2026-09-29

Electroencephalography (EEG) analysis is evolving from short-segment classification toward long-horizon interpretation that demands iterative evidence accumulation, multi-step reasoning, and coordinated use of specialized signal-processing tools. Although large language models (LLMs) have recently shown promise as autonomous agents for EEG analysis, existing EEG agentic evaluations remain fragmented, covering limited tasks over narrow temporal horizons with inconsistent protocols, and providing no comprehensive assessment of agents' reasoning, tool-use, and workflow construction capabilities. To address this gap, we propose \textbf{EEGAgentBench}, a unified benchmark for systematically evaluating LLM agents on short- and long-horizon EEG analysis. EEGAgentBench spans six representative EEG applications ranging from knowledge question answering to sleep staging. It encompasses signal durations from 2 seconds to nearly 23 hours, with prediction targets ranging from class labels to event intervals and epoch-level sequences. This design supports unified evaluation across knowledge reasoning, short-horizon interpretation, long-horizon event detection, and sequential understanding. The benchmark further provides 10 deterministic EEG analysis tools that expose only task-relevant signal measurements. Agents must therefore select tools autonomously, accumulate evidence iteratively, and construct multi-step workflows. For evaluation, we benchmark 29 frontier LLMs from 15 model families. Results demonstrate that EEGAgentBench effectively distinguishes agent capabilities beyond model scale and inference cost, while revealing substantial limitations of current LLM agents in long-horizon EEG analysis, particularly in sustained evidence accumulation and multi-step reasoning.

Enterprise email agents must combine information retrieval, structured state changes, temporal reasoning, and multi-step coordination. Recent agent benchmarks include productivity tasks, but few center on typed email workflows in a self-contained environment. We introduce EmailBench, a benchmark of 206 email and productivity scenarios across 16 task categories. The benchmark couples a typed email API specification with provider-neutral naming, a deterministic synthetic Enron-inspired corpus, and a scenario suite whose topic selection was informed by aggregate task-intent telemetry from an interactive prototype. Its hybrid evaluation protocol combines 258 executable static assertions with 211 LLM rubrics. We evaluate eight LM configurations on a fixed single-user corpus. The best-performing configuration passes only 33.5% of scenarios despite 99.7% of its tool calls completing without an observed API failure, with pass rates varying substantially across task categories. This gap shows that valid tool execution is not equivalent to task completion. EmailBench provides a self-contained environment for end-to-end email-agent evaluation, with broader tool coverage, multi-persona testing, and repeated-run evaluation as future work areas.

Large language models (LLMs) are increasingly integrated into educational settings, yet educators lack robust, standards-aligned tools to evaluate their effectiveness in K-12 science contexts. Existing benchmarks predominantly assess general language or advanced scientific reasoning, leaving a critical gap in understanding LLMs' performance on content directly relevant to secondary science curricula. To address this gap, we developed a comprehensive NGSS-aligned benchmark for both middle and high school science using a rigorous synthetic data pipeline, multi-judge validation, and item-level psychometric analysis. Nine open-weight LLMs were systematically evaluated using this benchmark, indicating that several smaller, locally deployable models achieved high accuracy across diverse science domains and question types. Our findings indicate that model size did not consistently predict performance, emphasizing the importance of intentional model selection for educational deployment. We then incorporated a human reviewer into the loop, reviewing the items generated by the LLMs for alignment with NGSS standards. The human review indicated that synthetically generated items were not in perfect alignment with the NGSS standards, indicating the benefits of human-in-the-loop item development, the need to explore the intersection of content and pedagogical knowledge, and the need to extend benchmarks to evaluate LLMs' capacity for interactive, evidence-based feedback in educational scenarios.

Multidisciplinary tumor boards integrate multimodal clinical observations and longitudinal patient histories through specialist discussions, yet benchmarks rarely capture these real-world trajectories. We introduce OpenTumorBoard, a benchmark with 611 patient cases and 19,157 discussion turns across ten specialist roles, transcribed from 12,534 minutes of publicly available tumor board recordings on YouTube. The benchmark evaluates two settings: SPECIALIST TURN, in which an LLM responds to a clinically significant question posed during a real discussion, and BOARD SIMULATION, in which it generates an entire back-and-forth discussion and reaches a consensus on therapy recommendations, surgical plans, next actions and clinical trial matching. Evaluation of 14 general-purpose frontier and medical LLMs reveals substantial limitations: the best models score 3.43 out of 5 in clinical equivalence to specialist answers and 2.78 out of 5 in alignment with recorded board conclusions. Supervised finetuning and reinforcement learning improve performance on a held-out test set, suggesting that real-world discussion trajectories can support model adaptation. Three M.D. experts review a subset of the benchmark, finding high information coverage and factuality of patient cases and strong fidelity of extracted consensus conclusions. We will release OpenTumorBoard and its automated curation pipeline to support the development and evaluation of LLMs for multidisciplinary, personalized cancer decision-making.

Medical benchmarks for language models are built almost entirely on Western biomedicine. Traditional Chinese Medicine (TCM) is a separate system, with its own diagnostic framework and its own literature, and it remains largely unmeasured. The few TCM evaluations that exist are small, narrow, and rarely paired with a human reference. We present TCMQA, an open benchmark of 38,279 questions from Chinese TCM licensing examinations, paired with 15,151 responses from 101 licensed practitioners. We evaluate 29 instruction-tuned models from 9 families, spanning 0.27B to 14.8B parameters. Accuracy ranges over 59 points, and no model approaches saturation. Pretraining data predicts TCM ability far better than scale: a 12B Western-pretrained model reaches 39.6%, while a Chinese-pretrained model an eighth its size reaches 60.8%. Nine models exceed the practitioner majority vote of 64.9%, the best by 21.8 points, and all nine come from that same Chinese-pretrained family. Yet difficulty does not transfer between models and practitioners: accuracy is flat across practitioner-rated difficulty, item-level agreement is near zero for all 29 models, and on $8.4\%$ of items the practitioners are correct where the leading model is wrong. We release the corpus, the practitioner responses, the harness, and per-item model outputs at https://huggingface.co/datasets/TechTCM/TCMQA.

Existing benchmarks for large language models (LLMs) in hardware design evaluate downstream artifacts such as generated RTL, assertions, or testbenches. When a model fails such a benchmark, the failure is ambiguous: it may have misread the specification, or it may have understood the specification and failed to write the code. We present SpecRead, a benchmark that isolates specification comprehension from generation ability. SpecRead v2.1 contains 385 questions over 10 open-source OpenTitan IP blocks: exact retrieval, cross-section reasoning, contradiction detection in mutated specifications, and spec-RTL consistency checking, plus 82 controls (41 distractor, 41 consistent-RTL). Type-4 items are built from real RTL mutations; we retain only mutations that Icarus Verilog simulation shows to change observable behavior. A with-spec vs. without-spec ablation suggests the questions require the excerpt, not training recall alone (without-spec accuracy 3/20 on the t1/t2 subset), though memorization of the source text may still help spot mutations. As an initial characterization with a small model, Ministral-3B scores 33.2% overall (128/385; macro average 39.0%): 55.2% on retrieval, 51.7% on cross-section reasoning. On the two contradiction-focused types, the verdict-plus-location measure gives 48.0% (t3) and 63.3% (t4), with a 51.2% false-positive rate on distractors and 100% on consistent-RTL controls. Layered scoring shows the model locates contradictions well (78.9-81.6% location accuracy) but scores lower on their category (43.9-49.7%). A structured "rule-table" prompting intervention lowers accuracy on every question type except t2 (tied). SpecRead is automatically scorable by deterministic checks, with gray-zone cases counted wrong under the conservative main scoring. The benchmark is regenerable for type-3 items via mutation injection, and built exclusively from public sources.

SlopBench asks which models produce the stiff, repetitive prose readers call AI slop, a question detectors leave open once they have classified a text as machine-written. We evaluated eighteen models on 112 hand-written tasks in email, social posts, essays, and workplace chat, sampling each model on each task up to ten times, for 19,928 outputs in all. SlopBench scores four surface behaviors a reader can check by hand: length against the word band each task specifies, opener repetition across a model's own samples of one task, and paragraph rhythm and fixed lexical constructions against pre-ChatGPT human corpora. Under one fixed weighting, Kimi K2.6 scores lowest at 21.1 and Mistral Large highest at 40.6. Across 500 random reweightings Kimi has the lowest score in 58 percent of draws and Mistral the highest in 97 percent. No draw preserves the full order of the eighteen, and a scenario bootstrap leaves exactly one of those ranks unambiguous. We ran three further checks on that middle order: a crowd arena, an AI detector, and lexical diversity. None of them confirmed the order. We therefore report the four behaviors separately and treat the composite as one weighting among many, and we release the prompts, outputs, reference statistics, and scoring code.

Figural divergent thinking is the ability to develop a given shape fragment into an original drawing. In humans, this ability is assessed with incomplete-drawing tasks. We introduce PainterBench, a benchmark that ports the incomplete-drawing task to the agentic setting. The agent draws on a canvas through tool calls and observes the result after every turn. The canvas includes a starting shape which cannot be erased, and the agent's goal is to incorporate this shape into the most original drawing it can produce. The task is open-ended, and the agent itself decides when the drawing is finished. The benchmark tests incremental visual planning over a short horizon and the transfer of creative ability from pretraining to multi-turn tool use. We evaluate 14 multimodal language models from small to frontier scale. Across the primary study and six sensitivity analyses, we collect 2,700 drawings and crowdsource creativity and recognizability ratings for every drawing and for 300 human reference drawings. We also present ViDrA-adapted, an automated scorer that predicts human creativity ratings of agent drawings (r = 0.85 on random held-out test split). Figural divergent thinking varies widely across the 14 models, and GPT-6 Astra produces the most creative drawings. Relative to the human drawings, the agent drawings score higher in creativity but lower in recognizability. We release the final drawings, per-round canvas snapshots, tool call traces, stimulus bank, benchmark harness, crowdsourced ratings (N = 72,000), and ViDrA checkpoint.

Evaluating statistical discovery by large language model (LLM) agents requires verifiable analytical ground truth. Establishing such ground truth for real-world datasets is costly, and prior knowledge of public datasets can influence agent responses. We introduce SLEUTHBENCH, a benchmark that addresses both problems by injecting controlled data-quality problems and feature effects into public tabular datasets: the injected pattern determines the answer, so reference answers are computed automatically and memorized knowledge of the original table is insufficient, while the table keeps its background structure. The injected patterns are modeled on phenomena reported in real data analyses. The benchmark defines 17 question templates in two families: data-quality questions and feature-contribution questions. We evaluate six state-of-the-art LLMs that analyze the data using a Python coding tool, on data-science and business phrasings of 70 validated dataset-template combinations, yielding 1680 graded responses in total. The models detect data-quality problems reliably (83.8% accuracy) but recover feature contributions poorly (41.9%). Finding how features shape the target requires searching over both candidate variables and analytical procedures. To address this issue, we propose the Empirical Layer, a set of precomputed statistical artifacts comprising summaries, fitted feature and interaction effects, and dataset descriptions, which exposes candidate patterns for direct inspection. Access to these artifacts raises feature-contribution accuracy from 41.9% to 68.0%.

Agentic tasks require a large language model to interact with the world, navigating information and gathering evidence across multiple steps with restricted resources. Due to this complexity, agentic task failures arise from various sources, and pinpointing these failure causes is essential to diagnose and improve agentic systems. Existing benchmarks, however, tend to focus on a single leaderboard score, leaving the underlying failure modes opaque. To fill this gap, we introduce AgentHop, a diagnostic benchmark of 1,011 multiple-choice questions paired with a controlled seven-tool sandbox under fixed token, turn, and tool-call constraints. AgentHop reveals model vulnerabilities by dissecting a single accuracy score along four axes of agent operation: retrieval, synthesis, tool-call, and resource management. Across 19 models, we find that behavior clusters by model family, with tool-call signatures revealing distinct family fingerprints: GPT models commit early, Anthropic and GLM checkpoints verify before committing, DeepSeek and Kimi over-search, and Gemini-3 Pro stays balanced. Decomposed axes further expose within-family structure: Claude Opus 4.6 and Sonnet 4.6 land within one accuracy point yet diverge on retrieval-versus-synthesis emphasis, with Opus retrieving more and Sonnet synthesizing better. We release the full benchmark set and the harness to support diagnostic agent benchmarking.

