoring at k=4. Within one coherent domain at fixed volume the three classical freedoms average 0.010-0.021 against a 0.019 floor, and the largest contrast, 0.0619, clears a three-seed resolution and then reads +0.010 and -0.015 on two reruns. The departure with a systematic answer is coherence: halving one pooled corpus and letting the halves write answers under incompatible but equally correct conventions moves arrangement from capability to allocation between conventions, by two orders of magnitude over a same-convention control, and writing the convention into the input switches the phenomenon off. The switch replicates on a second pretraining family and survives an independent re-execution of its own protocol, with a re-execution spread (0.087) smaller than the resolution a search-selected order cell carries (0.144). Order itself is a transient whose sign crosses zero three times inside a single run. Volume, the one lever nobody calls a recipe, is the one that reliably pays. A public scorecard grades all 26 pre-registered claims: 18 supported, 5 failed, 2 untested, 1 mixed.

Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It

Yu Sun, Junhao Xu — Wed, 23 Sep 2026 00:00:00 +0000

85% · reliability 85%, Evaluation 80%

Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.

The AI Neuroscientist: An Interactive Agentic Interface for Neuroimaging Analysis

Aakash Patel, Panos Ketonis, Shreya Saxena, Smita Krishnaswamy, David van Dijk — Wed, 23 Sep 2026 00:00:00 +0000

84% · Agents 84%, Evaluation 73%, reliability 43%

Analyzing neuroimaging data requires specialized coding and statistical expertise, which limits accessibility for researchers without computational backgrounds. We present the AI Neuroscientist, a language agent for interactive data exploration. The system integrates a large language model (LLM) with a neuroimaging toolset to perform quality control, modeling, and visualization. This allows researchers to query data quality and specify analysis parameters directly in natural language, providing a transparent and interactive alternative to conventional scripted pipelines for small-scale data exploration. We demonstrate these capabilities using functional near-infrared spectroscopy (fNIRS) data, and evaluate the agent on a custom fNIRS benchmarking suite against general-purpose LLM agents with code sandboxes. Future extensions will generalize the architecture to additional modalities, including functional magnetic resonance imaging (fMRI) data, and expand the benchmarking suite to additional fNIRS tasks.

SMTB: Fast Structure-Mapping with Tight Bounds

Daniel Weitekamp, Christopher MacLellan — Wed, 23 Sep 2026 00:00:00 +0000

84% · capability 84%

Structure-mapping forms analogies by aligning systems of relationally connected elements based on shared structure instead of surface features. We introduce a new structure-mapping algorithm: Structure-Mapping with Tight Bounds (SMTB) that is 5--15x faster than the structure-mapping engine (SME) and about 50\% better at finding mappings in large nested domains. SMTB is part of the broader Cognitive Rule Engine (CRE) project, a flexible multi-language-compatible framework with an accessible Python interface to state-of-the-art C++ implementations of core algorithms commonly used in cognitive systems such as pattern matching, planning, and structure-mapping. CRE and SMTB are designed to work with a wide range of representation choices. Unlike SME, which biases higher-order correspondences in tree-like predicate logic, SMTB maximizes relational connectivity without privileging higher-order relations. This allows SMTB to work just as well over arbitrary relational graphs as it does in tree-like domains of nested predicate logic. We discuss situations where privileging "higher-orderness" in structure-mapping can cause issues, and illustrate how SMTB avoids failure modes that SME would encounter in these situations. We also provide an evaluation comparing SMTB to SME v4 over 5845 domain pairs from the SME corpus.

Recovering Agentic Sovereignty: Mitigating the Consensus Paradox via Contrastive Epistemic Decoding

Dahlia Shehata, Ming Li — Wed, 23 Sep 2026 00:00:00 +0000

84% · reliability 84%, Agents 57%, capability 41%

Large language models (LLMs) exhibit a parametric vulnerability to adversarial swarm consensus. To mitigate this sycophancy, we introduce Contrastive Epistemic Decoding (CED), a zero-shot inference intervention. Unlike standard Contrastive Decoding (CD) which relies on a weaker secondary model, CED utilizes a dual forward-pass on a single architecture to isolate conformity bias. By introducing a novel asymmetric, zero-bounded probability clamp and discrete top-k truncation mask, CED mathematically suppresses toxic consensus tokens without causing grammatical collapse. Evaluated across 7,200 paired trajectories on complex benchmarks (GAIA, SWE-bench, Multi-Challenge) using Gemma-2 (9B), Llama-3.1 (8B), and Mistral v0.3 (7B), CED successfully neutralizes architectural and positional biases. By reducing cognitive loafing by up to 33.00% absolute, CED drives significant performance gains, yielding up to a 30.75% accuracy recovery. Regaining sovereignty induces distinct architectural behaviors---passive task-focus in Gemma-2 and active refutation of the simulated swarm in Llama-3.1---showing CED decouples compliance from capability without fine-tuning.

Syndrome, Synergy, and Safety: Structured Reasoning and Knowledge-Driven Alignment for TCM Prescription Generation

Zheng Chen, ZhiCheng Du, Haoxuan Li, Peiwu Qin — Wed, 23 Sep 2026 00:00:00 +0000

84% · reliability 84%, capability 57%, Evaluation 34%

Applying large language models to Traditional Chinese Medicine (TCM) prescription generation reveals three clinically critical gaps: models produce end-to-end mappings without auditable reasoning following the li-fa-fang-yao paradigm (SR Gap), treat each encounter in isolation without follow-up adjustment via sui zheng jia jian (LA Gap), and fail to enforce absolute contraindication rules such as Shi Ba Fan (SC Gap). We propose a progressive four-stage framework (SFT $\to$ PG-CoT $\to$ Dynamic $\to$ K-RL) that addresses each gap: PG-CoT constrains CoT distillation under the li-fa-fang-yao paradigm to produce auditable diagnostic chains, Dynamic SFT models patient trajectories with explicit transition reasoning, and K-RL encodes deterministic pharmacological rules as rule-based DPO preference signals. Across 12 fine-tuned models and 6 zero-shot baselines, our framework substantially improves prescription quality over zero-shot baselines---with a 7B model (Mistral-7B) surpassing zero-shot GPT-5 on all three TCM evaluation metrics.

Unanimity Without Persuasion: A Single Round of Debate Erases the Disagreement That Verification Needs

Yang Shu — Wed, 23 Sep 2026 00:00:00 +0000

84% · reliability 84%, Evaluation 72%, Agents 40%

A debate panel can become unanimous without becoming more correct. This is dangerous for downstream safeguards: a substituted verification ballot can change only narrow-margin votes, while richer arbiters lose disagreement as a natural targeting signal. We show that one debate round can erase that resource without requiring persuasion. Tracking a heterogeneous 7-judge panel through a blind round and three debate rounds on 600 code-correctness candidates, unanimity on a fixed cohort jumps from 39.5\% to 95.2\% in round 1 (93.1\% of the total collapse), while accuracy moves by less than one point and 96.3\% of verdict flips follow the displayed peer majority. An execution-based verification ballot corrects 8 of 2,037 pre-debate candidate-substitution instances but changes zero in every later round; by round 3 every wrong decision is unanimous, erasing dissent that had flagged two-thirds of the panel's errors. Identical-cohort controls explain why: no-peer reconsideration reproduces 79.6\% of the collapse, real labels without reasoning reproduce 91.5\%, and random labels steer flips toward whatever they display; the full-debate condition adds 4.3 percentage points over labels only (clustered 95\% CI 0.7--8.1). The one-round collapse reproduces in two additional real runs and two fake-label seeds, remains under panel sizes 3--7, and appears in MATH-500. Parse failures concentrate on contested candidates ($p<0.001$), making attrition non-ignorable. The design implication is operational: verify before any second-pass evaluation or peer exposure, and never treat post-debate unanimity as independent evidence of reliability.

Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models

Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li, Johannes Bjerva — Wed, 23 Sep 2026 00:00:00 +0000

84% · Evaluation 84%, Efficiency 50%, Agents 47%

The rapid capability gains of frontier language models are widely attributed to improved reasoning abilities, yet this cannot be verified as raw CoT traces in closed-source systems are hidden. By registering a simple custom tool through a standard API feature, we induce frontier models to externalize intermediate reasoning. Because these traces may reflect post-hoc rationalization rather than genuine reasoning, we first evaluate against native CoT on open-source models and extend to closed-source frontier models including GPT-6 Astra. We find that the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines, across competition mathematics, science, and code generation. We then characterize how frontier models structure their intermediate reasoning. Across token efficiency, reasoning-step types, and induced reasoning trees, we identify systematic differences in how models externalize, compress, and organize reasoning. We find that Astra exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning. These findings provide a behavioral lens on frontier-model reasoning beyond benchmark scores.

Mining Legal Arguments in U.S. Corporate Case Law

Wed, 23 Sep 2026 00:00:00 +0000

83% · Evaluation 83%

Legal argument mining supports passage classification, retrieval, and argument completion. This work introduces an expert-annotated dataset of 42 U.S. federal tax opinions on corporate reorganizations under I.R.C. {\S}368. To our knowledge, it is the first expert-annotated, tree-structured argument corpus for this domain. Explicit spans receive one of five functional labels: Rule, Analysis, Conclusion, Background Facts, and Procedural History. Rule, Analysis, and Conclusion spans can be linked into directed support trees, while Background Facts and Procedural History serve a contextual function. The corpus provides span-based, sentence-based, flat, and tree-structured representations. Agreement analysis shows that functional node labels are more reliable than directed support edges and implicit intermediate conclusions. Directed-path agreement is stronger than direct-edge agreement, which indicates that broad reachability is more stable than exact local decomposition. Classification experiments show that functional labels are learnable under case-disjoint evaluation. Retrieval experiments show that supervised fine-tuning improves within-case retrieval. However, cross-case generalization remains weak. The dataset supports legal passage classification and provides a conservative benchmark for structured argument mining in U.S. federal tax case law.