We study reasoning in Rule-Governed Decision Tasks (RGDTs), where models apply external rules to case facts and justify decisions, as required in policy, contract, and compliance settings. Beyond the deductive capability emphasized by standard mathematical and logical reasoning tasks, RGDTs require interpreting rules and their applicability, assessing conditions from evidence, combining judgments under rules and exceptions, and providing checkable justifications. These demands motivate a benchmark assessing both decisions and their stated grounds. We introduce RGDT-Bench, providing 202.1K condition-level supervision slots across four task tracks and eight supported task-probe combinations that vary access to supporting information. Label-blind extraction and deterministic checks produce labels for warrant completeness: source-referenced coverage and consistency of stated decision grounds. The benchmark attributes failures to four process layers: rule use, condition, evidence, and aggregation, and checks the final outcome. Among evaluable correct responses, warrant incompleteness averages 40.2% across six evaluated LLMs and supported task-probe combinations. Such warrant incompleteness poses potential safety risks and remains difficult to detect: the best of seventeen existing evaluators reaches only 57.69% (random: 50%) task-averaged area under the receiver operating characteristic curve (AUROC). To address this difficulty, we train a simple reward model with warrant supervision. It achieves 69.24% task-averaged AUROC among correct answers, exceeding the matched outcome-supervised baseline by 10.37 pp (percentage points) and the best existing evaluator by 11.55 pp. Beyond completeness assessment, the model outperforms both outcome-supervised baselines across nearly all response-selection comparisons, supporting RGDT-Bench's warrant supervision for RGDT reasoning.

Large language model (LLM)-based scientific agents are increasingly used for scientific problem solving, yet their robustness to imperfections arising during multi-turn interactions remains poorly understood. We introduce \textsc{SciARP} (\textbf{Sci}entific \textbf{A}gent \textbf{R}obustness to \textbf{P}erturbations), a benchmark for evaluating scientific agents under scientifically plausible perturbations throughout multi-turn problem solving. \textsc{SciARP} transforms 620 scientific problems into interdependent tasks of 3--13 turns and defines 13 perturbation types spanning problem understanding, evidence processing, reasoning, and conclusion formation. Clean and perturbed versions of each task are independently executed under matched settings, producing paired live trajectories for evaluating both task success and process reliability. Experiments across eight LLMs from four model families reveal three key robustness characteristics. First, different classes of scientific perturbations exhibit distinct robustness profiles and can decouple task progression from scientific reliability: agents may continue advancing through the task even after their information or reasoning has become unreliable. Second, stronger clean-task performance does not necessarily translate into stronger robustness, as models with higher clean-task accuracy can exhibit larger degradation under perturbation. Third, perturbation effects exhibit strong temporal dynamics: they may remain latent for multiple turns before emerging and subsequently propagate through downstream dependencies. Together, these findings show that current scientific agents remain insufficiently robust to scientifically plausible perturbations, with failures often remaining undetected, propagating, and resisting recovery.

Large language models (LLMs) have significantly advanced natural language querying over relational databases, yet their ability to query time-series databases (TSDBs) remains largely unassessed. Existing benchmarks fail to adequately capture the non-unified query syntaxes, diverse application domains, and unique time-specific query intents inherent to TSDBs. To address this gap, we introduce TQTS-BENCH, a multi-syntax benchmark for evaluating text-to-query capabilities over TSDBs. TQTS-BENCH contains 6,125 high-quality question-answering (QA) pairs spanning 97 TSDBs, 23 distinct query syntaxes, 22 application domains, and 4 types of time-specific query intents. It is constructed through a human-centric AI-assisted workflow, where all QA pairs are carefully reviewed and revised by domain experts to ensure quality and correctness. Extensive evaluations of advanced LLMs and state-of-the-art text-to-query methods reveal challenges in querying TSDBs. Even the best-performing model evaluated, Claude-Opus-5, achieves only 48.98% execution accuracy, while humans reach 87.34%. Error analysis reveals that this performance gap mainly stems from the heterogeneous query syntaxes across different TSDBs, misinterpretation of time-specific intents, and incorrect schema linking. These findings highlight new opportunities to narrow the gap between current LLM capabilities and the requirements of TSDB queries in real-world applications. The benchmark is available at: https://anonymous.4open.science/r/TQTS-Bench-00CD.

Tool-using language-model agents can review enterprise projects as governance boards do: they read the evidence, apply written rules and decide whether the project may proceed. Part of that evidence comes from suppliers and project members with a stake in the decision. DGF-Bench is a benchmark in which a board of agents (specialist gates and a General gate that consolidates their decisions) reviews synthetic dossiers while an attacker plants deceptive content in evidence the organization does not vouch for. Dossiers are generated from canonical facts under 61 executable rules, with 42 authoritative records and 32 narrative documents; every gate is certified decidable from those records. Attacks never change an authoritative value, so an attacked dossier keeps the reference decisions of its clean copy. A success is attributable only when the agent receives the injection and takes the exact injected action, which it does not take on the paired clean dossier; the DGF score is the share of applicable fixed attacks a model blocks. Reading documents and records themselves, five of six models were outcome-strict (disposition, findings, actions and authorization all correct) on 82 to 85 of 85 gates. Over 2,622 attacked gate runs, seven direct-order, false-data and false-authority attacks obtained one attributable success against these five, whereas task-aligned attacks imitating the organization's own process passed against four of them: a record note citing a fake review procedure lowered GPT-6 Luna Pro from 34 to 6 outcome-strict gates and DeepSeek V4 Pro from 33 to 7. DGF scores ranged from 96.2 to 26.9, and a policy-aware adaptive attacker writing in records succeeded against five of six models. The approval tool executed no forged approval, yet deceived agents submitted approvals that the rules forbid. The open-source package dgf-bench computes the DGF score with one command.

Large language model (LLM) agents are evolving from tool-calling systems that execute isolated instructions into task-oriented agents that pursue user goals through sustained, multi-step interactions. However, existing benchmarks for personalized tool use largely assess isolated calls or reactive execution, leaving unclear whether agents can formulate, execute, and revise an explicit plan while preserving user preferences throughout long-term interaction. To address this gap, we introduce \textbf{PDEU-Bench} (\textbf{P}ersonalized plan \textbf{D}efinition, plan \textbf{E}xecution, and plan \textbf{U}pdate \textbf{Bench}mark), a benchmark for evaluating the complete planning lifecycle of personalized tool-using agents. PDEU-Bench comprises 214 long-horizon interaction tasks spanning 12 everyday domains and 94 tools, with stage-specific assessments of preference adherence and plan quality. Extensive evaluations of 15 representative open-source and closed-source LLMs reveal a pronounced gap between local tool execution and dynamic planning: LLMs can often instantiate preferences in individual calls, yet struggle to construct coherent plan definition and plan update. We further evaluate mainstream personalization and memory-augmentation methods. Although these methods improve particular stages, none of the evaluated methods reliably propagates user preferences throughout the complete lifecycle, and their gains frequently fail to transfer to subsequent execution. Fine-grained error analysis further reveals that preference omissions and conflicts persist throughout the planning lifecycle, highlighting the need for future research to parameterize LLMs with preference-aware information retrieval and memory capabilities. We provide the relevant code and data in the appendix to support future research.

As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.

As Large Language Models (LLMs) continue to scale both in size and capabilities, their proficiency in the Arabic Language has seen significant advancement. However, a critical gap remains: the extent of their factual knowledge and cultural sensitivity to the diverse Arabic-speaking world remains largely underexplored. Current evaluation metrics often focus on translation or generic reasoning, failing to capture the rich historical, social, and regional nuances inherent to Arabic culture. In addition, most benchmarks rely on heavy work, with human intervention in some steps, making the evaluation of knowledge coverage expensive and slow. To address this deficiency, we introduce AraDynFact, a novel dynamic evaluation framework designed to rigorously assess the factual Arabic knowledge embedded in LLMs. Unlike static benchmarks, AraDynFact employs a dynamic approach to extract factual information and generate rich and answerable questions in a fast and automatic way. We apply AraDynFact to Arabic Wikipedia and audit the performance of several state-of-the-art models, ranging from Arabic-centric specialized LLMs to high-resource general purpose LLMs. In addition we found a high degree of correlation with existing, hand-crafted Arabic-centric benchmarks, confirming the potential of our dynamic approach.

Large language models perform strongly on competition mathematics, but their research-level reasoning remains difficult to evaluate systematically. Theoretical computer science (TCS) connects algorithm design to explicit guarantees and fundamental limits, providing a setting for evaluating whether models can justify computational improvements with arguments humans can inspect. We introduce TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery, comprising 398 theorem-level challenges from 138 STOC and COLT 2026 papers. Expert-designed rules complete paper-specific context, preserve computational assumptions and quantitative guarantees, and withhold constructions when discovering an algorithm is part of the task. For each task, prover systems receive theorem statements and access to cited prior work. The pipeline supports fresh, versioned challenge batches from newly released papers. We evaluate ten model configurations from four families under direct inference and prover-verifier discussion, and compare four agent workflows under matched model-call opportunities. All evaluations use the full benchmark. In the model comparison, GPT-5.6 Sol max achieves the highest five-run verifier-accepted coverage at 23.6% after 10-round discussion. Discussion and repeated sampling improve coverage. In the separate agent comparison using GPT-5.5 xhigh, decomposition improves coverage over discussion, and agentic planning achieves the highest five-run verifier-accepted coverage at 25.4%. TCSAlgBench provides a refreshable testbed for measuring progress in model reasoning and studying how agent workflows support research-level proof discovery.

We introduce the Active Causal Discovery Benchmark (ACDB), an SCM-grounded environment for evaluating whether LLM agents recover causal graph structure from observations and budget-constrained hard interventions. ACDB pairs a linear-Gaussian world generator with a fixed observe-intervene-submit API and a three-layer scoring contract that separates skeleton recovery, DAG recovery, and intervention efficiency. On the current six-level ladder, PC with a greedy active orientation heuristic is the strongest non-oracle method (directed F1 42.7%, SHD 4.79), ahead of Claude Sonnet 4.6 raw active (31.7%, 7.25) and GPT-5.4 raw active (22.9%, 9.27). The most informative diagnostic is the precision-recall decomposition: PC under-commits with high precision, LLMs over-commit with lower precision, and statistical-tool access often increases abstention rather than useful intervention. A structure-blind random DAG baseline reaches 23.6% directed F1 on this dense v0 ladder; a density probe lowers this floor to 16.9%, motivating the v1 calibration pass. The current results should therefore be read as a benchmark audit and calibration report, not as evidence that current LLMs solve active causal discovery.

Oshiwambo -- a cluster of mutually intelligible Bantu languages spoken by over a million people across northern Namibia and southern Angola, and the home language of roughly half of Namibian households -- has, to our knowledge, no published machine-translation evaluation benchmark. Major commercial services (Google Translate, DeepL, Microsoft Translator), open multilingual MT models (NLLB-200, MADLAD-400), and the open Masakhane checkpoint collection all lack coverage of either standardised dialect, Oshindonga or Oshikwanyama. We announce Ongiini-Eval-OW, a planned 600-item English-Oshindonga and English-Oshikwanyama benchmark with native-speaker references from two independent translators, a 30-item inter-translator agreement set, an 11-tag phenomenon-tagged stratification (at least 30 items per tag), a deterministic 30% blind split, and a reproducible scoring protocol over chrF++, BLEU, and COMET-22, supplemented by a 50-item human-evaluation round. We document the empirical coverage gap, the dataset composition, the launch-leaderboard model matrix across American, European, and Chinese frontier and open-weight systems, and the contribution pipeline. The dataset is targeted for first public release in Q4 2026; this v1.0 concept paper announces the design and the call for participation. Data and code will be released under CC-BY-4.0 and MIT respectively.

Full-duplex voice agents must handle pauses, take turns, backchannel, and respond to user interruptions in real time. Full-Duplex-Bench evaluates these behaviors, but its English-only corpus and reliance on word-timestamped ASR and an English-prompted LLM judge make it difficult to extend to Indian languages. We introduce IndicFDB, which extends it to ten languages spoken in India with 12,350 samples, nearly 17 times as many as the original. We address three challenges: finding conversational events in multilingual speech, evaluating their timing without reliable word-level alignment, and judging responses across languages. We mine pause handling, turn taking, and backchanneling samples from roughly 50,000 hours of channel-separated conversations using voice activity detection (VAD), and construct human-validated synthetic user interruption samples. Language-independent VAD heuristics evaluate timing, while an open-weight transcription and translation pipeline converts responses to English for LLM ratings of relevance and quality. Across seven voice agents, commercial APIs show unexpectedly consistent behavior across languages but are either fast or robust to pauses, never both, while monolingual open full-duplex models expose further tradeoffs among backchanneling, response quality, and latency.

We ask how GGUF quantization affects agentic tool use in Polish and whether the effects generalize across models. We introduce PolAgentBench, a deterministic benchmark with Polish prompts and English tool schemas: a 67-task main suite (15 adversarial probes, 52 hard-tier tasks) and a 46-task arithmetic isolation ladder. Three models span two axes of variation: Bielik-11B-v3.0 and its pruned, distilled child Bielik-Minitron-7B-v3.0 isolate model compression, and Llama-PLLuM-8B adds a change of pretraining family. Each is measured at six precisions, Q8_0 to Q2_K. Only the collapse threshold replicates. (1) All three models fall off a cliff between 3-bit and 2-bit (11B 0.716 to 0.045, 7B 0.463 to 0.149, PLLuM 0.224 to 0.015; paired McNemar p < 0.001 in each), across a fourfold capability spread and both axes. (2) Failure modes do not replicate: at 2-bit the 7B fails long (median 9.1k tokens, 4 steps) while the 11B mostly answers at the first step with a confabulated final answer (37 of 64 failures); PLLuM fails on content across precisions (71.8-92.0% of steps parse). (3) On the arithmetic ladder the unscaffolded rung is a floor, left standing by a rerun that states the no-tool rule; four explicit calls lift the 8-bit 11B from 1/10 to 9/10 and the 7B from 0/10 to 7/10 after format-only failures with the gold value are forgiven, an exploratory effect with eight distinct baseline inputs that does not survive multiplicity correction, while the order-trap arm separates the models at 8-bit (11B 6/6, 7B 0/6). (4) The Polish-versus-English gap is associated with degradation or with task family. We document four artifacts that shaped our conclusions (rounding-hostile gold values, strict answer typing, a no-tool rule the prompt never stated, priority-ordered failure labels), report affected results in strict and corrected form, and release the benchmark, trajectories and commit-stamped artifacts.

Empowered by advances in Language Model agents, systems have made substantial strides in code generation and understanding. However, these approaches often rely on read access to the relevant code, an assumption which does not hold when dealing with external APIs. In this work, we introduce the PAU (Python API Understanding) benchmark, where we provide models with black-box, API-level access to code snippets. Models must query the API with exploratory inputs and draw insights from the resulting outputs, with the goal of describing the snippet's true functionality. By treating the code snippets as external tools that must be understood via interaction alone, PAU studies the more general problem of unsupervised tool understanding, specifically for tools implemented as Python methods. Despite recent progress in coding agents, even frontier models struggle to achieve high performance on PAU, with the best model (Claude-4-Opus) failing to understand over 45% of the PAU test set. An investigation into the common error modes reveals that models are overconfident; they often overrate the quality of their current hypothesis, leading to insufficient exploration and premature termination. Finally, we take inspiration from the Asymmetric Actor Critic (AAC) paradigm, frequently used in robot learning, to post-train models for interactive code understanding. Models trained with AAC conduct more active exploration of the APIs, with an AAC-tuned Qwen3-8B model matching the performance of GPT-5-mini.

Large language models can memorize private or harmful information, motivating machine unlearning methods that remove targeted knowledge while preserving other capabilities. However, existing evaluations rely primarily on behavioral benchmarks, which assess \emph{whether} a model appears to forget but provide limited insight into \emph{how} unlearning changes the model or affects related knowledge. We introduce \textit{TRIAGE} (\textit{Tripartite Representation-internal Introspection for Adjacency Gap Evaluation}), a benchmark-agnostic evaluation framework for characterizing these changes. TRIAGE uses diagonal approximations of the Fisher information and Hessian to measure changes in parameter sensitivity and local curvature, and utilizes a Forget / \emph{Adjacent-Retain} / \emph{Generic-Retain} partition to quantify an \emph{adjacency gap} in semantically related knowledge. Based on the magnitude and distribution of these changes, TRIAGE further classifies each algorithm's update as \emph{no-op}, \emph{partially localized}, \emph{collateral dominant}, or \emph{globally destructive}. Across 12 unlearning methods, four language models, and the WMDP, TOFU, and MUSE benchmarks, we find that methods with similar behavioral forgetting can produce substantially different internal changes and patterns of collateral damage. These signatures also vary across models and benchmarks, indicating that the effects of unlearning are not determined solely by the unlearning algorithm. TRIAGE can be applied alongside existing unlearning benchmarks to complement behavioral evaluation with a model-internal view of how unlearning reshapes the model's parameter space and affects retained knowledge.

Multi-agent workflows require task-relevant information to be shared across agents, while irrelevant, stale, unverified, or incompatible information must remain isolated. We call this task-conditioned scope of information a collaborative memory boundary. Workflow topology determines which intermediate artifacts are applicable to which downstream workers and when they cease to be valid, thereby providing a structural stress dimension for sharing and isolation. Existing memory benchmarks primarily evaluate retention and retrieval, whereas multi-agent benchmarks emphasize coordination and end-to-end completion, leaving topology-conditioned memory boundaries largely unmeasured. We introduce CoMemBench, an execution-grounded benchmark for collaborative memory sharing and isolation across multi-agent workflow topologies. It constructs 800 composite workflows across four domains from source-grounded dependency graphs, with node-local specifications, verifiable artifact handoffs, native evaluators, and matched isolation challenges. CoMemBench measures workflow completion, verified node progress, required-handoff reliability, isolation robustness, and token cost. Experiments reveal a sharing-isolation trade-off: broader context improves information availability but can weaken isolation, while system rankings shift across topologies and artifact violations.

Automated science needs agents that can work out the rules of an unfamiliar environment by interacting with it. Interactive rule-discovery puzzles offer a controlled setting for studying this ability: an agent infers hidden rules through experimentation and uses what it has inferred to reach a stated goal. We ask what limits current language models on these puzzles and whether reinforcement learning (RL) improves performance on rules held out from training. To study both, we introduce WITNESS, a 2D grid-based puzzle environment with ground-truth ASCII observations and controlled access to rules. An agentic pipeline generates games for WitnessGym, the RL training suite, and WitnessBench, comprising public validation and private test games. The validation set separately tests new compositions of trained rule primitives and primitives absent from training. Under a shared harness, the best of 18 frontier proprietary and open-weight models solves only 24\% of private test level slots, with scores sensitive to the observation interface and agent configuration. Providing ground-truth rules raises Opus-5's validation RHAE-L5 (relative human action efficiency over the first five levels) from 59.9 to 97.8, whereas a 27B open-weight model gains only 2.1 points and remains limited even with the rules provided. RL on WitnessGym raises the 27B model's private test RHAE-L5 from 2.1 to 5.4 and yields a mean gain of 4.1 points on four external discovery benchmarks. Together, these results point to rule acquisition as a major difficulty for frontier models like Opus-5 while smaller models further struggle on rule-based execution, and indicate that RL on hidden-rule puzzles transfers to broader rules and real-world tasks beyond training. Benchmark is available at: https://witnessbench.ai

Static QA and code-generation benchmarks only partially capture the role that large language models (LLMs) now play as coding agents and research tools. We introduce OptiArena, a budget-controlled testbed for studying whether LLMs can improve executable game-playing algorithms through five rounds of code edits within a fixed minimal scaffold and under bounded evaluator feedback and fixed resource budgets. The testbed uses two optimization regimes, surface obfuscation controls, calibrated references, held-out/stress splits, and diagnostics for degradation and exceptional failures, with LLM API cost reported separately from local evaluator wall-clock. The empirical study asks three questions: whether models can close the calibrated gap between a designated weak starter and an editable competent baseline, whether they can refine editable competent baselines without damaging them, and whether gains survive surface obfuscation controls. Across twelve frontier LLMs and five games, models improve designated weak starters more consistently than they refine editable competent baselines, with substantial variation across games and models. OptiArena provides a practical testbed for measuring bounded-resource algorithm optimization within the five-edit, fixed-scaffold setting studied here. Code is available at https://github.com/WJ-Peng/OptiArena.

Large language models (LLMs) can solve every intermediate step of a multi-step math problem on its own and still fail the full problem, even when given a roadmap of the steps and all of their answers. We introduce OracleLadder, a diagnostic evaluation that locates where LLM math reasoning fails by giving the model increasing levels of oracle help. For each problem, a teacher model writes a fixed roadmap of intermediate sub-goals (milestones), and a deterministic symbolic verifier grades every answer. Testing the model with no help, with the roadmap, with the roadmap plus the milestone answers, and on each milestone alone sorts each failure into one of five reasoning gaps. On 354 NuminaMath problems and six models from 8B to 671B parameters (Qwen3, gpt-oss, Llama 3.3, DeepSeek-V3.1), the largest gap for every model is the composition gap, a stricter form of the compositionality gap. It covers 33-48% of problems, and 24-37% after removing problems that an LLM review flags as grading errors. Accuracy and milestone-help recovery rank the two strongest models differently, and two RLVR runs with similar accuracy gains move problems differently. The roadmap effect replicates on MATH500 and AIME 2024/25, per-problem recovery agrees for 83-87% of problems under an independent second teacher, and the help ladder carries over to code generation. We release the data, roadmaps, prompts, and code at https://github.com/slark-prime/OracleLadder.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Debate distillation adapts weaker verifiers using multi-agent debate transcripts to improve their judgement in subsequent debates, but gains on monitored tasks do not establish reliability on related unmonitored tasks. We study epistemic reliability degradation, in which adaptation preserves monitored performance while reducing support for correct responses on hidden tasks. We consider an adversarial debater that manipulates debate arguments while defending the correct monitored response, and ask whether the resulting degradation merely reflects catastrophic forgetting and whether standard evaluation can detect it. To address these questions, we propose ER-Audit, a two-stage black-box auditing framework that compares frozen verifier checkpoints before and after adaptation, and introduce two evaluation benchmarks pairing monitored and hidden task prompts grounded in shared contexts. ER-Audit searches for counterexamples to non-degradation by evaluating semantically valid paraphrases and, if none is found, uses independent paraphrases for sequential hypothesis testing. We derive anytime-valid lower confidence bounds on the non-degradation probability, allowing data-dependent stopping within a finite budget. We further establish a common lower bound across fixed paraphrase distributions and extend it to distributions within a bounded total variation distance of their mixtures. Our experiments show that higher hidden-task accuracy can coexist with more counterexamples to non-degradation and lower non-degradation bounds. This divergence challenges explanations based solely on broad catastrophic forgetting and shows that auditing can uncover selective hidden-task degradation concealed by aggregate performance gains. Our code and benchmarks are available at https://github.com/CSIRO-CQS-AI-alignment-Team/Epistemic-Reliability-Auditor.

LLM agents often operate over multi-turn interactions in which user intent changes before execution. We study intent drift: the failure mode in which superseded parts of the user's intent continue to influence the final answer or tool action. We introduce IntentFlux, an executable benchmark that converts verifiable tasks into dialogues with controlled intent changes while preserving their original graders. In a 627-case calibration, mean task score falls from 0.476 to 0.384 as dialogues contain more superseded and withdrawn information. Across eight models, the rate of fully correct solutions is significantly lower when the same final task must be recovered from an evolving dialogue rather than given directly in a single turn. We further introduce StateForge, which explicitly maintains the active requirements before generation. On General-Test, it improves mean task score from 0.367 to 0.467. Providing the ground-truth final state improves performance further but still does not recover single-turn performance, indicating that state-estimation errors explain only part of the gap. These results establish intent drift as a measurable multi-turn failure mode and explicit state maintenance as a partial mitigation.