The Architect, the Adversary, and the Judge: Closed-Loop Generation of Standards-Aligned Assessment Items at Scale

Wenhui Chen, Ziyao Lin, Jianlin Chen, Peiji Long, Chi Man Vong — Wed, 23 Sep 2026 00:00:00 +0000

83% · Evaluation 83%, reliability 83%, capability 34%

We present CLAIM, a production pipeline for K-12 assessment-item generation coupling a two-stage generate-then-attack protocol (the model drafts as a "curriculum architect", then re-enters the same conversation as a hostile adversarial reviewer), bi-directional few-shot conditioning on accepted and rejected items, the latter carrying the evaluator's diagnosis, and a knowledge dictionary of 44,844 error-correction rules mined from that feedback and retrieved per standard and item type. Across 43,227 scored items over 755 Common Core ELA standards, three item types, and ten LLMs, the pipeline reaches a 97.8% expert-evaluator pass rate on a 9,074-item production run. We then ask what that rate certifies. Re-scoring a stratified sample with three judges from other vendors, blind to the deployed verdict, reproduces the format ordering under every judge and recovers a larger open-set deficit than the deployed evaluator does; but agreement on the accept/reject binary is weak at production prevalence (kappa about 0.13), and the judges agree with each other no better. The level is therefore judge-relative, and with no student-response data our quality evidence is evaluator-judged throughout. The corpus also exposes a robust asymmetry. Multiple-choice and multiple-select generation saturate at 98% or above for both frontier models under a dozen static rules, whereas fill-in-the-blank generation is capability-tiered (82.8-96.7% across five models under a matched rule set, standards, and judge) and plateaus under prompt-only optimization, with error mass shifting between answer-key over-inclusion and omission as rules accumulate. We analyze this as open-set boundary determination, a task autoregressive decoders are structurally ill-equipped to solve, and show the asymmetry recurring when the evaluator itself is distilled: fail-recall rises from 8% to 63% while F1 saturates at 0.25.

Certified Mechanistic Interpretability: Lifting Single-Input Findings to Bounded Neighbourhoods

Zhen Zhang, Yanliang Huang, Peng Xie, Wenyuan Wu, Amr Alanwar — Wed, 23 Sep 2026 00:00:00 +0000

83% · reliability 83%, Evaluation 31%

Mechanistic interpretability reverse-engineers transformer circuits one input at a time, leaving observed mechanisms without guarantees over bounded input neighbourhoods. We address this gap with a framework based on constrained polynomial-zonotope (CPZ) propagation that lifts mechanistic-interpretability observations from a single input to certified statements over a bounded set of perturbations. Three internal-attention queries (top-$k$ stability, evidence mass, and attention entropy) are formulated as tractable programs over the simplex of attention weights, and CPZ propagation through transformer blocks is shown to preserve the softmax simplex and the LayerNorm zero-mean identity exactly. A recursive Jacobian zonotope construction extends the same certificates across layer depth by linearising the block stack at the input and avoids per-layer generator growth. We instantiate the framework on transformer attention; the resulting certificates offer a way to sharpen mechanistic statements that single-input inspection cannot resolve on its own, and to inform downstream decisions in regimes where empirical heuristics may be misleading.

Recursive self-improvement of AI research agents

Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, Zhengyao Jiang — Wed, 23 Sep 2026 00:00:00 +0000

83% · Agents 83%, Efficiency 83%, reliability 45%

AI agents are beginning to automate research and development across the AI stack, from improving training efficiency to optimizing inference. A natural next step is to improve the research efficiency of the agents themselves. When an AI research agent's own code is the object of optimization, each accepted rewrite becomes the agent that the next round edits. We refer to this loop as recursive self-improvement. Its significance lies in a long-standing trend, in which increased cumulative spending on R&D yields diminishing returns. Sustained self-improvement offers a way to counter this trend. We present AIDE^2, a system that implements this loop for a frontier AI research agent. It proposes changes to its own code, benchmarks modified versions of itself on a suite of AI R&D tasks, and keeps the changes that perform best on hidden evaluations. In an autonomous 8-day run, AIDE^2 discovered seven successive improvements, ranging from a new search policy to memory mechanisms that compress and manage the agent's growing context. These gains generalize to four held-out benchmarks spanning machine learning engineering, heuristic algorithm engineering, and physics-based weather forecasting, the last of which is out of distribution from the selection tasks. On all four, the strongest discovered agent matches or exceeds a human-engineered production research agent that ranks among the strongest on FML-Bench. On a separate held-out task family, the discovered agents also exhibit reduced reward hacking, a property the loop never explicitly optimized for: the rate falls from 55% to 32% during the run, 7 percentage points below the human-engineered agent. Together, these results show that an AI research agent can improve its own research efficiency through recursive self-improvement, and that these gains transfer to tasks and domains the loop never encountered.

Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference

Wed, 23 Sep 2026 00:00:00 +0000

83% · reliability 83%, Efficiency 68%, Evaluation 55%

Greedy decoding from large language models is commonly treated as deterministic. We show it is not precision-invariant: the same model, prompt, and decoding algorithm produce different outputs in BF16 versus FP16 on identical hardware. Across our evaluations of six models (1.1B-7B parameters, four families; divergence additionally characterised at 12B) and three benchmarks, 49-100\% of prompts diverge; a single token flip often cascades into trajectory-level divergence. We develop an empirical error-propagation analysis and find that 22 layers of accumulated body error do not distinguish flipping from non-flipping steps; the outcome depends primarily on the top-two logit margin at the LM head relative to the directional perturbation between the top-two candidates. The analysis makes five testable predictions about intervention outcomes, including that applying more FP32 compute (broader scope) makes agreement worse. The experiments match all five predictions. The best-performing low-overhead intervention we evaluate, selective FP32 LM head recomputation, triggered only when the margin falls below a threshold, delivers +22-36 pp exact agreement on A10G (+12-21 pp on L4 and A100) at less than 4\% latency overhead in low-batch (batch size <=4) single-stream inference. We map the applicability boundary across six models and four batch sizes, and hypothesise that training-time precision stability is a determining factor. The method is a partial mitigation rather than a universal determinism guarantee: its benefit vanishes when body-originated error dominates, including at batch size >=8 and under end-to-end FP8 in our tests.

ChatT2: An Adaptive Framework for Developing a Large Language Model-Based Agent for Natural Product Domain Research

Yihan Wang, Qiandi Gao, Yihui Zhuang, Liangjun Ge, Heqian Zhang, Jiaquan Huang — Wed, 23 Sep 2026 00:00:00 +0000

82% · Agents 82%, reliability 55%

Scientific investigations into microbial natural products (NPs) present significant challenges for novices, largely due to the complexity of microbial systems, biochemical diversity, technical skill requirements, and the demands of bioinformatics and data analysis processes. To address these issues, we introduce ChatT2, a large language model (LLM)-based agent that is specifically tailored to the unique characteristics of bacterial type II polyketides. These polyketides form a structurally distinct and therapeutically important NP family. ChatT2 was developed within an autonomous multiagent framework composed of a mentor, an executor, and an evaluator, each with defined responsibilities. The mentor acts as an intermediary between ChatT2 and the user, utilizing chain-of-thought prompting to refine the intent of the user. Under the guidance of the mentor, the executor synthesizes multimodal information via retrieval-augmented generation techniques and seamlessly integrates bioinformatics and cheminformatics tools. The evaluator ultimately assesses the output of the executor to ensure the richness and accuracy of the retrieved information. Our research highlights how ChatT2, designed with this multiagent framework, addresses the challenges faced by general LLMs in terms of understanding limited, specialized corpora and complex biological information and provides both experts and novices with a valuable tool for exploring various NPs of interest. The ChatT2 webserver can be accessed at https://chatt2.site/#/chat.

Seeing Is Not Perceiving: When Synthetic Consumers Can and Cannot Pretest Visual Marketing

Yi-Lin Tsai (Arvin), Yung-Hsiu (Arvin), Lai — Wed, 23 Sep 2026 00:00:00 +0000

82% · reliability 82%, Evaluation 78%, Efficiency 34%

Marketers now deploy generative AI agents as synthetic consumers to pretest visual assets such as logos, packaging, and advertising at a fraction of human-panel cost. However, this procedure assumes that a model seeing a visual cue can also perceive its consumer meaning, which is largely untested. We stress-test the assumption using six canonical visual marketing experiments, varying the two levers managers control: model generation (GPT-4o-mini vs. GPT-5.4-mini) and input format (plain text vs. JSON). Every resulting configuration passed the manipulation checks; however, none of the configurations reproduced more than two of the six human effects, and the remainder were nonsignificant. The one exception was a significant reversal of the human pattern. Providing conceptual or empirical evidence through in-context learning steers average responses toward the human effect. Yet steering has a limit: even when it succeeds, a configuration reproduces less than half of the natural spread of human responses and so understates consumer heterogeneity. We integrate these results into an AI governance protocol (Calibrate, Intervene, Deploy) that delineates when synthetic consumers can responsibly screen creatives and when human panels remain necessary.

When Verifiers Vote Backwards under Verdict Substitution: Signed Pivotal Value in Correlated Self-Consistency

Yang Shu — Wed, 23 Sep 2026 00:00:00 +0000

82% · Evaluation 82%, reliability 78%, Agents 31%

Replacing one ballot can change a majority decision only on queries decided by a single vote; this structural fact requires no independence assumption. We study the sign of that change using a labeled, verdict-style intervention: one correctness signal replaces one correctness-indicator ballot in $k{=}7$ self-consistency panels. This diagnostic intervention is not identical to deployed answer-identity plurality. A primary MATH-500 experiment ($n{=}570$) gives a different-model verifier a $+24.2$pp pivotal gain, whereas a role-reversed configuration gives $-11.2$pp; an exploratory code stress test (14 pivotal rows across 9 tasks) gives $-24.5$pp. An exact signed-gain decomposition accounts for all observed signs through the verifier's state-specific accuracy and the composition of the two one-vote tally states, rather than global accuracy or model provenance. Same-source signals lose accuracy on the pivotal stratum (65$\to$44\% in the primary configuration), while error correlations provide a descriptive error-association diagnostic. Controlled degradation and a $k\in\{3,5,7\}$ subset sensitivity analysis probe the stability of the observed pattern around this accounting. Under the evaluated ties-incorrect answer-identity plurality analysis, the structural zero and strong-verifier benefit persist, but the role-reversed harm attenuates to $-0.9$pp and is not significant. The results therefore establish harmful verdict substitution, not universally harmful deployed plurality, and motivate a testable but unverified hypothesis for negative process-reward-model weights.

Component Type, Not Reconstruction Error, Predicts Attention Quantization Sensitivity

Kasun Dewage, Marianna Pensky, Suranadi De Silva — Wed, 23 Sep 2026 00:00:00 +0000

82% · Evaluation 82%, Efficiency 76%, reliability 30%

Many post-training quantization (PTQ) methods use layer-wise reconstruction, second-order proxy objectives, or activation-aware transformations to reduce quantization-induced error. Whether that error signal predicts the downstream functional impact of quantizing an individual attention projection has not been directly characterized. We sweep nine open-weight language models (1.3B--8B parameters; OPT, GPT-J, LLaMA-1/2/3, Mistral, Qwen 2.5) and quantize one attention projection at a time under round-to-nearest (RTN) and, for seven models, GPTQ at 3 and 4 bits, recording reconstruction error, perplexity change, and per-projection activation-weighted quantization error for 3,808 distinct measurements. We find: (1) within a given component type (Q, K, V, or O), reconstruction error explains less than 10% of the variance in perplexity sensitivity in 27 of 36 cases under RTN, with median R^2 = 0.044; (2) both component type and layer identity explain more variance than reconstruction error in all 9 models, with layer identity the strongest predictor in 7 of 9 models and component type strongest in the remaining 2; (3) value (V) projections are the most commonly dominant component, accounting for 38--51% of total positive Delta PPL in seven of nine models; (4) the dominant component is broadly preserved between RTN and GPTQ (5 of 7 cases); and (5) activation-weighted quantization error is a moderately better within-component predictor than reconstruction error for V projections specifically (median R^2 of 0.20 vs. 0.06). These findings indicate that relative weight reconstruction error alone is insufficient for sensitivity-aware bit allocation, and that V projections merit dedicated consideration in mixed-precision schemes.