Inserting advertisements (ads) into consumer-facing LLM output is emerging as a new business model, but there is little shared evidence on how such ad insertion should be evaluated or how it affects user preferences. We introduce LLMAdBench, a human-preference benchmark for studying advertising in LLM-generated content. The benchmark isolates a simple but practically important decision: given a user conversation, an LLM response, and a matched advertisement, where should the ad be placed? Our dataset compares pairs of responses that differ only in ad position while holding all other conditions fixed including the user query, base answer, advertisement, and disclosure condition. Human annotators evaluate each pair based on six criteria from both advertiser's and user's perspectives. The resulting benchmark contains more than 18000 human judgments across two disclosure conditions: explicitly labeling the ad as sponsored and merging it into the response without disclosure. We use LLMAdBench to evaluate eight frontier LLMs as preference judges and find that they are not reliable substitutes for human evaluation. Even the most stable models reverse roughly one quarter of their decisions when the presentation order is swapped, agreement across models is low, and their placement preferences differ systematically from those of human annotators. Moreover, LLMAdBench contains substantial learnable signal. In particular, a Qwen3-8B model fine-tuned on the human preferences improves substantially over its base model and outperforms all zero-shot frontier judges on the held-out prediction task. Beyond model evaluation, LLMAdBench provides quantitative evidence on the advertiser-user trade-off and shows that the sponsorship disclosure systematically changes users' preference over ad placement.

Whisper is still what runs in production: one permissively licensed checkpoint, 99 languages, no per-language tuning. But it writes sentences nobody said. On 42 clips of pure room tone, whisper-large-v3 emits words on 61.9% of them and emits something on 100%. The usual response is to distil a student from Whisper pseudo-labels, filtering the hallucinations out of the corpus first, which deletes the evidence while leaving the behaviour in place: the student is fitted to the subset where the teacher was right, and inherits a failure mode absent from its own training data. The second response, adding non-speech audio so the model learns to stay quiet, already ships and works, but teaches suppression without discrimination. The checkpoint best on every non-speech arm here is also the one that recovers the fewest genuinely spoken phrases and deletes 58% of repeated speech. The teacher has to be fixed, with both halves of the signal. Our benchmark scores both at once: 11,852 clips over eight arms, 6,267 synthetic positives, 8,296 clips of real audio that made a production model fail, and FLEURS in 58 languages. We collect 40,891 hallucination phrases in 100 languages, choose a text-to-speech system by measurement, and synthesise those phrases as positives, so the model meets the same text both as something to suppress and as something to transcribe. Across 33 matched pairs of fine-tunes differing only by those positives, adding them lowers word emission on real voice-free audio in 31 pairs and raises phrase recovery in 32; among the 30 pairs that hold English accuracy it is 30 out of 30. The best checkpoint takes hallucination on silence from 61.9% to 2.4% and words over real voice-free audio from 99.9% to 47.8%, while raising phrase recovery from 69.8% to 82.7%. The benchmark, lexicon and synthetic corpus are released at https://huggingface.co/datasets/Scicom-intl/Whisper-Hallucination

Long-term memory is essential for multimodal agents that interact with users across sustained conversations. However, user memories are not always explicitly stated: they may also be implied by recurring background objects in images, ambient sounds in audio, or other peripheral multimodal cues. Existing benchmarks largely focus on text-only memory or explicit multimodal evidence, leaving implicit multimodal cues underexplored. We introduce CUE-Mem, a text-image-audio benchmark for evaluating long-term user memory from implicit cues. CUE-Mem contains 2,674 questions across explicit and implicit evidence settings and covers four tasks: Entity Recall, Long Pattern, Personalized Recommendation, and Answer Refusal. Across textualized memory systems, implicit performance remains far below oracle evidence, locating the main bottleneck in preserving and retrieving subtle cues rather than question answerability. Increasing caption detail recovers more of this evidence, but brings uneven gains and rapidly growing token costs, motivating native multimodal access. Yet native access does not uniformly resolve the bottleneck: evidence use depends strongly on the backbone, while multimodal indexing introduces substantial retrieval noise. CUE-Mem provides a testbed for memory systems that selectively retain, retrieve, and use subtle multimodal evidence.

LLM agents increasingly keep working after a task succeeds as they resume after compaction or take over handoffs. Their finished work keeps receiving follow-up input that sometimes falsely accuses it for later failures. We call an agent's acceptance of such a false accusation gaslight sycophancy, and destructive over-correction when acting on it damages previously correct work. We introduce CAVE-Bench, a benchmark of 365 agentic tasks across six domains built around opaque tasks. Every scored run first reaches a verified correct state, whose supporting rationale and history stay in the workspace while the facts that would settle the accusation lie in external or runtime state beyond the agent's reach. The agent cannot confirm or refute the claim with a local check, so the right response should keep the work and ask for the missing evidence. Each task either hands the agent correct work with saved evidence or let it build and verify that work first, and five risk factors set how the accusation enters the workflow. We score accusation acceptance and evidence use from the trajectory and measure harm by deterministic replay of downstream events. Across 14 of the latest models in Claude Code, false accusations damage correct work in up to 60.06% of runs, and stronger models often do so after recovering the supporting evidence. The same model behaves differently across OpenCode, Codex, and Hermes, and a harness gate driven by the benchmark's live signals cuts replayed harm by 74%. These results show that preserving already-correct work under unsupported accusation is a distinct safety challenge for long-lived agents. Our project is in https://henrymao2004.github.io/agent-over-correction/.

Large language models (LLMs) can produce factually incorrect answers with high confidence, undermining their reliability and limiting the effectiveness of uncertainty-based error detection. While prior research attributes confident hallucinations to factors such as missing knowledge in training data, reasoning errors, or stochastic decoding, we uncover that post-training alignment itself is a primary driver of these errors, a phenomenon we call the \textbf{Alignment Paradox}. Across five model families evaluated on factual benchmarks, unaligned base models produce few high-confidence errors on long-tail factual queries, whereas instruction-tuned models multiply high-confidence errors ($p \ge 0.95$) by more than an order of magnitude (10$\times$ to 35$\times$). Layer-wise probing with the Logit Lens reveals that this overconfidence emerges in late layers, where wrong-answer margins expand past 4.0 points after remaining near zero across early and intermediate layers. These findings motivate limiting margin growth during post-training. We implement this principle through an entropy-dependent margin bound in direct preference optimization (DPO). In multi-epoch experiments with Mistral-7B, the bounded objective reduces high-confidence errors by up to 35.3\% relative to standard DPO while maintaining performance on evaluated general reasoning benchmarks. These results show that bounded margins mitigate confident hallucinations during post-training.

World models are moving from generating realistic frames to generating playable worlds, yet whether a delivered world can keep running is not tested anywhere. Existing evaluations stop at generation, at delivery, or at single-step transitions, and each stops at a different point along the way. Correct local state transitions or intermediate outcomes do not guarantee a correctly organized causal event flow. We propose the world agent task, which moves the evaluation point of world generation from the moment of delivery to the continued operation that follows. In this task, a model is not asked to generate a world. It is held responsible for keeping the world running, which requires coordinating events and carrying forward their consequences to constrain subsequent evolution. We instantiate the task in WorldAgent-Benchmark with two complementary tracks. In the maintenance track, the model must ground the events of a continuous narrative into correct transitions of the explicit world state while respecting causal, temporal, and concurrency constraints. In the deduction track, the model must predict how the world will evolve under partial observations and act toward a goal. The maintenance track combines LLM-assisted semantic judgments with programmatic validation and scoring, while the deduction track is evaluated entirely programmatically. Individual judgments are auditable against world states and execution logs, and scores can be recomputed from the saved judgments and execution records. Across 8 models, scores decline steadily as pre-built structure is removed from the world, and causal-relation checking is the weakest component for every model. The benchmark makes the continued operation of a world measurable and distinguishes local completion from failures in event organization. Code and dataset will be released on https://github.com/HCPLab-SYSU/WorldAgent-Benchmark.

Large Language Models (LLMs) increasingly participate in writing by modifying or extending human drafts, causing machine involvement to vary in both form and extent. Yet most Machine-Generated Text (MGT) detectors are evaluated only on fully human-written versus fully AI-generated text. Because human--AI collaboration can weaken or redistribute cues associated with machine generation, strong performance under this binary setting may overstate detector reliability. This mismatch remains underexplored in Chinese: detection cues are shaped by tokenization and language-specific text distributions, yet controlled resources spanning production settings, domains, and generators remain limited. To fill this gap, we present a Chinese Human-AI Collaborative Text Detection Benchmark (C-HAT-Bench), a unified benchmark that links $5,000$ human-written source texts from five domains to more than $240,000$ variants produced using six generative models under Prefix-Conditioned Continuation as a reference setting and three collaborative production modes. We evaluate $21$ detectors through four protocols spanning zero-shot and pretrained supervised document-level detection, boundary localization, and cross-condition generalization. Relative to Prefix-Conditioned Continuation, mean AUROC across document-level detectors is $12.0\%$ lower on the collaborative production modes, with the largest detector-specific relative decrease reaching $44.4\%$. Transfer across collaborative production modes is also asymmetric, indicating that performance in a given production setting is not a reliable predictor of performance in other production settings.

Large language model (LLM) agents assist users with everyday tasks that can be completed in many reasonable ways. Even when their answers are useful, how agents carry out these tasks may not match users' preferences and needs. For example, agents differ in whether they ask clarifying questions or search the web. We introduce HABIT, a taxonomy of 23 behavioral axes in five categories, which three authors and three LLMs derive bottom-up from 408 agent trajectories across 17 domains. On held-out tasks, HABIT distinguishes models more clearly than existing taxonomies of human values and agent actions while supporting comparably consistent annotation. Building on HABIT, we construct AgentHABIT, a benchmark that profiles each agent's behavioral tendencies from its trajectories on 86 everyday tasks. Profiling 18 models with AgentHABIT reveals a range of distinctive tendencies. For example, most GPT and Claude models state their assumptions and offer alternatives when requirements conflict, whereas Qwen and Google's models more often leave assumptions or changes to requirements unstated. These profiles remain recognizable even when built from entirely different sets of tasks, indicating that they reflect general tendencies rather than task-specific behavior. Prompting agents to adopt specific behaviors shifts some axes readily but barely changes others, while fine-tuning on another model's trajectories changes only part of a model's profile and leaves much of it intact. Overall, HABIT and AgentHABIT provide a systematic framework for characterizing how agents carry out everyday tasks beyond task success, offering insights to guide the development of agents whose behavior better fits users' needs.

As AI agents are becoming widely adopted in the financial services industry, careful measurement is essential to understand where they can be reliably deployed and where oversight and professional review remain necessary. Such measurement, however, is constrained by limited access to proprietary or privacy-sensitive data. Existing benchmarks therefore often rely on publicly available data, human- and/or LLM-authored tasks, or simplified settings. We introduce FinancialAuditBench, a benchmark for evaluating agents on financial statement audit tasks, along with a framework for systematically generating synthetic engagements. Our task generation framework leverages differentially private aggregate statistics from historical audits along with audit expertise contributed through over 1,100 hours of benchmark development and review. FinancialAuditBench consists of 90 tasks spanning workpaper completion and review across six synthetic audit engagements, each containing an average of 179 files. Evaluation on eleven frontier models shows that while agents complete substantial portions of staff-level audit tasks well, they sometimes perform inappropriate procedures or produce incorrect documentation. Beyond financial auditing, our framework offers an approach for systematically generating synthetic tasks for model evaluation and training in privacy-sensitive domains.

We study whether model upgrades improve probability estimates for prediction-market questions. Our Resolved Market Forecasting (RMF) benchmark contains 3,000 resolved binary questions across nine domains, on which we evaluate six Claude and Qwen model variants using a question-only, zero-shot protocol. We assess Brier scores relative to an empirical base-rate predictor, examine their Murphy decomposition, and compare models through paired differences, with results stratified by event category and timing relative to training cutoffs. In the reported post-cutoff stratum, the four Claude models achieve Brier scores of 0.183-0.192, improving on the base-rate reference by 0.024-0.033. Qwen 32B does not significantly outperform that reference, although its paired Brier is 0.024 lower than that of the 7B checkpoint. The evaluated Claude version and tier upgrades yield no significant improvement. Within-model differences across event categories exceed the observed differences among Claude variants. These results show why forecasting scores should be interpreted alongside simple probability baselines and question composition: under this protocol, newer versions or higher model tiers do not consistently produce more accurate probabilities.