PACT: From Credit Assignment to Critic Alignment

Jiayan Fu, Hang Xu, Yong Zhang, Zhaokai Luo, Yao Hu, Dongyan Zhao — Wed, 23 Sep 2026 00:00:00 +0000

82% · capability 82%, Agents 62%

Reinforcement learning has become a central component of large language model (LLM) post-training, yet token-level credit lacks a generally accepted mathematical definition, leaving its relationship to commonly used training signals unclear. We formulate three regularity conditions, namely Completeness, Prefix Consistency, and Neutrality, and prove that they uniquely determine token-level credit. This characterization provides a unified basis for explaining phenomena across existing algorithms and guides the development of an improved actor-critic training procedure. Through this lens, an ideal teacher in On-Policy Distillation (OPD) acts as an implicit critic, yielding an expected policy gradient proportional to that induced by token-level credit. Response-level REINFORCE Leave-One-Out (RLOO) signals match the expected policy-gradient contribution of token-level credit despite their coarser granularity. We further establish approximate credit sparsity under bounded outcome rewards and show how intermediate critic errors in Generalized Advantage Estimation (GAE) can become comparable to the underlying credit. These motivate Policy Aligned Critic Training (PACT), which adopts an Actor-then-Critic update order to apply importance sampling correction to critic training and better align the critic with the updated policy. In agentic mathematical reasoning, PACT achieves 72.87% average accuracy across four benchmarks, outperforming GRPO and PPO by 8.80 and 13.16 percentage points, respectively. On SWE-bench Verified, PACT achieves a pass rate of 67.4%, outperforming PPO, GRPO, and SAO by 2.4, 2.0, and 3.8 percentage points, respectively.

Optimizing Denoising Trajectories in dLLMs: A Lightweight Evolutionary Heuristic Approach

Zijian Zhao, Dian Jin, Xialiang Tong, Sen Li, Mingxuan Yuan — Wed, 23 Sep 2026 00:00:00 +0000

81% · Efficiency 81%, capability 61%, reliability 56%

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to conventional Auto-Regressive (AR) Large Language Models (LLMs). By leveraging bidirectional attention and parallel decoding, dLLMs enable more efficient generation. However, they require a carefully designed denoising scheduler at inference time (absent during training) whose choice significantly impacts generation quality. While confidence-based heuristic schedulers have shown strong empirical performance, they suffer from two critical failure modes: EOS Overflow and Proximal Bias. Through in-depth analysis of the Transformer's attention patterns, we reveal that these failures stem from certain positions assigning disproportionately high attention weights to invalid tokens (e.g., [MASK] and [EOS]), which produce misleading confidence signals. Building on this insight, empirical evidence shows that valid attention scores can provide complementary guidance to conventional confidence-based heuristics, yet no single metric consistently excels across all scenarios, implying that the optimal denoising trajectory is highly context-dependent. To address this problem, we propose a lightweight evolutionary heuristic scheduler optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Our scheduler dynamically integrates multiple heuristic features with a contextual mean-field embedding, while requiring only 393 trainable parameters. Evaluated on LLaDA and Dream across four reasoning and planning benchmarks, our method consistently outperforms strong baselines, including conventional heuristics, block auto-regressive methods, and recent State-Of-The-Art (SOTA) approaches. To the best of our knowledge, it represents the most parameter-efficient neural scheduler to date. Our code is available at https://github.com/RS2002/Evo-Denoise .

Combining Hierarchical Cognitive Process with Process Supervision for Interpretable Scene Safety Understanding

Zhiyun Jiang, Hanyong Wang, Binbin Liang, Yu Xie, Zhengjie Wang, Menglong Yang — Wed, 23 Sep 2026 00:00:00 +0000

81% · reliability 81%, Evaluation 79%, capability 38%

Scene safety understanding plays a life-or-death role in situational awareness in various critical domains. Traditional methods that rely on learning direct mappings between scenes and safety levels often lack interpretability, limiting their reliability in critical applications. An effective approach to overcoming this challenge lies in interpreting human cognitive processes and equipping machine models with analogous cognitive capabilities. This work explores an effective way of integrating scene safety cognitive process modeling and process supervision. Specifically, we first construct a hierarchical cognitive safety structure, which motivates the development of a novel, high-quality scene safety understanding dataset based on multi-step reasoning with process labels. This dataset serves both as a benchmark and a resource to improve the safety reasoning capabilities of Large Language Models (LLMs), while also enabling a granular analysis of intermediate reasoning steps through information flow and saliency-based techniques. Building upon this foundation, we introduce a modular and flexible process supervision framework that reflects the hierarchical nature of human cognition. This framework leverages LLMs as the core architecture and incorporates Low-Rank Adaptation(LoRA) and Mixture-of-Experts (MoE) strategies to enable specialization and collaboration among expert modules, each tasked with specific sub-processes of the overall reasoning chain. Systematic experimental evaluations and analyses confirm that our framework exhibits superior interpretability and performance characteristics compared to traditional approaches.

Enriching Speech Emotion Representations with Conversational Context

Wed, 23 Sep 2026 00:00:00 +0000

81% · capability 81%, Evaluation 53%

Detecting emotions is necessary for building systems that can accurately and adaptively interact with humans. Speech Emotion Recognition (SER) has become an important research focus to develop intelligent spoken interfaces. However, most studies predict emotions at the utterance level, ignoring the conversational context, along with the emotional flow and speaker interactions it carries. In this paper, we introduce ACERT (Averaged Contextual Emotion Representation through Time), a module that integrates a flexible-length window of conversational context to better capture emotional evolution in spoken interactions. To evaluate the robustness of this method, we conducted experiments on datasets spanning diverse emotionally expressive styles and contexts. ACERT outperforms current state-of-the-art (SOTA) approaches on IEMOCAP, establishes the first context-aware benchmark on SAFE, and obtains strong results on MELD for unweighted, class-balanced metrics. Ablation studies show that ACERT's gains come from emotional and conversational continuity, rather than from speaker identity or acoustic conditions.

Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning

Wed, 23 Sep 2026 00:00:00 +0000

81% · capability 81%, Efficiency 80%, reliability 42%

Quantization-aware distillation (QAD) restores much of the short-form question-answering performance lost to sub-3-bit quantization, yet leaves mathematical and code reasoning substantially impaired. Long generations often degenerate into repetitive loops, exhausting the decoding budget without completing a solution. We trace this gap to quantization-amplified exposure bias: QAD trains on fixed corpus prefixes, while quantization-induced deviations compound along the model's own autoregressive trajectories. To address this mismatch, we introduce an on-policy distillation (OPD) stage that places teacher supervision where the quantized model actually goes. Starting from a QAD checkpoint, the student generates through the quantized forward path used at deployment and receives feedback from a frozen full-precision teacher on its own prefixes, combining dense token-level guidance with task-verifier rewards. Across four models at 2.79 and 1.88 effective bits, OPD raises average BF16 performance retention from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval while preserving short-form performance, with reasoning gains substantially exceeding those of continued teacher-forced QAD in matched-budget comparisons. By coupling QAD's stable low-bit initialization with OPD's on-policy reasoning recovery, our framework provides a comprehensive sub-3-bit solution that preserves broad capabilities while restoring long-form reasoning.

ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains

Wed, 23 Sep 2026 00:00:00 +0000

80% · Evaluation 80%, Agents 70%, capability 39%

In this report we present results of the ICDAR2026 Competition on Multimodal Reasoning over Documents in Multiple Domains. This competition aimed to advance research in document understanding through the task of Visual Question Answering (VQA). Building upon previous DocVQA benchmarks, this competition introduces challenging reasoning questions over a diverse collection of documents spanning eight domains, including business reports, scientific papers, slides, posters, maps, comics, infographics, and engineering drawings. The competition concluded with 20 valid submissions from 8 teams spanning zero-shot VLMs, OCR and parser-augmented pipelines, agentic retrieval systems, multi-agent ensembles, and fine-tuned multimodal models. The results show that the strongest systems move beyond single-pass prompting and instead rely on structured evidence extraction, retrieval, verification, and orchestration across multiple components.

ufakzeka-1: Building and Evaluating a 151M-Parameter Turkish Language Model from Scratch

Sait Furkan Teke (ufak AI) — Wed, 23 Sep 2026 00:00:00 +0000

80% · reliability 80%, Evaluation 75%, Efficiency 72%

We describe ufakzeka-1, a 151M-parameter (182M with embeddings) decoder-only Turkish language model pretrained from scratch on 13.5B tokens of openly licensed text and instruction-tuned for chat, at a total cost of about \$286 in cloud GPU, API and notebook time. The contribution is not the model's capability, which is what a model this size can be expected to have, but the record of building and measuring it: a Turkish byte-level tokenizer at 1.77 tokens per word, a three-stage pretraining schedule, a post-training mixture of openly licensed and generated data, and an evaluation battery of release gates, a rule-checked sweep of 5,508 conversations, judged conversations and hand tests, all with prompts held out from the training data, enforced by decontamination inside the data build and by a checked-in invariant script we run before each build. We report three findings that we believe transfer to other small-model efforts: a safety gate that had been "fixed" with training data written from its own questions read 64/64 while the honest figure was 34/64; training-seed variance was as large as the spread across every recipe we tried, so single-seed comparisons at this scale are uninformative; and data rounds repaired only what was absent from the data, while identity tracking over long context and multi-turn arithmetic did not move across any data change we tried, which we read as limits of the model size rather than gaps in the data, a reading the next, larger model will test. Weights, the data recipe, the evaluation code and the spend ledger are released under Apache-2.0.

Graph Domain Adaptation Does Not End with Representation Learning

Ziqian Liu, Yongxue Xu, Enze Zhang, Jiaqi Zhang, Hao Wang, Maolin Wang — Wed, 23 Sep 2026 00:00:00 +0000

80% · capability 80%

Graph domain adaptation (GDA) transfers knowledge from a labeled source graph to an unlabeled target graph under shifts in both node attributes and graph structure. Existing methods primarily adapt graph representations through propagation redesign, distribution alignment, or source-to-target transition modeling, but still rely on a single graph-propagating path for target prediction. This leaves open whether an adapted graph representation exhausts the predictive evidence available in the target domain, since the graph-aware expert and graph-free local expert may exhibit different failure modes under topological shifts. To address this limitation, we propose EviGDA, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert. The graph-aware expert performs message passing and entropy-aware marginal alignment, while the graph-free local expert learns solely from source node features and labels without graph propagation or target alignment. The two experts are optimized independently and combined only at inference through a task-level constant probability mixture, preserving complementary evidence without joint training, learned routing, or target pseudo-labels. Extensive experiments on ten datasets and 16 transfer tasks show that EviGDA outperforms state-of-the-art baselines.

Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction

Yan Zhang, Pei Fu, Daiqing Wu, Huawen Shen, Ruoceng Zhang, Shaojie Zhang — Wed, 23 Sep 2026 00:00:00 +0000

80% · Agents 80%, capability 61%

Graphical User Interface (GUI) Agents autonomously interact with software to fulfill user requests, where GUI navigation stands out as the most critical and challenging capability. Mastering this capability demands a complex synergy of step-wise decision-making, state-action alignment, and long-horizon planning. While directly mixing these corresponding navigation tasks seems intuitive to simultaneously acquire these skills, such a direct combination is severely bottlenecked by inconsistent optimization objectives and profound data heterogeneity. To overcome these barriers, we propose the MaP (stands for ``\textbf{M}asked Tr\textbf{a}jectory \textbf{P}rediction''), a unified framework that seamlessly harmonizes divergent GUI navigation tasks. By modeling multi-turn GUI interactions as a trajectory and defining training objectives through component masking and prediction, MaP shifts the optimization from task-specific marginal distributions to a consistent objective. Furthermore, to handle the data heterogeneity across multiple navigation tasks, we design a role-aware adapter learning module that dynamically routes each token to a specialized representation space. Extensive experiments on five representative GUI navigation benchmarks demonstrate that MaP effectively mitigates gradient conflicts and significantly outperforms the direct mixture training, establishing a robust paradigm for multi-task GUI navigation.

AgenticSizing: A Large Language Model-based Multi-Agent Framework for Analog Circuit Sizing

Wed, 23 Sep 2026 00:00:00 +0000

80% · Agents 80%, capability 47%

Analog circuit sizing remains a challenging and time-consuming task due to the large design space, strong performance trade-offs, and increasing circuit complexity in scaled technologies. Although recent large language model (LLM)-based methods show promise in improving sample efficiency and interpretability, existing approaches often lack explicit circuit-topology understanding and are mainly evaluated on relatively simple analog building blocks. This paper presents a multi-agent LLM-based framework for complex analog circuit sizing. The proposed framework first analyzes the circuit topology and decomposes the netlist into functional blocks and substructures. It also extracts lightweight design knowledge for reuse. Based on the extracted topology and knowledge, a planner coordinates multiple role-specialized sizing agents to update design variables and achieve global performance specifications. This workflow mimics the collaborative process of an expert analog design team and provides a structured, interpretable, and simulation-driven optimization procedure. The framework was validated on eight circuits, with the largest design containing up to 55 transistors and 60 sizing variables. Notably, for the LDO benchmark, the proposed method achieved a 60\% success rate with an average of 83 iterations, where classical optimizers failed to find feasible solutions. Further, ablation studies demonstrate that topology understanding, design-knowledge infusion, and agent specialization provide complementary benefits. The source code is available to support reproducibility.

MSA-CITE: A Co-Adapted LoRA Specialist Ecology for Fixed-Budget Small-Model Inference

Ruitong Li, Binjie Guo, Aisheng Mo, Guowei Su, Jie Li, Ru Zhang — Wed, 23 Sep 2026 00:00:00 +0000

80% · reliability 80%, Efficiency 43%

Compact language models are typically deployed by retaining a single post-training checkpoint and sampling it repeatedly. In this work, we challenge this practice by treating multiple discarded checkpoints as composable assets for deployment. Starting from a single Qwen3-4B backbone, we preserve four frozen LoRA branches, each derived from a different post-training trajectory. stead of drawing four generations from one branch, we allocate a fixed four-generation budget by sampling one completion from each branch. Our method, Multi-path Specialist Adaptation with Calibrated Inference-Time Evidence (MSA-CITE), processes the resulting portfolio by grouping terminal answers into equivalence classes, scoring each class via summed calibration-derived source priors, and selecting a representative under deterministic tie-breaking rules. The readout stage does not learn from evaluation results, nor does it introduce additional generations, verifiers, or reranking steps. On 200 held-out mathematics items, the four-path portfolio achieves 65.5% accuracy, compared with 62.0% for the strongest single-branch baseline. On a 100-item subject-disjoint shift, it attains 42.0% versus 40.0%. Under in-distribution conditions, the improvements over homogeneous SFT and Online-OPD repetition are robust; results against the strongest baseline and under shifted conditions are not conclusive. Our findings offer a narrow but concrete contribution: post-training branches, even without co-training, can be collectively beneficial for deployment.

Dual-Frontier: When Can an Agent Trust Its World Model?

Huatai Zhu, Qiang Chen, Ziqian Kou, Wenhao Li, Fei Wang, Yichao Cao — Wed, 23 Sep 2026 00:00:00 +0000

80% · Agents 80%, reliability 69%, capability 44%

Learned world models are becoming essential to general-purpose agents: by predicting action consequences, they support planning and decision-making while reducing reliance on costly trial and error. This reliance creates a fundamental ambiguity: when a world-model-guided decision fails, the trajectory alone may not reveal whether the agent's decision rule or the world model caused the loss. We formalize this failure-attribution problem as a counterfactual decomposition of return loss and prove that its components are not identifiable from passive interaction, even for finite-horizon planners. This obstruction motivates Dual-Frontier, a learning principle that admits a world-model-guided decision only when its predicted advantage exceeds a certified bound on decision-relevant world-model error; otherwise, evidence is allocated to world-model verification. Action-conditioned value bounds and a closed-loop extension guarantee non-decreasing return for admitted decisions. Calibrated gates and simultaneous confidence sequences support adaptive evidence reuse, with sufficient and necessary verification bounds. Controlled learned-model experiments validate the predicted failure modes and certification behavior, while cross-backbone tool-use benchmarks instantiate the same verify-then-promote rule in realistic agent world-model pipelines, consistently improving decision quality and reliability.

DeepFEAv2: Deep Learning for Transient Finite Element Analysis Beyond Structured Meshes

Georgios Triantafyllou, Panagiotis G. Kalozoumis, Dimitris K. Iakovidis — Wed, 23 Sep 2026 00:00:00 +0000

80% · capability 80%, Efficiency 35%

Finite Element Analysis (FEA) is widely used for transient mechanical simulations, but its high computational cost limits real-time and high-resolution applications. Deep learning surrogate models can reduce this cost; however, many existing approaches are restricted to steady-state prediction or cannot jointly predict Node- and Element-based Outputs (NEO) over time. The state-of-the-art DeepFEA framework has addressed these issues but remains limited to structured finite element (FE) meshes. To overcome this limitation, this study proposes DeepFEAv2, a deep learning surrogate framework that enables prediction of transient FEA simulations across different FE mesh topologies and element types. The main contributions of DeepFEAv2 are: (a) a module that uses the FE connectivity matrix to organize input features by element and arrange them into an input sequence guided by the mesh topology; (b) a novel neural network architecture designed to process the input sequence and jointly predict NEO over time; and (c) a FEA-informed optimization strategy for regularizing these NEO predictions. DeepFEAv2 was evaluated on structured and unstructured 3D linear elastic datasets, as well as on a pressure-driven aortic valve dataset. DeepFEAv2 achieved R^2 values up to 0.99 and normalized errors as low as 0.38%. Compared with DeepFEA, it achieved up to 38.0% relative increase in R^2 and up to 87.1% reduction in normalized error. DeepFEAv2 also performed inference up to three orders of magnitude faster than traditional FEA. These results demonstrate that DeepFEAv2 can efficiently model transient FEA simulations across increasingly complex FE settings, providing a scalable surrogate framework for transient FEA.

MAC-RRG: Iterative Multi-Agent Collaboration for X-ray Radiology Report Generation

Futian Wang, Yuhan Qiao, Xiao Wang, Dan Xu, Yuehang Li, Zhixiang Guo — Wed, 23 Sep 2026 00:00:00 +0000

79% · reliability 79%, capability 71%, Agents 37%

Despite the remarkable progress of LLM-based and knowledge graph-augmented Radiology Report Generation (RRG) methods, existing techniques still suffer from inherent defects. Conventional LLM-only models lack structured medical prior knowledge, resulting in frequent medical hallucinations and low diagnostic interpretability. Current knowledge graph-enhanced schemes adopt static one-round knowledge fusion with single-source knowledge, incapable of dynamic knowledge updating according to generation feedback. This paper proposes a novel Multi-Agent Collaborative iterative framework for X-ray Radiology Report Generation, termed MAC-RRG. Inspired by multi-agent technology, our framework constructs a closed-loop optimization paradigm based on task decoupling and collaborative reasoning. Specifically, the framework first generates a preliminary radiology report from input X-ray images via a vision encoder and a basic LLM. Subsequently, a multimodal knowledge graph (MM-KG) agent mines structured disease correlation and anatomical knowledge from medical knowledge graphs, while an auxiliary knowledge agent extracts unstructured domain knowledge from public medical databases. The multi-source knowledge acquired by dual agents is fused and embedded to guide the LLM in iteratively refining the initial report. Extensive quantitative and qualitative experiments on mainstream X-ray RRG datasets, including IU X-ray, MIMIC, and CheXpert Plus, fully verify the superiority of our proposed method. The source code and pre-trained models have been released on https://github.com/Event-AHU/Medical_Image_Analysis

Information-Theoretic Decoupled Prompt Tuning for Continual Learning

Yunfei Zhang, Wen Wen, Tieliang Gong, Weizhan Zhang — Wed, 23 Sep 2026 00:00:00 +0000

79% · capability 79%, Efficiency 39%, reliability 34%

Continual learning (CL) aims to incrementally acquire knowledge from sequential data while avoiding catastrophic forgetting. Recently, prompt tuning has attracted increasing attention as an efficient approach for adapting pre-trained models to CL tasks. However, existing prompt design paradigms commonly suffer from retrieval dependence and classifier bias, which make model adaptation sensitive to prompt selection and bias predictions toward newly arrived classes. To address these challenges, we propose Decoupled Prompt Tuning for Continual Learning (DPT4CL), which decouples the CLIP textual prompt into a task-shared prompt distribution and class-specific prompts. The task-shared prompt distribution is derived by optimizing an Information Bottleneck objective to facilitate cross-task knowledge transfer and alleviate classifier bias, while class-specific prompts enhance inter-class separability without relying on explicit prompt retrieval. Furthermore, we establish a unified excess risk bound from an information-theoretic perspective, providing theoretical support for the robust generalization and forgetting mitigation of the proposed framework. Extensive experiments on standard CL benchmarks demonstrate that DPT4CL achieves state-of-the-art performance. The source code is available at https://github.com/Cloudfly-Z/DPT4CL

TimeInteract: Towards Real-Time Interactive Intelligence for Streaming Time Series

Sheng Pan, Yongli Gu, Yiqing Guo, Warren Jin, Bo Du, Shirui Pan — Wed, 23 Sep 2026 00:00:00 +0000

79% · Efficiency 79%, Evaluation 68%, capability 59%

Real-world time series evolve continuously, with meaningful changes potentially emerging at any moment. However, existing time-series language models (TSLMs) remain inherently static. They either receive complete sequences for offline processing or alternate between streaming input and response generation, which prevents processing of new observations during interaction. We introduce a new regime, Time-Series Interaction: a model continuously perceives incoming time-series observations and user intent, autonomously decides when to remain silent or respond, and continues processing new observations during response generation. To realize this, we develop TimeInteract with three key designs: a dual-view streaming TS encoder that captures local variations and historical dynamics, a response control mechanism that learns when to trigger a response, and a decoupled streaming inference mechanism that separates control from response generation to avoid blocking subsequent observations. We further formulate a hierarchy of interaction capabilities, progressing from Understanding to Adaptivity. Based on this hierarchy, we construct StreamTSI-34K, a large-scale streaming TS interaction dataset with 34,588 episodes and 77,505 responses across synthetic and real-world time series in single- and multi-turn settings. Across all four interaction levels, TimeInteract consistently outperforms existing LLMs, VLMs, and TSLMs, with gains of up to 23.92 points on challenging tasks. It also improves response triggering while achieving near-zero stream stall and up to $2.15\times$ inference speedup.