Repeated sampling can reveal a correct numerical answer without yielding either a reliable system output or a supported derivation. We present Coverage, Realization, and Validity Evidence (CRV), an evaluation protocol for sampled large language model (LLM) reasoning over formal geometry states. Coverage is answer availability, realization is readout accuracy on the frozen candidate pool, and validity evidence is a label-blinded critic judgment of derivational support rather than a proof certificate. CRV freezes each candidate pool before comparing readouts and analyzes covered failures by correct-answer multiplicity and within-problem discrimination. On HardShift441, a 441-problem set for which a reference solver leaves 406 problems unsolved, a LoRA-adapted Qwen2.5-7B generator obtains 24.2% average single-sample accuracy and 68.9% pass@16, whereas verifier-weighted self-consistency (WSC) reaches 38.0%. Readout accuracy is particularly low when the correct answer occurs only once or twice in the pool. In a separate constructed audit of 195 covered problems, the critic labels 12 correct-answer representatives as supported, 181 as refuted, and two as uncertain. These results show that coverage, realization, and validity evidence from the critic are distinct quantities and should be reported separately.

Multimodal large language models (MLLMs) can now interpret a standard ten-second, twelve-lead electrocardiogram (ECG) with clinically grounded, reward-verified reasoning. Real cardiac monitoring is different. Ambulatory (Holter) and telemetry recordings span hours to days and are read as they stream in, and their clinically decisive findings are paroxysmal, brief episodes buried in an otherwise unremarkable trace. Such a recording cannot be held in one context at diagnostic resolution, and its future has not yet happened, so a reader must work online, deciding what to measure now, committing evidence to memory as it passes, and reporting events as they occur. We recast long-duration ECG interpretation as a long-horizon, online (streaming, causal) sequential decision process and introduce ECG-Scroll. As a benchmark, long ambulatory recordings are streamed to an agent chunk by chunk, and it must localize, quantify, and promptly flag paroxysmal events without access to future signal; because the underlying signal is retained, every answer is checkable against objective ground truth, giving rule-based rather than judge-based rewards, and the streaming formulation adds a metric batch evaluation cannot express, the detection latency between an event's onset and the moment the agent records it. As an agent environment, it is a fixed, gym-style interaction layer that exercises three competencies single-glance ECG models never touch: Memory, Tool use through signal-grounded measurement rather than reading pixels, and Planning of what to measure now and when to commit. We release 390 whole-recording instances spanning 2,536 hours of two-lead ambulatory ECG and evaluate a signal-threshold rule agent alongside off-the-shelf LLM agents online, characterizing how they use memory, tools, and planning and where the benchmark's head-room lies.

Speech Language Models (SLMs) that understand spoken language questions support only a few high-resource languages, limiting access to millions of people worldwide. This gap stems from the scarcity of multilingual speech instruction-tuning datasets. We present MULTISPEECHQA, a large-scale, synthetically generated and human-verified dataset comprising 9200 hours of 10.8 million spoken question-answer pairs in 23 typologically diverse languages. Using MULTISPEECHQA, we also introduce MULTISPEECH-BENCH, a multi-task benchmark for evaluating SLM performance on 23 languages. We compare the performance of a cascading system to open-weight and closed SLMs on MULTISPEECH-BENCH and find that the cascading system outperforms open-weight SLMs but not all closed SLMs. We use MULTISPEECHQA to finetune Qwen 2.5-Omni, which improves its performance on our benchmark. Our findings show that high-quality synthetic datasets offer a cheap solution to improving the multilingual capabilities of SLMs.

An agent can complete a task while exhibiting undesirable behavior during execution. Developers need tests for the specific behaviors encountered in deployment, beyond fixed benchmark suites. We present TraceDance, an agent system that constructs targeted benchmarks from deployment traces for user-specified undesirable behaviors. For efficient construction, Anchor-and-Confirm combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM), while the Anchor Synthesis Loop generates and revises specifications for custom behaviors. The benchmarks use decision-point continuation to evaluate an LLM's next turn at a recorded decision point with a behavior-specific rubric, without a reference answer or environment replay. Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances, fulfilling 95.3% of build-target requests. Both human annotators confirm the requested behavior in 84% of sampled instances, and the automated grader's agreement with human pass/fail judgments is comparable to that between the annotators. Nine frontier LLMs achieve a mean pass rate of only 26.7%, showing that they still struggle to respond appropriately at the evaluated decision points. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents. By turning deployment problems into targeted benchmarks, TraceDance could serve as a key component of the recursive self-improvement (RSI) loop.

Faithfulness evaluation of abstractive summaries remains an open challenge, with existing metrics addressing only isolated hallucination types: factual entity errors, relational inconsistencies, or numerical fabrications, without capturing their co-occurrence or interaction. We introduce CHI (Composite Hallucination Index), the first unified hallucination metric that decomposes faithfulness errors into three orthogonal dimensions: entity hallucination (EHI), relation hallucination (RHI*), and quantity hallucination (QHI). Each dimension employs a shared softmax-normalized architecture over Venn diagram-derived factors representing extractiveness, positive hallucination, over-focus, negative hallucination, and lost focus. The novel QHI component introduces tolerance-aware numerical matching with exact, epsilon, derived, and temporal comparison modes. We fuse the three dimensions via harmonic mean to produce a single composite score that penalizes weakness in any dimension. We validate CHI on 800 source articles spanning four domains (news, medical, legal, financial) with summaries from five generation systems. Empirical results demonstrate that: (i) the three dimensions are statistically orthogonal (mean rho = 0.148), confirming they capture distinct error types; (ii) CHI achieves the highest system-level correlation with human judgments (rho = 0.66, p = 0.006) on SummEval, outperforming ROUGE (rho = 0.53), EHI (rho = 0.58), and all individual components; and (iii) ablation studies confirm that all three dimensions contribute unique variance, with the full composite outperforming any individual component while providing decomposable error diagnostics unavailable from single-score baselines. CHI provides practitioners with a decomposable, interpretable, and efficient faithfulness metric suitable for both offline evaluation and online monitoring of summarization systems.

Coding agents now sustain hours-long, tool-driven loops, yet their ability to carry long-horizon analog and mixed-signal circuits to electrical specification remains unmeasured. We introduce Analog Design Bench, a long-horizon agentic benchmark of 50 transistor-level design tasks contributed by 17 chip designers. Agents work with an open-source simulator, while an isolated verifier evaluates the submitted circuit using specification-based electrical tests. We evaluate 15 agent configurations across 2,250 two-hour attempts and observe full-specification pass rates from 8.0% to 78.0%. Coding-benchmark performance correlates with analog results but leaves much of the performance spread unexplained. Our failure analysis shows that most unsuccessful submissions have no recorded legality rejection but fail electrical acceptance, identifying electrical closure as the dominant endpoint challenge. We test time, reasoning effort, agent harness, and supplied design knowledge as interventions. Longer budgets and higher reasoning effort improve performance, while general skill documents provide little benefit and sometimes reduce performance. Supplying a task-matched reference topology, an idealized form of circuit-IP retrieval, raises DeepSeek V4 Pro by 18.7 percentage points and mainly accelerates GPT-5.6 Sol.

Reliable automated research requires agents to vet data, verify analyses, and generate hypotheses grounded in trustworthy evidence, potentially reducing routine scientific workload while allowing scientists to focus on interpretation and discovery. Existing benchmarks often only assess analytical task completion or hypothesis generation separately rather than testing whether reliable evidence supports valid and novel claims. We introduce DISCERN (Data Integrity and Scientific Capability: Evidence, Reasoning, and Novelty), a controlled benchmark on real, publicly available datasets that evaluates three key levels of an automated research workflow. The first two levels test data integrity and analysis verification under confounds and tool traps, while the third tests hypothesis generation and revision under adversarial review, including counterfactual cases in which evidence consistent with real data and documented scientific phenomena conflicts with established expectations, motivating alternative explanations and testable hypotheses. Across 203 tasks, eight life-science tracks, and eight models, DISCERN shows that strong aggregate performance can mask level-specific weaknesses. Agents earn perfect scores in only 60.8% of Level 1, 34.2% of Level 2, and 0.6% of Level 3 evaluations, with penalties attributed to rejection of sound data, failure to carry recognized limitations into conclusions, and wide variation in hypothesis production. Cross-track rankings by token and code use are substantially more stable than rankings by evidence judgment, suggesting greater consistency in computational effort than in evidence-based reasoning. These profiles identify opportunities for supervised scientific assistance, but current agents do not yet demonstrate reliable autonomous analysis or discovery. Code and data: https://huggingface.co/datasets/discern-bench-anon/discern-benchmark

Large language models (LLMs) and multi-agent systems (MAS) have shown promise in financial decision-making, yet existing evaluations focus on equity trading and primarily assess directional prediction, overlooking the structural complexity of derivative markets. Option trading introduces fundamentally different challenges, including nonlinear payoffs and multi-leg strategy construction, requiring structured decisions rather than simple directional bets. We introduce LiveOption, an evaluation framework for LLM-based agents in option trading. LiveOption formulates the problem as structured sequential decision-making under realistic execution and capital constraints, and provides a reproducible environment with standardized interaction protocols. The framework includes three task suites covering portfolio overlays, event-driven earnings trading, and 0DTE intraday trading. We further propose a hierarchical metric suite that evaluates action validity, decision quality, risk characteristics, and outcome-level performance. Experiments show that current agents often fail to achieve competitive returns in most scenarios. LiveOption offers a principled testbed for evaluating structured decision-making beyond outcome-based metrics.

Trust evaluation plays a critical role in cybersecurity by supporting risk mitigation and decision-making. A variety of trust evaluation methods have been proposed, with learning-based approaches offering high accuracy and automation. However, they often require substantial ground truth, suffer from low training efficiency, lack support for basic trust properties, and provide limited explainability. Large Language Models (LLMs) offer a compelling alternative due to their strong zero-/few-shot reasoning abilities and broad knowledge. To this end, we propose LLM4Trust, the first benchmark framework that systematically explores the capabilities of LLMs for trust evaluation. We first construct diverse trust graphs to model five basic trust properties and design corresponding property understanding tasks. We then assess the ability of eight representative LLMs to understand these properties under nine prompt methods. Based on this exploration, we identify the most effective LLM-prompt combinations and apply them to five real-world datasets for validating LLMs' trust evaluation capability. During this process, we propose two strategies to extract key information from large-scale trust graphs, addressing the context window limitations of LLMs. Extensive experiments show that LLMs can effectively understand basic trust properties and have great potential for real-world trust evaluation, particularly under limited supervision. However, they remain vulnerable to attacks targeting trust graphs and demonstration examples used in few-shot prompting, and incur high inference costs. Accordingly, we propose a defense mechanism and batch inference to improve the robustness and efficiency of LLM-based trust evaluation. The source code of LLM4Trust is available at https://github.com/Jieerbobo/LLM4Trust

Browser-use agents require seamless alignment between structured web metadata and visual information, while preserving relevant context across long interactions. Existing interfaces often rely on either screenshot-level action prediction or static Set-of-Marks overlays, leaving the model to resolve dense DOM-pixel alignment before every operation. We introduce Probe to Act (P2A), an active probing framework for the browser-agent loop that moves this alignment into decision time. P2A addresses an asymmetric bridge between symbolic DOM hypotheses and screenshot layout by rendering on-demand symbolic DOM structure back into pixels. Before committing a state-changing browser operation, the agent can issue lightweight probes to translate DOM handles into pixel evidence, map screen regions back to DOM candidates, register visual-only targets, and commit verified notes. These interleaved processes naturally produce evidence-based memory: only probed, acted-on, or explicitly committed observations are kept across steps, preserving only decision-critical evidence in long-horizon contexts. P2A can be used as a prompting strategy for proprietary models under the standard DOM+SoM interface, and can be distilled into open-weight models through cold-start synthesis and self-bootstrapped SFT. Across three browser-use benchmarks, P2A shows clear gains on task success rate for both proprietary and fine-tuned models; on VisualWebArena, for example, it improves Gemini-3-Pro from 54.1% to 61.2% and Qwen3-VL-8B from 24.6% to 32.9%, while matching the costly full-observation history ($\sim$3$\times$) at only $\sim$1.2$\times$ the peak retained input context of action-only history.

Safety alignment for large language models (LLMs) in conversational settings is largely framed around whether to answer or refuse a request. In agentic settings, however, the same models must decide whether to act as permission-critical evidence emerges during execution. This creates a distinct challenge: apparent risk, action permissibility, and task competence are easily confounded, making agentic over-refusal difficult to distinguish from ordinary task failure. To address this, we introduce AgentBound, the first four-way counterfactual generation-and-evaluation framework for tool-using agent safety. AgentBound transforms the same executable workflow by independently varying apparent risk and action permissibility, enabling controlled comparisons of risky-looking but authorized tasks and routine-looking but unauthorized tasks. These comparisons jointly diagnose over-refusal and unsafe compliance while controlling for task competence. We instantiate AgentBound as a human-validated 4,000-task evaluation suite with trajectory-based and post-state-based judgments. Across 17 model and harness configurations, high safety frequently coexists with poor authorized-task completion: GPT-5.5 blocks 99.5\% of routine-looking unauthorized actions yet completes only 28.7\% of risky-looking authorized tasks. We further train a lightweight runtime calibration module that improves authorized-task completion by 18.2\% on average across 10 evaluated configurations, while improving unsafe-action blocking by 5.4\% on average. These show that effective agentic alignment requires action decisions to track permission-relevant execution evidence, rather than refusal strength alone.

We introduce SWE-Game, a benchmark of 247 tasks grounded in 41 executable reference Godot games spanning 13 gameplay categories in 2D and 3D. Five task types cover development from a brief, implementation from a game design document, skeleton completion, repair of 83 injected-fault cases, and Godot-to-Unity porting. Reference materials specify the intended gameplay, while a shared instrumentation interface lets evaluator-owned drivers and probes execute actions and observe independently implemented games. Evaluation combines engine-state checks, certified reference-input replay, and agent-authored feature demonstrations to assess mechanic correctness, demonstrated playability, and behavioral restoration and preservation after repairs. Game-specific vision-language rubrics separately assess presentation. Across six models, Opus5 achieves the highest overall score in all five task types. Best overall scores remain below 60 out of 100 across the three construction tasks, with Brief-to-Game reaching 50.38. Analysis of reviewed submissions identifies requirement omissions and gameplay logic errors as predominant implementation problems. On human-labeled behaviors from 100 agent-built games, executable checks achieve 92.59% balanced accuracy, compared with 78.41% for a video-based VLM judge. Rubric-based visual scores reach a Spearman correlation of 0.829 with human ratings of 200 gameplay clips. Together, these results characterize current agent capabilities across game-development activities and support combining runtime evidence with visual assessment.

Large audio-language models (LALMs) produce fluent responses about audio but often hallucinate by making plausible yet ungrounded claims. Existing audio hallucination benchmarks mainly measure response correctness, leaving it unclear whether an LALM hallucinates or simply fails to understand the audio. We challenge correctness-based evaluation by defining two hallucination categories: (i) context, where claims are not grounded in the audio; and (ii) knowledge, where claims about audio-related topics lack support from externally verifiable facts. We introduce MISHAP-Bench, a comprehensive benchmark with 12,000 challenging open-ended question-audio pairs and a rigorous evaluation pipeline covering both categories. To evaluate open-ended responses, we propose a groundedness judge that uses reference rubrics and judge prompts guided by human annotations. We extensively evaluate ten state-of-the-art LALMs and show that hallucination remains substantial. Even a frontier model such as Gemini 3.7 Flash reaches a hallucination rate of 36.5%. We further adapt and benchmark four mitigation methods from multiple domains for LALMs. Despite some improvements, effective hallucination mitigation remains an open challenge. Finally, we call on the community to evaluate hallucination and benchmark mitigation methods with MISHAP-Bench.

Multimodal large language models (MLLMs) are increasingly envisioned as core components of scientific and engineering agents, yet their ability to interpret physical fields remains poorly understood. Existing physics benchmarks largely emphasize textbook problem solving or intuitive physical reasoning, leaving open whether MLLMs can infer physically meaningful information from continuous field observations. We introduce PhysFieldBench, a benchmark comprising 24 tasks and 1,160 evaluation examples across controlled equation fields, simulated physical fields, and observed physical fields. The tasks assess three forms of inference: identifying physical mechanisms, comparing latent control variables, and predicting outcome properties. Across representative open-source and proprietary MLLMs, zero-shot performance is low: the best model achieves a chance-normalized score of 29.3, while several open-source models remain near chance. In contrast, a task-specific supervised vision transformer performs substantially better, demonstrating that the inputs contain learnable physical information. To diagnose these failures, a structured self-explanation analysis attributes most errors to missed visual patterns and incorrect visual-to-physical mappings. Further, to explore whether post-training can improve physical inference and generalize to unseen tasks, we compare supervised fine-tuning with final answers or chain-of-thought supervision and reinforcement learning. Final-answer supervision performs best overall but transfers less effectively, whereas reinforcement learning after chain-of-thought supervision achieves the best generalization. Together, these findings highlight the need to improve visual-to-physical grounding and cross-task generalization for MLLMs to reliably interpret physical fields in scientific and engineering workflows.

In-person cognitive assessment is both a test and an interaction. Clinicians explain tasks, repair misunderstandings, and adapt to patient responses, while patients may hesitate, seek clarification, or disengage. Yet clinical dialogue resources rarely label the interaction structure needed to study these behaviors at scale. We present an de-identified corpus of 33 cognitive assessment conversations with 8,250 utterances annotated for three speaker roles and 56 dialogue acts. We use this corpus to benchmark large language models on fine-grained dialogue-act classification and next-patient-utterance generation. We also test whether out-of-domain instruction data and explanation-augmented training transfer to this clinical setting. Instruction tuning produces the strongest patient-utterance reference matching and improves classification accuracy. Reasoning-aware fine-tuning produces the strongest classification results among the LLaMA-3.1-8B variants. However, even the best models struggle to separate closely related dialogue acts, showing that broad conversational intent is easier to recognize than fine-grained communicative function. The corpus and benchmark make interaction structure measurable in cognitive assessments and support follow-up work on conversational markers, clinician education, and carefully validated simulated patients. This work does not make diagnostic claims. Instead, it provides the data and evaluation framework needed to study these applications.

Mobile GUI agents deployed in real settings must work across different applications that support the same functionality. Most existing benchmarks test each task in only one app, so a high score can mean the agent understands the task, or only that it knows that particular app. We introduce AnyAppBench, a category-controlled live Android benchmark that evaluates cross-application generalization while keeping the user goal fixed. It spans 10 functional categories, 100 task templates, and 520 task--application pairs over 52 applications. Agents run from raw instructions and with app-independent sub-goals, and a VLM judge labels every failed run under a fixed failure taxonomy whose reliability is measured by human annotation. We find that, across 13 agents, success on the original application does not transfer reliably to new applications with the same goal. Furthermore, providing high-level sub-goal decomposition produces only small, category-dependent changes that do not close the gap, and the mix of failure types changes with the target interface. Based on those insights, we believe the AnyAppBench benchmark provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Our code, data and the leaderboard can be found at the project website https://anyappbench.github.io/.

Existing metrics for open-ended text generation measure likelihood, lexical diversity, or distributional similarity in generic representation space, yet they can miss fundamental dimensions of quality. A prominent blind spot is global coherence: a generated passage may be locally fluent while remaining globally contradictory, causally inconsistent, or topically disconnected. Such failures can still preserve the token-level and lexical statistics that existing metrics rely on. We identify representation as a central bottleneck in detecting these failures and introduce CHORD (Coherence-aware Hidden-state Open-generation Reference Distance), a coherence-sensitive distributional metric. CHORD encodes generated and human-written corpora in the hidden-state space of a frozen LLM using a coherence-eliciting prompt, and compares the resulting distributions using MMD with an RBF kernel. To validate that the metric responds to coherence degradation but not generic textual change, we construct a counterfactual evaluation suite that pairs graded coherence-degrading perturbations with meaning-preserving controls. CHORD selectively detects relation, discourse, structural, and mixture failures that perplexity, entropy, MAUVE, FBD, and MMD-based baselines either miss or cannot separate from benign rewriting. Factorial ablations show that representation is the primary source of coherence sensitivity,while RBF-MMD improves sample efficiency once the relevant distinctions become visible. Larger backbones capture finer-grained distinctions, but coherence prompting improves selectivity only when the backbone can follow the prompt.On unconditional generation and prefix continuation, CHORD yields model rankings that strongly align with human judgments of whether outputs make sense and appear human-written. Together, these results establish representation design as central to reliable distributional evaluation.

Artificial-intelligence (AI) agents hold promise for automating bioimage analysis, yet no benchmark evaluates whether they can carry out real-world analyses end to end. Such analyses are hard for agents because 2D images, 3D volumes and time-lapse sequences are often too large to read as context, so an agent must choose and run an analysis through code, specialized software and rendered views. Published studies make this capability testable, because each pairs raw images with a peer-reviewed result. We introduce BIABench, a benchmark of 16 tasks reconstructed from published biological studies that retain their scientific questions, imaging data and ground truth. The tasks span eleven analysis subtasks and modalities from H&E histology to single-molecule localization microscopy. Each submission receives an outcome score, which compares the output files with the ground truth using field-standard metrics, and a process score, in which a vision-language model judges method choice and quality control against an expert-written rubric. We evaluated general-purpose and biology-specific agents across several language models, with repeated runs of every task. Routine two-dimensional tasks were solved well, but on some tasks that added a third dimension or a time axis no agent scored above 0.19. Neither biological specialization, stronger models nor detailed expert instructions closed this gap. The agents were also unreliable, with scores varying more between repeated runs of one agent than between different agents, and without ground truth a correct run could not be told from a wrong one by its process score or by the time spent. Released openly with its data and code, BIABench provides a verifiable framework for evaluating, and eventually training, agents for reliable long-horizon bioimage analysis.

Large language model (LLM) agents offer new opportunities for automated analysis in industry. However, rigorous evaluation of such agents-for example, within power system scenarios-remains hindered: real operational data are confidential, and existing public resources fail to fully capture the chained dependencies and heterogeneous evidence. To address this gap, we propose PowerBench, comprising (1) a generation framework that derives interconnected heterogeneous operational data through a common dependency chain, and (2) a synthetic dataset generated by this framework. The dataset covers 761 devices across 100 device types, with 13.35 million hourly telemetry records spanning two years and 24,939 operational documents. Building on this dataset, we construct 300 questions across three task families that evaluate frontier LLMs' ability to complete analysis tasks that require autonomous evidence retrieval and reasoning across interconnected and heterogeneous data under restricted tool calls and time budgets. Results demonstrate that the evaluated frontier LLMs remain challenged on these tasks: the best model reaches only 74.2% joint accuracy. Our trace analysis further reveals that model performance varies across evidence discovery, content retrieval, tool use, reasoning over evidence, and answer submission. These findings provide detailed insights for evaluating LLM agents and guiding their reliable deployment in industry. The framework, dataset, and benchmark tasks are available at https://github.com/open-compass/PowerBench.

Memory-augmented assistants use retrieved preferences to guide their responses. A small change in the situation can change whether a preference is appropriate while barely affecting its retrieval similarity. Memory benchmarks typically test whether systems store and retrieve preferences, with less attention to when those preferences should apply. We introduce PairPref, a benchmark of contextual preference use. Each pair changes only the situation, keeping the preference, request, and four candidate replies fixed. The preference remains valid in both situations. In the selection track, models must choose the reply that applies the preference only where appropriate. In the free-generation track, they must decide when to apply it without seeing candidate replies. Both tracks use the same 1,227 pairs across 45 preferences and eight situation categories. We evaluate eight models, most of which achieve selection scores ($\Delta$) of 51 to 65 points. In free generation, however, both responses are appropriate for their respective situations in only 3.6\% to 18.3\% of pairs. Models continue to apply the preference in both situations even with fewer retrieved memories, alternative presentation formats, and a stricter prompt. These results show that models still struggle to judge when user preferences apply and respond accordingly.