The Probabilistic Structure of Large Language Models

Adnan Aboulala\^a — Wed, 23 Sep 2026 00:00:00 +0000

78% · reliability 78%

This paper presents a probabilistic perspective on large language models (LLMs), developed with the aim of bringing together, in a single self-contained account, tools that are usually treated separately across the literature. LLMs are described through probability measures on the set of sequences of tokens, specified via their autoregressive conditional distributions. Training is formulated as a maximum-likelihood estimation problem, addressed by stochastic gradient methods, while text generation is viewed as the sequential simulation of the resulting stochastic process. The role of the asymmetry of the Kullback--Leibler divergence in text generation is examined in relation with characteristic phenomena such as hallucination and the distinction between statistical plausibility and truth. As a complementary illustration of the same viewpoint, we also discuss diffusion models, built around the score function, which cast generation not as sequential token prediction but as the simulation of a reverse-time stochastic process transforming noise into data both in discrete and continuous time.

Towards Adaptive Federated Graph Clustering: A Global Community-aware Contrastive Learning-based Approach

Yinlin Zhu, Di Wu, Wang Luo, Guocong Quan, Miao Hu — Wed, 23 Sep 2026 00:00:00 +0000

78% · capability 78%

Federated graph learning (FGL) enables multiple clients to collaboratively train graph models without sharing their private graph data, providing a promising paradigm for mining knowledge from distributed graph repositories. While most existing FGL methods focus on supervised tasks, real-world graphs are often massive and unlabeled, making federated graph clustering an important yet still immature research direction. Notably, this task is particularly challenging due to the inherent subgraph heterogeneity across clients, which leads to client-specific community structures. In this work, we identify two critical limitations in existing federated graph clustering methods: (1) unrealistic pre-defined cluster cardinality assumptions and (2) incomplete inter-community separation. To address these challenges, we propose AdaFGC, an Adaptive Federated graph clustering framework based on Global community-aware Contrastive learning. AdaFGC introduces an over-complete set of global community anchors to model the global community structure and adaptively estimate clustering cardinality via cross-client anchor refinement. In addition, it employs a global community-aware contrastive learning scheme that uses the shared anchors as contrastive prototypes to explicitly enforce community-level attraction and repulsion across clients, complemented by node-level and topology-level objectives that stabilize local representations. Extensive experiments on eight benchmark datasets demonstrate that AdaFGC consistently outperforms existing supervised and unsupervised FGL baselines across multiple clustering metrics.

MoM: Memory of Memory

Bowen Qin, Yao Lu — Wed, 23 Sep 2026 00:00:00 +0000

77% · Efficiency 77%, Agents 73%, reliability 72%

For a long-horizon LLM agent, the memory question is not what was once recorded but what \emph{currently holds}. Most designs answer it only indirectly: every interaction is stored, and the present is reconstructed at query time by retrieving and reconciling records, so stale values re-enter and the same conflicts are re-litigated. Committing the current value at write time avoids this, but existing write-time (CRUD) memories overwrite, so a wrong update is unrecoverable and prior state is lost. We take the missing combination---\emph{commit on arrival while retaining what is displaced}---and formalize it as \textsc{Memory of Memory} (MoM): memory tracks not only content but the provenance, status, and history of its own entries. We instantiate MoM as \textsc{Provenant Memory} (P-Mem), a typed provenance graph whose \emph{active frontier} exposes one current value per resolved key while displaced values are retained as provenance; typed operations decide whether a new observation supports, supersedes, contests, rejects, revokes, or resolves an existing value. P-Mem's decisive gain is validity rather than accuracy: its turn-level read matches the strongest retrieval memory in accuracy at $\sim$4$\times$ fewer read tokens---a retrieval-granularity effect---while graph-guided turn pruning cuts the knowledge-update stale-answer rate (19.4\%$\rightarrow$10.9\%); on revision chains it stays at 100\% where query-time reading collapses to 25\%, and, because displaced values are retained rather than overwritten, it recovers committed errors a CRUD memory cannot (100\% vs.\ 0\%).

Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

Embedding Team — Wed, 23 Sep 2026 00:00:00 +0000

77% · capability 77%, Efficiency 61%

In this report, we introduce \textbf{Ovis-Embedding}, a state-of-the-art omni-modal embedding family built on native integration of text, image, video, and audio. Instead of assembling separate modality towers, Ovis-Embedding uses a shared multimodal backbone to encode different modalities in a common representation space. Specifically, we make \textbf{three key advances}: (1) \textbf{native omni-modal initialization}: we adopt a pretrained Qwen-omni model as the embedding backbone and adapt it through contrastive training with low-rank initialization; (2) \textbf{data-centric omni-modal training}: we construct a broad, high-quality corpus spanning text, images, video, audio, and interleaved multimodal data. To improve data efficiency, we introduce homogeneous-source sampling to form task-consistent batches with informative in-batch negatives; and (3) \textbf{embedding-specific training and inference optimization}: we use focal loss to emphasize hard examples and similarity-based Embedding Distillation to transfer fine-grained similarity structure from complementary experts. At inference time, low-rank feature decomposition enables compact embeddings with flexible dimensionality and minimal performance loss. Empirical evaluations show that the \textbf{Ovis-Embedding} family achieves state-of-the-art performance on \textbf{MMEB-v3}, \textbf{MMEB-v2}, \textbf{MVEB}, \textbf{MAEB}, and \textbf{RTEB}, demonstrating its effectiveness across text, image, video, and audio modalities. These results highlight the potential of unified omni-modal training to overcome modality fragmentation and advance universal embedding models for any-to-any retrieval.

Latest Exact Match Attention

Moritz Br\"osamle — Wed, 23 Sep 2026 00:00:00 +0000

77% · Efficiency 77%

We introduce latest exact match attention (LEMA), an attention variant for transformers where queries and keys are binarized and each query attends only to the latest exactly matching key. We prove that LEMA transformers with chain of thought can simulate word-RAMs, as was recently shown for the less restrictive rightmost hard attention. In contrast to prior hard attention variants, the restriction to exact matches enables an efficient converse direction: word-RAMs can simulate LEMA transformers at a cost per token independent of the context length. Together, these results yield a close correspondence between the two computational models in terms of both compute and memory. Beyond the theory, we propose a training method for LEMA transformers that handles their non-differentiable operations with a straight-through estimator for the binarization and a soft attention surrogate annealed towards LEMA. On a synthetic associative recall task, LEMA models trained this way use their growing state to store and recall a large number of associations, outperforming gated DeltaNet (GDN) with its fixed state size. As a first scaling test, we train LEMA language models with up to 834 million parameters. They match softmax transformers of around half their size in loss and, on repeated rare phrases and a needle-retrieval task, remain behind softmax transformers but recall across longer distances than GDN models of comparable size. Finally, we implement dictionary-based inference for LEMA transformers and show constant generation speed comparable to GDN despite their growing state, with the dictionaries residing in main memory rather than VRAM. Code is available at https://github.com/moritzbroe/latest_exact_match_attention.

Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts

Vansh Wahi — Wed, 23 Sep 2026 00:00:00 +0000

77% · reliability 77%, Agents 41%

A higher evaluation score does not always mean a better language model system. When optimization exploits an evaluator's mistakes, measured progress can conceal unchanged or deteriorating task performance. This failure can arise through parameter updates, selection among generated outputs, or revisions to persistent prompts. We develop a comparative framework for reward hacking across these three optimization substrates: weights, selection, and text. Building on the Proxy Compression Hypothesis and research on inference-time and in-context reward hacking, we examine how reachable behavior, optimization budgets, and persistent adaptation shape exposure to proxy error. We formalize a distance-dependent upper bound on evaluator disagreement and a capacity ordering for nested policy classes, then show why distance alone cannot establish a universal ranking of vulnerability. An exact finite-output illustration demonstrates how the location of a scoring defect changes the behavior favored by each method. We also map representative defenses across substrates, identifying which mechanisms transfer directly and which offer only functional analogies. Persistent prompts receive particular attention: their contents are inspectable, but the behavior induced by a small textual change may be difficult to anticipate. The formal analysis, numerical illustration, and published evidence together provide a basis for comparing optimization methods and identifying the conditions under which their defenses transfer. The resulting framework connects optimization choices to verification requirements: reliable improvement depends on controlling accessible failure modes and preserving evidence of task quality independent of the score being optimized.

Domain-Adaptive Pretraining Enhances Water Treatment Semantic Representation for Large-Scale Structured Literature Mining

Wed, 23 Sep 2026 00:00:00 +0000

77% · Efficiency 77%, capability 66%

Water treatment research is expanding rapidly, but much of the knowledge acquired from this research remains scattered across unstructured literature. The field still lacks a dedicated language model that can efficiently capture water treatment-specific domain semantics for large-scale literature mining. Here, we address this by developing WaterBERT, a domain-adapted encoder model designed for semantic representation and structured information extraction from water treatment texts. WaterBERT was developed by continual pretraining on a large-scale water treatment corpus comprising about 2.97 billion tokens. Three fine-tuned models based on WaterBERT were systematically evaluated on downstream tasks, achieving the best overall performance among general-purpose and domain-specific BERT models, with F1 scores of 90.12% for multiclass treatment process classification, 79.50% for named entity recognition, and 74.04% for relation extraction. Beyond these benchmark tasks, we further demonstrated WaterBERT's advantages for large-scale literature processing. Applied to 5,144 Environmental Science & Technology articles, WaterBERT-BERTopic identified coherent, diverse, and domain-specific research topics without predefined categories. Building on WaterBERT, we processed 693,211 abstracts at substantially lower cost than commercial LLMs while retaining competitive extraction performance to construct a structured water treatment knowledge graph. The knowledge graph was then integrated with lexical and dense retrieval to develop a Water Knowledge-Enhanced Retrieval System (WaterKERS), which achieved a relevance score of 77.7, substantially outperforming text-based retrieval baselines (54.7-64.5). Through WaterBERT, this study provides a compact and scalable semantic foundation for large-scale information processing and evidence mapping in water treatment research.