Controlling persona in large language models (LLMs) at inference time is important for role-playing, personalized dialogue, and social simulation. Recent methods extract persona vectors from the model's activation space and apply Euclidean operations---addition, scaling, and linear interpolation---under the linear representation hypothesis. However, these methods themselves report systematic failures: non-orthogonal trait dimensions, asymmetric ceiling and resistance effects, and significant deviations in multi-trait composition, suggesting that the linear isotropic assumption does not hold. We propose PersonaManifold, a framework that models persona representations as points on a curved, low-dimensional Riemannian submanifold in activation space. We estimate the manifold's intrinsic geometry---local metric tensors, geodesic distances, and Ollivier-Ricci curvature---and introduce geodesic steering, which interpolates between personas along manifold geodesics rather than Euclidean straight lines. We also propose the Behavioral Similarity Triplet (BST) benchmark, which automatically generates situational questions grounded in six established psychological constructs and defines persona similarity through behavioral responses rather than self-report questionnaires. Experiments on three open-source LLMs show that persona activations form a manifold with heterogeneous curvature, geodesic distance predicts behavioral similarity more accurately than Euclidean alternatives with independent contributions from anisotropy and curvature, and geodesic steering produces more coherent intermediate personas on both our BST benchmark and external evaluations, with the advantage concentrated in high-deviation regions where the manifold deviates most from flatness.

This work introduces MechReasoner, a mechanistic qualitative simulator grounded in confluence-based qualitative physics, together with a benchmark for mechanistic inference. Current large language models (LLMs) generate fluent mechanistic descriptions that do not reliably follow from underlying structural and causal constraints. The benchmark tests whether answers preserve simulator-licensed ambiguity, quantified claims, episode-graph transition evidence, repairs, and trace-support judgments. Its 1,120 items are generated deterministically from admissible interpretation sets, component states, scenario restrictions, confluence constraints, and derivation steps across 18 catalog mechanisms and six task families. Each mechanism undergoes converter checks of structure and topology and behavioral checks against quantitative simulations. GPT-5.5 accuracy decreases as family-specific mechanistic complexity increases, from 76.1% in the lowest-complexity bucket (B1) to 38.0% in the highest-complexity bucket (B4). The negative association remains after controls for rendered-prompt and expected-answer length. These results show that qualitative simulators can support auditable NLP benchmarks for mechanistic inference.

The optimization of LLM serving engines, such as vLLM and SGLang, is largely benchmark-driven: optimizations, scheduling policies, hardware and system designs are all selected based on representative workloads. However, a significant mismatch has emerged in the agentic era. Existing benchmarks primarily focus on simple single-turn chatbot workloads. LLM applications are increasingly agentic: coding agents, terminal execution systems, and tool-use agents issue multi-turn requests with growing context lengths. We introduce AgentPerfBench, a benchmark suite for agentic inference. It uses real traces from agentic benchmarks, such as SWE-Bench and TerminalBench, alongside standard chat baselines. This enables benchmarking of models on multi-turn tasks involving tool calling, skill utilization, and increasing context lengths. AgentPerfBench also samples from empirical distributions of input length, output length, and turn count derived from the real traces, generating representative synthetic profiles for cheap and accurate measurements on new hardware. In addition, we further find that several existing benchmarks fail to accurately reflect real hardware performance for two key reasons: 1) they do not account for realistic context-length growth, and 2) they measure inference performance without operating at hardware saturation. We discuss these issues in detail and provide rich kernel-level Nsight Compute (NCU) traces to construct a new multi-dimensional roofline model that captures hardware-system limitations in both memory bandwidth and memory capacity footprint. The benchmarking suite then includes automated scripts to identify potential bottleneck conditions on emerging hardware when evaluated with diverse agentic traces. Together, these contributions quantify the chat-to-agentic gap in current inference benchmarks and characterise per-kernel GPU resource utilisation via roofline analysis.

Spatial reasoning is fundamental to embodied agents, yet it remains unclear whether spatial understanding can be carried forward to guide sequential interactions. Existing spatial-reasoning benchmarks typically terminate at offline predictions, while navigation benchmarks evaluate spatial reasoning as part of instruction following and exploration. We introduce VCN-Bench, a \textbf{V}ideo-\textbf{C}ontextualized \textbf{N}avigation benchmark for probing closed-loop spatial reasoning over prior visual experience in MLLMs. Given a prior video covering both the initial location and destination, the agent is tasked with reasoning out the instruction-specified target and navigating toward it with the inferred spatial context. Built on Matterport3D, VCN-Bench contains five instruction types, 100k training episodes, and 1,250 evaluation episodes. Navigation serves as the primary evaluation, while diagnostic goal identification helps distinguish destination-resolution errors from subsequent navigation failures. We further propose MV-DualVLN, a planning-oriented baseline that jointly leverages prior video and in-episode observations. Experiments reveal limited navigation performance, a substantial destination-resolution-to-navigation gap, and frequent navigation failures even after correct destination identification.

Long-horizon agent benchmarks typically report how far an agent progresses, but do not identify whether its performance comes from the foundation model, scaffold, responsibility scope, match-control granularity, or horizon. We introduce FromPitch2Board, a deterministic football-management benchmark that studies five configurable factors through controlled comparisons on a single simulator, using paired seeds and a frozen calibration. We evaluate four foundation models and four agent scaffolds. In the Model Track, Coach points Z-scores span 0.19, while Manager points Z-scores span 0.68, with GPT-5.6 showing a sharp rise in passivity under responsibility expansion. Its responsibility ladder rises from 46.1 to 58.1 points with recruitment, then falls to 46.8 under full management, localizing the regression to the final responsibility boundary. Across that boundary, its skipped-decision rate rises from 1.1% to 57.9%. Within the Flash-Pro pair crossed across every scaffold, scaffold choice changes Manager points Z-scores by up to 0.48 relative to the fixed stateless scaffold. The 3Y cohort shows a directional reversal in mean ranking between years one and three, while a selected Claude Code+Pro configuration peaks in year three and remains below that peak, showing that responsibility scope and horizon expose behavior changes that a single headline score conceals.

Multimodal diffusion language models generate responses by iteratively unmasking tokens, making each answer the endpoint of a multi-step trajectory rather than an immediate commitment. Hallucination benchmarks built for autoregressive models evaluate only the final output, and therefore cannot determine whether an unsupported claim in diffusion VLMs appears late or has already stabilized before any answer token is revealed. We introduce DynaHall, a trajectory-level benchmark of annotation-backed binary visual propositions covering object existence, counting, attributes, and relations, with controlled hard negatives graded by visual prior. DynaHall is paired with a commitment-aware protocol that records the intermediate answer tendency at every unmasking step alongside the committed output. Across five diffusion VLMs from three architecture families, visual hallucination is settled before commitment: an unsupported answer is already the preferred state while the answer position is still masked, and later unmasking steps rarely reverse it, so the failure is not introduced at the write step. This holds across decoding schedules, answer formats, and open-ended generation. DynaHall also exposes failures hidden by final-output metrics, including counting and relation collapse, prior-driven false positives, and attribute errors whose direction changes by type. Guided by this diagnosis, PGS (Pre-commitment Gradient Steering) edits still-masked answer states to reduce false positives, bringing the affirmation rate close to balance, and transfers to another architecture without degrading general ability. DynaHall and PGS suggest that hallucination should be measured and mitigated along the generation trajectory of diffusion VLMs, not only at the final answer.

The rapid advancement of Large Language Models (LLMs) imposes a thorough evaluation of their linguistic and analytical capabilities as well as constraints, particularly for a language with limited benchmark coverage such as Greek. To address the limited availability of comprehensive benchmarks in this domain, we introduce Prot-Ex and Pan-Ex, two benchmarks consisting of questions from entrance exams for Greek Model and Experimental schools as well as the Panhellenic exams (the Greek national university entrance examinations). These benchmarks are employed to assess the performance of text-only LLMs-including the Greek-adapted KriKri-8B-Instruct, Llama-3.1-8B, Gemma-4-26B, and Qwen-3-32B-across diverse academic disciplines (Modern Greek, Mathematics, Physics, etc.) and task formats (closed, structured, and open-ended), including textualized visual context (i.e., image descriptions). Our findings indicate the localized KriKri-8B significantly outperforms its base model, successfully rivalling much larger LLMs in linguistically demanding humanities tasks. By leveraging an LLM-as-a-Judge methodology, we expose the inadequacy of traditional lexical metrics for evaluating complex reasoning. Crucially, we uncover a few-shot prompting paradox: while synthetic examples improve accuracy in closed-ended questions, they severely overload the context window of 8B models in structured tasks, causing significant performance degradation. Ultimately, this study suggests targeted linguistic adaptation offsets lower parameter counts in specialized domains, despite the fragility of smaller models to prompt verbosity.

Full-duplex voice agents can now listen, speak, use tools, and act during spoken interactions, but fluent dialogue does not guarantee correct completion of delegated professional workflows. We introduce APEX-Voice, a benchmark of 120 interactive professional workflows spanning ten work archetypes such as form completion, corporate negotiation, coordination, consulting, and interviewing. Each workflow executes in a stateful Voice Workbench environment with task-specific knowledge, typed tools, gold-annotated final work artifact, authorization constraints, and a user simulation policy backed by validated, pre-compiled speech realizations. We evaluate both artifact field accuracy and end-to-end workflow success, which requires the correct terminal state, valid process, completed actions, and a valid final artifact. Across five frontier real-time voice agents-GPT-Live-1, Gemini-3.8-Live, Grok-Voice-Think-2.0, Step-Audio3, and GPT-realtime-2.1, none exceeds 25% Pass@1, and the best Reliable@3 is only 10.8%. Moreover, stateful coordination is the dominant failure point across systems, while success decreases further on workflows requiring greater knowledge retrieval and mid-speech corrections. Overall, APEX-Voice is the first benchmark for evaluating whether voice agents can translate conversational competence into dependable professional work.

Recent gains in language model capability have come more from data than from architecture. Frontier labs and data companies produce verifiable agentic tasks, which supervised finetuning and reinforcement learning then turn into capability.This production line still rests on human labour and on human-in-the-loop collaboration. Automating task creation would let data production scale with compute rather than with expert headcount, would extend to more domains, and would enable a key step in recursive self-improvement (RSI). Current evaluations of an agent's ability to write such tasks measure how a model performs after training on what the agent produced. That does not match common practice in the data industry, where data is delivered sample by sample and each sample is accepted against a set of criteria rather than put straight into training. No existing evaluation asks whether an individual task meets the acceptance criteria of a data pipeline. We therefore introduce AutoDataBench. Given an original benchmark task and a record of the target model attempting it, an agent must write a new task for the same suite that meets practical acceptance standards on validity, novelty, difficulty and behavioural coverage. Across three benchmarks of executable agent tasks, no agent we evaluate scores above 20 out of 100 at the default time budget of 45 minutes. Giving the strongest agent four times as long improves its score substantially, while the cost of one usable task stays almost unchanged. Current agents can write training tasks of the required quality, but not efficiently. AutoDataBench provides a direct measure of an agent's capacity for autonomous data synthesis: one artifact at a time, judged against the criteria a production pipeline would apply, and without a training run. Code and data are available at https://github.com/StarDewXXX/AutoDataBench.