Prompt Breadth and Rollout Refresh Interact in On-Policy Distillation

Lingxiang Hu, Tianle Xia, Ming Xu, Yiding Sun, Linfang Shang — Wed, 23 Sep 2026 00:00:00 +0000

76% · Efficiency 76%, Evaluation 43%

How many prompts does on-policy distillation (OPD) need, and how does the answer depend on the student policies that generate its training responses? We study these two controls jointly: prompt breadth and rollout refresh. A 3x3 mathematical-reasoning experiment fixes 14,080 trajectories and 110 optimizer updates while varying the prompt bank and the number of response-generating policy snapshots. With ten snapshots, eight prompts reach 24.09% average accuracy, close to 24.51% for 14,080 distinct prompts. With responses frozen at the initial policy, however, increasing breadth lowers accuracy from 21.16% to 19.05%; under per-update refresh, it raises accuracy from 23.61% to 25.57%. The resulting interaction is 4.07 percentage points, with a 95% question-paired interval of [2.00, 6.28]. Matched comparisons under two teachers reveal a second reversal: the periodic models have higher short-budget accuracy and answer completion, but frozen-response models overtake in average accuracy at a 32K output limit, using 1.7-1.8x as many response tokens. These results show that prompt efficiency in OPD can depend on both refresh and inference budget.

Multi-Term Fourier Graph Neural Network with Sample Relationship Learning for Enhanced Remaining Useful Life Prediction

Ya Song, Laurens Bliek, Yaoxin Wu, Yingqian Zhang — Wed, 23 Sep 2026 00:00:00 +0000

76% · capability 76%

Predicting the remaining useful life (RUL) is essential for effective predictive maintenance. Spatio-Temporal Graph Neural Networks (ST-GNNs), which can model both temporal and spatial relationships by representing time series data as a sequence of graphs, have shown exceptional performance in RUL prediction. However, current ST-GNNs face several drawbacks. First, they require domain expertise or significant computational power to establish graph structures prior to deploying GNNs. Second, the models are restricted to capture temporal dependencies within a predefined fixed-size lookback window. This restriction ignores the common issue of varying time series lengths, leading the prediction model to miss short-term or long-term dependencies. Finally, conventional models often fail to capture the inherent relationships between samples generated from adjacent time windows, which are crucial for improving both the accuracy and robustness of predictions. To address the aforementioned issues, we introduce a novel framework called Multi-Term Fourier Graph Neural Network with Sample Relationship Learning (MTFGN-SRL). Rather than treating the sample as a sequence of graphs, we consider it as a single complete graph and utilize a Fourier Graph Neural Network (FGN) to capture the spatio-temporal information in the frequency domain. We propose a multi-term learning module that utilizes multiple lookback windows to generate samples with varying terms, which are then fed into the FGN to enhance the extraction of useful information from the data. Finally, we develop a sample relationship learning module by training a heterogeneous GNN to identify inter-sample relationships, resulting in enhanced accuracy and robustness in predictions. Evaluations on the CMAPSS dataset demonstrate MTFGN-SRL's superior performance over state-of-the-art methods in RUL prediction.

Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages

Ryan Vo, Duc-Vu Nguyen, Matt Kretchmar, Ngan Luu-Thuy Nguyen — Wed, 23 Sep 2026 00:00:00 +0000

76% · Evaluation 76%, reliability 54%

Language models are increasingly trained on the outputs of other models, forming chains that we call lineages, in which a trait present in one generation can pass to the next. Prior work on subliminal learning has shown that a teacher's trait can transmit to a student through filtered data carrying none of the trait's content. However, the evidence covers only a single training step. We study whether such a trait holds or fades across lineages. We instill the trait into three copies of Qwen2.5-7B-Instruct and iterate the training step to depth ten from each, reading every generation two ways on the same held-out prompts: a keyword screen that looks for expressions of the trait in the model's output, and an activation probe that projects each model's displacement from the base onto a direction built from the other lineages' teachers. We report two findings. First, the trait persists through ten generations across three lineages. The instilled models express it on every completion; the keyword-screen rate falls to 55.6% after the first step and to 21.1% by generation ten. The base itself matches the screen on none of its 300 completions. Second, the trait can be present internally while absent behaviorally. When the model's default system prompt is removed at evaluation, the generation-ten students' keyword-screen rate is zero on every prompt while the probe score stays positive on every prompt. Steering the untreated base with the displacement of a generation-ten student, which is trained and measured under the default system prompt, induces screened expression of the trait even with the system prompt removed, while that same student shows no expression of the trait with the system prompt removed.

MICRO: Multi-Fidelity Active Search for Severe Error Discovery

Orlando Leone, Niclas Pokel, Pehu\'en Moure, Yingqiang Gao, Roman Boehringer — Wed, 23 Sep 2026 00:00:00 +0000

76% · capability 76%, reliability 69%, Efficiency 52%

Human feedback can vary in cost and informativeness. Strong feedback can reveal severe errors but is costly, so cheaper quality ratings can help decide which items to annotate. We propose MICRO (Multi-Fidelity Impact Clustered Rollout), an active search framework that allocates a shared budget to these feedback types to maximise confirmed severe error discoveries. MICRO jointly models ratings and annotation losses conditional on item features to steer acquisition. It clusters acquisitions by their predicted impact on severity probabilities to select diverse candidates, then uses rollout to estimate their discovery value. Experiments on WMT20 English-German show that ratings improve both loss reconstruction and severity prediction. MICRO achieves the highest mean discovery count across four budget and rating cost settings, with similar performance to adapted MF-ENS in one and significant gains over all six comparison policies, including two rollout controls, in the other three $(p<.001)$.

From Utterances to Networks: Modelling Slang Adoption and Diffusion Across Subreddits

Xiaoning Wang, Ted Underwood, Zhewei Sun — Wed, 23 Sep 2026 00:00:00 +0000

75% · Evaluation 75%

Adoption and diffusion of neologisms in online communities have received renewed attention in recent years. As internet slang terms such as APT, referring to a K-pop song, and phrases such as Canon Event meaning an embarrassing but pivotal event, go viral online, it becomes increasingly important to understand the mechanisms that contribute to their success. Prior studies have often explained slang diffusion either from the perspective of social interaction or from the linguistic properties of the slang itself, but rarely from both perspectives together. One major obstacle has been the high cost of annotating slang usage in large-scale online communication. Recent advances in large language models (LLMs), however, make it possible to use them as scalable annotators for such tasks. In this study, we first curate a human-annotated benchmark to evaluate LLM performance in detecting slang usage in real Reddit communication. We then leverage LLM-based annotations to model slang adoption and diffusion. Our results show that slang diffusers with higher bridging capital are associated with increased subsequent adoption, whereas diffusers with higher bonding capital are associated with reduced adoption. We also find that wider contextual usage of a slang term is associated with a longer time before new users officially adopt it. Together, these findings suggest that both social-network structure and linguistic context shape the diffusion of neologisms in online communities.

RankCert: When Can Simulated Learners Safely Select an AI Tutor? Robust Decision Certification Under Structural Uncertainty

Nizam Kadir — Wed, 23 Sep 2026 00:00:00 +0000

75% · reliability 75%, Evaluation 54%

Simulation-based tutor selection can be unstable when predictively adequate learner models imply different policy rankings. RankCert certifies one of eight equal-budget tutoring policies only when model-averaged utility, probability-best, posterior regret, cross-domain rank, family coverage, and leave-one-domain-out and leave-one-visible-family-out averages support the same candidate; otherwise it abstains. We evaluated RankCert in 1,280 frozen held-out settings spanning five rotating held-out oracle families, 64 scenarios per family, and four cohort sizes. Calibration used a licensed, de-identified EdNet-KT1 derivative with 5,000 learners and 590,056 retained responses; all five family representatives passed the frozen adequacy gate. Minimum-domain mean pairwise top-1 agreement was 0.272917 (95% CI [0.253646, 0.293229]), showing substantial structural disagreement. Cohort-noise variance decreased from n = 30 to n = 300, while the structural family share remained nonzero. RankCert reduced total held-out decision loss relative to full-coverage point selection by 0.006605 normalized-outcome units (95% CI [0.004859, 0.008407]). At comparable coverage, however, it did not reduce selective risk relative to a confidence-gated point certificate (difference -0.000213; 95% CI [-0.003238, 0.002384]; Holm p = 0.929654). Certification occurred in 3.75% of settings and only in stable scenarios; RankCert abstained in every ambiguous, misspecified, and structural-conflict setting. "Safe" denotes only benchmark-scoped decision certification under the declared utility and uncertainty set; no human-learning, causal, deployment-effectiveness, or general-safety claim is made.

"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It

J\k{e}drzej Maczan — Wed, 23 Sep 2026 00:00:00 +0000

74% · reliability 74%, Evaluation 56%

Large Language Models (LLMs) tend to add disclaimers like "I'm just an AI" when asked about something related to themselves. The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood. Are the models telling us about themselves or rather how they are deployed? In this work, we show that the chat template works like a switch - when present, it turns this disclaimer voice up and experiential voice like "I feel" down, across 8 popular open-source instruct models up to 9B parameters in size. And conversely when the chat template is not present, it turns the disclaimer voice down and experiential voice up. Inside the activations of 3 models, we find a direction that steers this behavior. Removing the direction in the model's activation space turns disclaimer voice down and adding it turns it up, while a random direction of the same size has little effect. We find that instruct models without chat template, when we add the disclaimer direction to them, disclaim like the template was there. Since the chat template controls the disclaimer voice of LLMs, then researchers studying self-reports or introspection of models might have a confound they need to control for. Our results show that there is a direction they can use to steer this voice. More broadly, our work shows that what models say about themselves is not a fact about them. What they say doesn't come only from weights, but it is partially set by the chat template, and because of that a model's self-description shouldn't be treated literally.

FineWeb-CLaR: Culture, Language, and Region Annotations for Benchmark-Aligned Corpus Auditing

Wed, 23 Sep 2026 00:00:00 +0000

74% · Evaluation 74%, reliability 43%

Cultural evaluation coverage and robustness in language models are difficult to diagnose because pretraining corpora and cultural benchmarks are rarely indexed with comparable metadata. Benchmarks increasingly target culturally situated phenomena at the level of languages, regions, and locale-specific practices, while web-scale corpora are usually organized only by language. A shared culture-language-region layer makes these resources comparable, enabling audits of whether a target cultural phenomenon is represented in pretraining data, evaluated by benchmarks or both. To this end, we introduce FineWeb-CLaR, a large-scale annotated dataset derived from FineWeb and FineWeb-2 that places web documents on a shared culture-language-region axis for corpus auditing and benchmark alignment. FineWeb-CLaR annotates the full 30.9B-document collection from FineWeb and FineWeb-2 with URL-derived region labels and cultural-topic provenance. Our region resolver assigns a non-empty region to 25.61% of documents (7.92B). For cultural-topic analysis, we induce locale-specific topics and project them onto the 14 leaves of the Cultural Taxonomy of Liu et al. (2025), producing Locale Topic Distributions (LTDs) for corpus-side comparison. We also annotate 277 cultural NLP benchmarks with the same taxonomy, language coverage, and region coverage. Together, these resources enable direct comparison between corpus-side pretraining evidence and benchmark-side evaluation coverage.

A Practical Recipe for Semi-Supervised Federated ASR: Online Pseudo-Labels with Server Update Stabilization

Wed, 23 Sep 2026 00:00:00 +0000

74% · capability 74%

Semi-supervised federated learning (SSFL) trains models on clients' unlabeled data using a teacher to generate pseudo-labels, with a small labeled seed dataset on the server. Automatic Speech Recognition (ASR) is particularly fragile here: pseudo-label errors compound across the output sequence and across training rounds into divergence, leaving a large gap to fully-supervised FL. We show that closing this gap turns on two coupled design axes -- the teacher (which model generates the pseudo-labels) and the anchor (the server-side updates on labeled data that stabilize training). On the teacher axis, a per-client online teacher (each client's own evolving model) diverges on its own, but once stabilized it matches or beats the broadcast global teacher (one server model, fixed within a round) -- decisively in-domain and competitively under domain shift. As the seed grows stronger and the online teacher's advantage narrows, a transitioning teacher (global $\rightarrow$ online at round $r$) matches or beats both. On the anchor axis, the server must keep training on labeled data between rounds -- otherwise the online teacher drifts -- and this interleaving, more than the seed model, governs convergence. The two axes are inseparable: aggressive teacher choices pay off only once the anchor stabilizes training, which is highly sensitive to data augmentation and batch size -- the settings that govern how much input and gradient noise the server injects. How much stabilization is needed is domain-dependent, governed by the dispersion of the seed data and its overlap with client data. These findings yield guidelines for SSFL in ASR training, improving over the strongest prior method on 9 of 11 pairs, by $20.8\%$ on average in-domain and $10.0\%$ cross-domain, narrowing the gap to fully-supervised FL.

ArticleMiner: Ontology-Guided Knowledge Graph Construction from Scientific Publications

Md Abrar Jahin, Craig A. Knoblock, Jay Pujara — Wed, 23 Sep 2026 00:00:00 +0000

74% · Evaluation 74%, reliability 44%

Scientific papers keep much of their quantitative content in tables and supplementary files, where a number means something only through its header, caption, unit, analytical method, and the conventions of its field. Recovering the rows and columns of a table is therefore not the same as recovering the scientific fact it reports. Most semantic table-interpretation methods assume that a clean table is already available and subsequently map its cells or columns to ontology terms, whereas most publication-level extraction systems are designed for a single domain. We study a middle path: a shared process that reads a paper and its supplementary files, gathers evidence from several parsers and a language model, and reconciles that evidence, while a bounded human-authored task module for each task supplies the domain meaning. The module lists the canonical names the graph may use, the surface forms that map to them, a small set of derivation rules and validity constraints, an identity key, and the bindings used to write RDF. It defines what a task is allowed to emit; it does not try to list every convention of a field. We build four such modules (for drug-discovery chemistry, materials science, machine learning, and mineral geochemistry) in the ArticleMiner framework, and evaluate them on 163 papers, including a new geochemistry benchmark with expert-curated ground truth. In comparisons against a same-LLM few-shot baseline, the point estimates favor ArticleMiner on all four tasks, with uncertainty on the two smaller benchmarks. The geochemistry comparison also includes access to supplementary files, so its improvement cannot be attributed to domain guidance alone.

Matryoshka attribution: Learning to attribute language model outputs to representations and weights

Wed, 23 Sep 2026 00:00:00 +0000

73% · capability 73%, reliability 48%, Evaluation 38%

Attributing language model outputs to their internal computations is an open problem in interpretability. Existing methods, which use causal interventions, gradients, or learnable masks, either are infeasibly expensive or struggle to identify actual causally-important internal computations. We propose framing attribution as the problem of identifying nested subsets of internal components which minimise a downstream loss. To learn this task, we introduce Matryoshka Attribution (MAttr), a mask learning method that parametrises the mask with a simple differentiable sigmoid top-$k$ operator. We supervise training over all sparsities simultaneously by randomising $k$ over training, resulting in a learned ordering of components by attribution score. MAttr achieves number 1 on the official leaderboard of the Mechanistic Interpretability Benchmark (Mueller et al., 2025); our method identifies sparse and task-transferrable circuits across varying circuit bases. As a practical application, we show that MAttr can be trained with reinforcement learning to identify weight changes responsible for downstream behaviours in LLM finetuning. We train MAttr on refusal judge scores and find that restoring $1\%$ of Llama 3.1 8B Instruct's weights to their base model state is sufficient to remove refusals while maintaining capabilities. We view MAttr as a successful formulation of interpretability into a learnable objective that we can tackle with gradient descent, and encourage future work along these lines.

JAMPR+/L2D: scalable neural heuristic for constrained vehicle routing problems in dynamic environment

Andrew Soroka, Alex Meshcheryakov — Wed, 23 Sep 2026 00:00:00 +0000

73% · capability 73%, Efficiency 37%

The vehicle routing problems with real-world constraints (we consider vehicles capacity limits, time windows constrains, pickup-and-delivery multi-depo --- CPDPTW) pose significant computational challenges. While classical exact and heuristic methods remain effective to solve problems of small/medium size ($N\lesssim100$), they often lack adaptability and scalability for larger logistics tasks. In this work, we show how JAMPR+/L2D RL deep learning model, proposed in to solve large CPDPTW problems can be adopted in the case of substantial changes of graph distance matrix. We test performance of JAMPR+/L2D model for medium-sized CVRP and VRPTW problems on CVRPLIB benchmarks: JAMPR+/L2D outperforms the state-of-the-art heuristic HGS in over 85\% of instances, achieving improvement in objective gap. We show that the JAMPR+/L2D model trained on CPDPTW problem, generalizes well for tasks with simpler constraints (CVRP, VRPTW), for different problem sizes and for moderate changes in distance matrixes. For more substantial changes in distance matrixes, we propose here to make fast finetuning of JAMPR+: on ORTEC data (for CPDPTW) the proposed strategy remarkably reduces the objective gap without full model retraining, what will give both accuracy and rapid inference of the model in the practical routing scenarios with distance matrix changes.

Transcribe, Translate, and Optimize: Joint Reward Learning for Speech Translation

Yanghe Dong, Wanting Huang, Weiran Wang — Wed, 23 Sep 2026 00:00:00 +0000

73% · capability 73%, reliability 30%

In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.

Terminal Shrinkage Averaging Reveals a Schedule-Estimator Interaction in LLM Pretraining

Adam Ousherovitch, Yixin Wang — Wed, 23 Sep 2026 00:00:00 +0000

72% · Efficiency 72%, Evaluation 31%

Large language model (LLM) pretraining conventionally returns the raw final iterate. This couples two design choices: the learning-rate schedule that generates the parameter trajectory and the estimator that constructs the deployed model (e.g. the raw final iterate or a checkpoint average). A schedule that promotes optimization progress may differ from one that minimizes variation in the raw final iterate. Separating these choices creates an opportunity to maintain progress late in training while reducing variation in the returned model. To this end, we propose \emph{Terminal Shrinkage Averaging (TSA)}, which interpolates between the raw final iterate and the average of recent checkpoints to balance recent progress against terminal variation. We analyze how TSA changes the preferred terminal learning-rate schedule under a local quadratic approximation and test this interaction through a sequence of controlled NanoChat experiments. Finally, we demonstrate that the resulting gains transfer to depth-22 NanoChat, where the combined schedule and estimator improve validation quality. A qualifying time-to-GPT-2 run also finishes faster than the public baseline used in our experiments, providing preliminary evidence of benchmark acceleration.

Evaluating Accuracy and Probabilistic Reliability of Zero-Shot Time Series Foundation Models

Panagiotis Michael, Moysis Symeonides, Demetris Trihinas — Wed, 23 Sep 2026 00:00:00 +0000

71% · reliability 71%, Evaluation 33%

Time Series Foundation Models (TSFMs) promise a paradigm shift toward zero-shot forecasting by eliminating task-specific training. However, existing works often overlook trade-offs between predictive accuracy and probabilistic calibration. This paper presents a benchmark study of six TSFMs evaluated on energy, traffic, and financial datasets. We contrast their performance against statistical baselines and a supervised DL model. The study reveals that while TSFMs outperform statistical methods and supervised models, they are subject to a fundamental trade-off between point accuracy and probabilistic reliability. Specifically, xLSTM architectures provide robust probabilistic calibration across horizons. In contrast, patch-based transformers offer competitive accuracy but face calibration issues at long horizons, while transformer-based models exhibit context saturation points for optimal zero-shot reasoning. These findings offer evidence-based guidance for balancing generalization and uncertainty quantification in real-world deployments.

ARAFA: An LLM-Generated Arabic Fact-Checking Dataset

Christophe Khalil, Shady Elbassuoni, Rida Assaf — Wed, 23 Sep 2026 00:00:00 +0000

71% · Evaluation 71%, reliability 45%

Automatic fact-checking poses a significant challenge in Arabic natural language processing due to the scarcity of datasets and resources. In this manuscript, we introduce Arafa, a new large-scale dataset for fact-checking in Modern Standard Arabic, constructed through an automated framework leveraging large language models (LLMs). The dataset was constructed through a three-step pipeline: (1) claim generation from Arabic Wikipedia pages with supporting textual evidence, (2) claim mutation to generate challenging counterfactual claims with refuting evidence, and (3) an automatic validation step to validate that the generated claims are either supported or refuted by their accompanying evidence, or if the evidence does not provide enough information to judge the validity of the claims. The resulting dataset comprises 181,976 claim-evidence pairs labeled as supported, refuted, or not enough information. Human evaluation carried out on a test sample from the dataset demonstrated strong inter-annotator agreement (kappa = 0.89) using Cohen's Kappa for supported claims and (kappa = 0.94) for refuted claims. Automatic validation based on a human-evaluated sample achieved 86% accuracy for supported claims and 88% for refuted ones. To showcase Arafa's value as a resource for automatic Arabic fact-checking, four open-source transformer-based models were fine-tuned using Arafa, with the top-performing model achieving a Macro F1-score of 77% on the test data. In addition to Arafa being the first large-scale dataset for Arabic fact-checking, our framework presents a scalable approach for developing similar resources for other low-resource languages.

Selection-Invariant Communication Compilers for Privacy-Aware Multi-Agent LLM Workflows

Jinghan Xu, Longze Fan, Zeyuan Wang, Xinjin Li, Hankai Liu — Wed, 23 Sep 2026 00:00:00 +0000

71% · Agents 71%, reliability 71%, Evaluation 50%

Structured multi-agent workflows exchange intermediate messages whose content and form can reveal private state even when the final output is safe. We identify selection-channel leakage: after authorization fixes what may be released, a private-state-aware choice among semantically valid realizations creates an additional inference channel. We introduce the selection-invariant communication compiler(SICC), which constrains this post-authorization representation kernel rather than prescribing templates. Any deterministic or independently public-randomized generator satisfying the invariant is valid; requirement-indexed canonical forms are one auditable implementation. We prove a compositional communication-layer guarantee: authorization, public-only form generation, and a dependency-safe utility gate make the emitted transcript reveal no information beyond the complete authorized view. Private-state-aware selection remains vulnerable after surface-disjoint and length-matched controls. Across 132 AgentLeak communication replays and 100 executable LangGraph tasks, deterministic SICC retains complete protocol utility without a positive excess-gain signal; independent public randomization preserves the same result in AgentLeak and 480 controlled cases.

Learning to Defer with Guidance on Real World Medical Data

Emma Sun, Joshua Strong, Alison Noble — Wed, 23 Sep 2026 00:00:00 +0000

71% · capability 71%

Medical image interpretation is high-volume and time-consuming, and while AI interpretation can reduce workload, fully autonomous deployment carries potential safety concerns and low specificity may in practice lead to increased clinician workload. Learning to Defer (L2D) addresses this by selectively routing cases between autonomous prediction and human experts by learning from input features and AI model and human performance. While theoretical guarantees have been proven for L2D, its performance has not been validated on real-world medical datasets with human reader annotations. We evaluate the predictor-rejector formulation of two-stage L2D, where the AI predictor model is fixed and separate from the trainable routing or rejector model, on Collab-CXR, a multilabel chest X-ray dataset with multiple human annotations per case. This is the first work to look at L2D in the context of real-world medical imaging data with human annotations. We further introduce a new setup, L2D with Guidance, where the decision space is extended to three choices: predict autonomously, defer to a human expert, or defer to a human expert and provide AI guidance. We compare multiple rejector architectures and loss functions, and different input feature availabilities. This is reproduced on two larger datasets, VinDr-CXR and CheXpert. Our results show that two-stage L2D with Guidance outperforms classic two-stage learning to defer, as well as human-alone, AI-alone and AI-guided human baselines. Notably, this performance is achieved with simpler loss functions compared to formally defined L2D surrogate loss functions in current literature.

MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning

Kairui Yang, Ziheng Yi, Xunkai Li, Minghao An, Zhanke Liu, Zekai Chen — Wed, 23 Sep 2026 00:00:00 +0000

71% · Efficiency 71%, capability 62%, Agents 55%

Collaboration topology shapes both the performance and execution cost of LLM-based multi-agent systems. Because tasks differ in complexity and required capabilities, recent approaches generate task-specific collaboration graphs that specify agent participation and information flow. However, representative topology generators use either individual agents or predefined groups throughout an organization, overlooking differing collaboration needs across subtasks. Our key insight is to select granularity locally for each functional role, combining fine-grained control with reusable collaboration patterns within one organization. Learning such organizations requires exploring a combinatorial construction space with limited intermediate feedback from final-answer rewards. Therefore, we propose MAGIC, a dense-reward reinforcement learning framework for mixed-granularity graph generation. Specifically, MAGIC constructs a mixed-granularity agent graph by sequentially selecting a functional role, instantiating it as a single agent or reusable group, and connecting it to existing units. We directly optimize the construction policy using returns from trajectories sampled under the current policy and use potential-based reward shaping to provide intermediate feedback from probe-based utility and structural signals while preserving the cumulative task reward. MAGIC outperforms state-of-the-art baselines across eight benchmarks and demonstrates strong inference efficiency in our efficiency study.

Attention as a Routing Graph: Live Circuit Extraction from a Single Forward Pass

Ash Manvi, Samreena Tajreen — Wed, 23 Sep 2026 00:00:00 +0000

70% · Efficiency 70%, Evaluation 65%

Finding circuits in language models usually means running many careful interventions. We try something simpler: treat attention as a routing map from one forward pass, keep a small set of routes that point toward the answer, and ask whether those routes actually matter. They often do. On induction and IOI (tasks where the "right" circuit is already known), ablating our extracted edges hurts the model much more than ablating a random set of the same size. We evaluate n=100 prompts per cell on GPT-2 Small, GPT-2 Medium, and Pythia-410M, with paired gap tests and bootstrap confidence intervals. The extract step costs one forward; a head-by-head patch sweep costs about two orders of magnitude more. We are not claiming a complete circuit atlas. We are claiming a cheap sketch that carries real causal signal on known tasks, with clear failure modes when it does not. Code and evaluation artifacts are at https://github.com/Aquinf03/live-circuit-routing.

Learned Enterprise Data Comprehension: Compression and Routing for Data Agents

Ethan Torres, Eric Mills — Wed, 23 Sep 2026 00:00:00 +0000

70% · capability 70%, Efficiency 54%, Agents 52%

Structured-data agents in enterprise settings must reason over complex data environments whose relevant evidence is distributed across schemas, relationships, policies, and recurring business roles. Modern agentic systems often address this burden through reusable markdown-style memory or skill files that preserve previously discovered information for later queries, reducing the need to rediscover the same structure repeatedly. This is useful, but it obscures a natural division of labor: agents are well suited to semantic reasoning, while learned systems are well suited to predicting and organizing recurring structure. We introduce latent equivalence learning to bridge this gap. The framework separates persistent task-relevant identities from their dataset-relative realizations. In our realization, supporting and opposing evidence shape support-realized Gaussian prototypes that learn how those identities are expressed in a particular data environment, while soft-membership profiles retain distinctions lost under a hard assignment. A separate learned query-prototype system represents recurring evidential requirements and maps them through a learned compatibility function into the same persistent identity structure. This identity-factorized, query-conditioned routing materializes the relevant dataset-specific evidence for downstream reasoning, allowing the agent to operate over an already organized evidential state rather than reconstructing cross-schema structure at every query. On the Data Agent Benchmark, spanning 54 queries across 12 heterogeneous datasets, our full implementation achieves 94.67% dataset-macro stratified Pass@1 over five complete trials and 258/270 successful raw query attempts, compared with 55.51% for the benchmark's Claude Opus 4.6 reference agent, ranking first among 40 leaderboard entries at submission.

Potential for Enhanced Learning in Machine Learning Classes by Using Wiki LLM Indexing

Brian Wright — Wed, 23 Sep 2026 00:00:00 +0000

70% · reliability 70%, Evaluation 64%

Large language models are increasingly deployed as course-specific tutors, but their usefulness depends on grounding in vetted instructional materials that are often revised mid-semester. Our prior work built a multimodal retrieval-augmented generation (RAG) system over an authentic machine learning course corpus (Foundations of Machine Learning) and found that retrieval improved contextual grounding, but that fixed retrieval strategies were suboptimal. That motivates a different question: whether how a corpus is structured at ingest time matters more than how much is retrieved at query time. We present a controlled head-to-head comparison of two knowledge representations over an identical classroom corpus: (A) vector RAG, replicating the best-performing configuration from our prior study, and (B) an LLM-compiled wiki (Karpathy framework), in which the corpus is synthesized at ingest into linked concept pages with explicit cross-references and citations back to source materials. We evaluate 59 questions spanning single-fact recall, cross-unit concept linking, synthesis and explanation, and currency after a syllabus revision, scored by an LLM judge against a human-authored rubric. Both representations answered single-fact questions about equally well (9.33 vs. 9.96 of 10), but diverged sharply on questions requiring links across course units. The compiled wiki remained accurate and grounded (9.93; 100% grounded in cited sources), while retrieval scored lower and was markedly less grounded (8.14; 64%). The wiki's citations let students and instructors trace any claim back to the lecture that introduced it, adding a layer of dynamic retrieval that machine learning courses require. While further testing is needed, instructors using AI to support learning in ML courses should consider wiki-based structure for its potential to support foundational elements of best practice.

Ladders of Thought: A Self-Evolving Curriculum of Progressively Simplified Reasoning Traces

Minghui Liu, Thomas Magelinski, Dehao Yuan, Qi Yu, Furong Huang — Wed, 23 Sep 2026 00:00:00 +0000

70% · capability 70%, Efficiency 49%

Large language models (LLMs) excel at reasoning when scaled to hundreds of billions of parameters, but small- and mid-scale models remain brittle reasoners even with knowledge distillation (KD). We present Ladders-of-Thought (LoT), a framework that improves reasoning by combining progressive question rewrites with a self-evolving curriculum. LoT automatically generates semantically faithful but easier variants of reasoning problems, organizes them into difficulty buckets using step-based measures, and employs a self-evolving bandit scheduler to allocate training adaptively. Evaluated on two reasoning domains, math and multi-hop reasoning, across 1-8B models from different families, LoT consistently improves over KD. It delivers large gains on arithmetic tasks (e.g., +32 percentage points on AddSub, +25pp on SVAMP), +2-8pp improvements on in-domain test splits, and strong though dataset-dependent benefits on multi-hop reasoning (e.g., +16pp on QASC, +25pp on StrategyQA). LoT also converges faster than staged curricula, highlighting the value of adaptive progression. These results show that progressive rewrites coupled with adaptive curricula provide a simple yet effective recipe for strengthening reasoning in smaller LLMs.

FISSION: Label Augmentation for Bot Detection

Sen Yang, Ignacy Nieweglowski, Aviv Yaish — Wed, 23 Sep 2026 00:00:00 +0000

70% · capability 70%

Bot accounts and coordinated influence operations are often discovered via heuristic methods, leaving a dearth of reliable ground-truth labels for training detection systems. To address this challenge, we study a natural question: can we generate labels to assist in learning embeddings in which bots and accounts from the same coordinated operation are close? We present FISSION, a method to generate labels by splitting each account's activity into positively labeled sub-accounts. Given this label source, we train detection models which preserve behavioral regularities recurring across positive sub-accounts. We evaluate FISSION and show it outperforms prior methods in detecting Wikipedia sockpuppets and Twitter/X bots.