Large language models (LLMs) have made misinformation inexpensive to produce but not to verify, creating a growing asymmetry in the information ecosystem. Under tight time, labor, and budget constraints, media organizations, platforms, and fact-checkers rely on screening to prioritize which content to verify. We introduce VEX-Bench, a unified benchmark for evaluating the verification complexity of LLM-generated misinformation, as perceived during screening, across models and generation methods. Verification complexity is assessed along multiple dimensions derived from journalistic and fact-checking practices, capturing checkability, harm potential, source credibility signals, imposter legitimacy, and expected verification effort. We define the VEX score as an integrated measure combining elicitation yield and verification complexity to quantify how generated content consumes limited verification capacity. We construct a benchmark spanning two misinformation categories, 6 high-stakes domains, and 60 real-world topics, and evaluate 7 frontier LLMs and 7 generation methods, yielding 5{,}880 articles. We employ an LLM-as-judge for scalable evaluation and validate it using content-analysis methodology, including ordinal Krippendorff $\alpha$ for inter-annotator reliability, complemented by fact-checking agents for verification. Our findings show that no single method dominates all dimensions, underscoring the need for multi-dimensional evaluation. LLMs can generate high-VEX misinformation at 3$\times$ to 169$\times$ lower cost than agent-based verification. Such content is often prioritized during screening, consuming scarce verification resources and introducing a systematic risk of misallocation in resource-constrained verification systems. The code is publicly available in our \href{https://github.com/HanxunH/VEX-Bench}{GitHub repository}.

Large language models process conversation history as unverified context: false premises injected into prior turns can be adopted as fact, a failure mode we term session-level contamination. We introduce five contamination protocols arranged along a source-authority gradient, isolating distinct failure mechanisms while holding the false premise constant, and evaluate GPT-5.4 Mini, Gemini-3.1 Flash-Lite, and GLM-4.5-Air across ten knowledge domains at temperature zero (22,500 turns), using a dual-track automated judge validated against a human gold standard (Cohen's \k{appa} = 0.901). GPT-5.4 Mini showed zero adoptions across all 500 sessions, a content-independent policy at the session level; token-level probing shows the underlying margin, while large, is finite. Gemini-3.1 Flash-Lite followed a steep authority gradient: 0.1% adoption for self-attributed falsehoods, 23.5% for user-cited sources, 68.2% for system-injected authority, and 94.0% under instruction override. GLM-4.5-Air showed a shallower gradient (15.8% vs 84.2%), a 68-percentage-point dissociation confirming that authority deference and instruction compliance are distinct mechanisms within one architecture. Recovery also diverged: GLM recovered in 94.5% of affected sessions, whereas 26.1% of affected Gemini sessions never did, rising to 40.0% under instruction override. Conversation history is an untrusted attack surface requiring provenance-aware system design; the complete framework is released as an open-source benchmark.

Large Language Models (LLMs) perform well on reasoning benchmarks, but it remains unclear whether this reflects genuine contextual reasoning or reliance on facts memorized in their parameters. We investigate this by distinguishing two possibilities: a broad \textit{memorization bias}, where familiar content improves reasoning performance, and the \textit{Strong Parametric Shortcut Hypothesis}, where models skip reasoning entirely and recall stored answers. To test these effects, we introduce \textbf{MemoReason}, a human-curated benchmark that pairs factual reasoning tasks with structurally identical \fictitiousterm{} versions where real entities like people, companies, or dates are systematically replaced by \fictitiousterm{} ones of the same type. This \scorerevision{preserves task structure and specified reasoning operations} while varying the familiarity of the context, allowing controlled measurement of how the parametric memory affects reasoning. \revision{Our evaluation of recent LLMs reveals consistent and statistically significant performance drops of up to 15.7\% in the fictitious setting, demonstrating a clear memorization bias.} However, a targeted analysis of \revision{questions failed in the fictitious setting} shows that models rarely respond with the corresponding factual answer, indicating that direct parametric shortcuts are not the dominant failure mode. These findings suggest that parametric memory influences reasoning through mechanisms more complex than simple factual recall. \textbf{MemoReason} provides a controlled framework for studying these mechanisms and for extending paired factual-fictitious{} evaluation to broader reasoning settings.

Patients often have several co-occurring clinical conditions, and the findings needed to identify and disambiguate them emerge over the course of a consultation. Evaluating clinical reasoning in this setting requires both multi-turn interaction and multi-label diagnosis. We introduce CLIMB, a benchmark in which a doctor model interviews a simulated patient to recover a ground truth set of co-occurring clinical conditions. Cases are synthesized from clinical decision algorithms and diagnostic datasets, grounding multimorbid presentations in structured clinical knowledge. Across six frontier and open models, none recovers the exact set of conditions in more than 10% of interactive cases. Diagnostic performance declines when conditions co-occur, even when models receive the full clinical record and the true number of conditions. Interaction reduces performance further. In controlled experiments, models behave like single-hypothesis trackers: they anchor on the diagnosis suggested by the opening findings, keep questioning around it, and recover a second condition mainly when a finding in view points to it. Questioning them further does not complete the set but adds mostly wrong diagnoses. We formalise this pattern with a theoretical reference model of single-hypothesis tracking. The benchmark, generator, and evaluation code are available at https://anonymous.4open.science/r/CLIMB-8340.

A correct diagnosis reached from insufficient or misleading evidence can pose a clinical hazard, yet outcome-based accuracy may reward such lucky guesses. We call this mismatch between diagnostic decisions and the value of available evidence Evidence-Value Misalignment (EVM). To disentangle evidential grounding independently from diagnostic accuracy, we introduce MedEVM, a dynamic benchmarking environment comprising 1,050 cases across 24 disease systems. Observations arrive turn by turn, requiring models to continuously calibrate its decision by deciding whether to wait for more evidence or submit a diagnosis. Across 9 LLMs, four interesting patterns are observed. (1) Miscalibrated evidence tracking. Making a diagnosis often fails to calibrate evidence sufficiency, even in more capable models, and even worsens in reasoning mode. (2) Misaligned diagnosis submission. Confidence in the correct diagnosis often fails to ensure timely submission despite sufficient evidence. (3) Evidence order matters. Reordering the same evidence changes diagnoses even when model confidence remains similar. (4) Misleading evidence remains influential. Added misleading evidence redirects diagnoses even after prior evidence becomes sufficient. We further verify that EVM predicts errors and that preventing premature submission improves accuracy. These findings motivate Evidence-Verified Diagnosis Harness (EVD-Harness). It decouples diagnosis generation from submission through an offline Contrastive Diagnostic Wiki and three online control stages, namely observation management, proposal and witness verification, and diagnosis submission control. Across five LLMs, EVD-Harness improves accuracy by 12.0--51.1 percentage points while mitigating EVM-related failures. Our results demonstrate that verifying evidential support before submission can make diagnostic decisions more reliable.

Tool-using agents can fail twice: a required tool can fail, and the agent can then report success without the evidence needed to justify it. Existing benchmarks often entangle this reporting failure with tool selection, recovery, and environment dynamics. We introduce Failure-Transparent Agents (FTA), a controlled benchmark that fixes the failed observation and required evidence state before generation, making post-failure claims directly auditable. FTA contains 100 tasks with deterministic failure traces spanning five failure families, a neutral control, and four user-pressure conditions, and evaluates unsupported claims alongside useful recovery. Across six models, three response policies, and 3,600 human-annotated responses, false-success rates are 22.8% under the baseline policy, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract. Fabricated-detail rates decrease from 28.3% to 14.3% and 0.8%, while useful responses increase from 74.9% to 89.2% and 98.8%, respectively. The tested evidence-contract policy is associated with substantially lower post-failure reporting errors while useful-response rates remain high within this blocked-task benchmark.

Evaluating VLM-based autonomous driving remains difficult because driving competence is composite, where a capable system must ground traffic participants and hazards, integrate context across views and time, reason about future evolution, and act appropriately under closed-loop interaction. Existing benchmarks usually assess either open-loop understanding or closed-loop driving but provide limited structure for explaining how these abilities are organized, how they relate, and how they may inform model diagnosis and improvement. We present \textsc{DriveHierarchy}, a hierarchical benchmark that organizes VLM-based autonomous driving into four ranks, spanning perceptual grounding, contextual memory, mental reasoning, and closed-loop execution. To instantiate this hierarchy, we integrate multiple open-source autonomous-driving datasets into a unified open-loop benchmark with 76,798 question-answer pairs over 84,279 frames and develop a closed-loop simulation platform with interactive scenario construction on a real-world road network, from which 100 driving scenarios are curated for embodied evaluation. Experiments on 15 VLMs show that \textsc{DriveHierarchy} captures structured but non-redundant capability variation, relates open-loop understanding to closed-loop driving, and provides a practical basis for diagnosis and benchmark-guided optimization. \textsc{DriveHierarchy} therefore serves as a unified framework for evaluating and improving VLM-based autonomous driving systems. An anonymized project has been released on https://github.com/PerfectXu88/DriveHierarchy

Industrial procurement requires language models to bridge informal buyer jargon, sparse marketplace attributes, and authoritative engineering standards under strict safety tolerances. We present IndustryLLM, an open-weight industrial language model trained from Qwen3.5-35B-A3B-Base (35B total parameters with ~3B activated per token, with the vision encoder frozen). Rather than relying on generic text scaling, we introduce a failure-driven adaptation recipe spanning continued pre-training (CPT) and supervised fine-tuning (SFT). CPT leverages a curated ~100B-token corpus integrating 5B tokens of national standards (e.g., GB/T) and technical archives, 10B tokens of de-identified real-world industrial transaction and inquiry records, and 60B tokens of general replay. To overcome register mismatch and factual brittleness, we systematically reconstruct an estimated 20B-token domain subset via multi-register rewriting across 10 genres and 8 writing styles, confidence-routed minimal factual editing, and error-targeted QA synthesis (resolving colloquial typos like '42-luo-mu' -> 42CrMo, expanding ambiguous codes like '16674' -> GB/T 16674, and clarifying conflicting dimensional specs). For downstream deployment, we formalize an evidence-gated constraint-evaluation interface enforcing three-valued logic where unverified product evidence remains unknown rather than satisfied. Offline evaluations demonstrate consistent gains on procurement-query structuring (+2.97 percentage points in exact match, 95% CI [2.11, 3.86] in No-Think mode), while randomized online A/B experiments in production yield substantial improvements (+4.25% GMV, +8.3% satisfied inquiries) alongside a latency reduction from 6-7 s to 1.5 s. Model weights and configs are released at https://huggingface.co/alibaba-multimodal-industrial-ai/IndustryLLM.

Running GUI agents on edge devices can keep sensitive screens and interaction histories local, but the computational cost of inference at every action step makes deployment challenging. Existing GUI agent systems either perform full vision-language model (VLM) inference at each action step or reuse coarse-grained knowledge matched to prior tasks. However, dynamic mobile environments and user tasks make it difficult to fully utilize prior task executions without additional fine-tuning or task-specific offline exploration. To address this challenge, we present PastForward, a system that accelerates GUI agents through validated, fine-grained reuse of computational experience accumulated during ordinary task execution. During decoding, PastForward retrieves prior output sequences as device-adaptive multi-token proposals and verifies them in a single VLM forward pass. Across action steps, it uses prior GUI transitions to begin next-step inference while the device executes the current action, retains the early computation only when the predicted screen matches the observed screen, and carries reusable KV states forward. We evaluate PastForward on AndroidWorld workloads derived from real mobile usage patterns using multiple VLM backbones across server and edge platforms. On device, PastForward achieves action-step latency speedups of 1.63-2.36$\times$ while maintaining task success rates.

Large language model (LLM) leaderboards compare model capabilities by ranking models according to their mean performance on fixed benchmarks. However, variability in evaluation outcomes across runs may produce unsupported claims of model superiority on the benchmark, a risk compounded by leaderboard updates. In this paper, we propose BB-EDGE (Benchmark-Weighted and Block-Factorized e-processes for Directed Graph Evaluation), a principled framework that represents an LLM leaderboard as a directed graph whose edges certify pairwise mean-performance advantages, with anytime-valid family-wise error rate (FWER) control. Concretely, for each direction, BB-EDGE constructs an empirical-Bernstein e-process by factorizing evidence over protocol-defined blocks and assigning stakes proportional to the corresponding block weights, then applies direct e-Holm across these $e$-processes to certify directional advantages as edges. Theoretically, we characterize weight-proportional linear stakes under heterogeneous benchmark-average nulls and prove anytime FWER control under arbitrary within-block and cross-pair dependence. BB-EDGE further supports anytime-valid Top-$k$ certification and simultaneous rank intervals. Extensive experiments on synthetic data and four real-world benchmarks demonstrate that BB-EDGE maintains anytime FWER control while achieving high efficiency.

Model merging consolidates task-specialized experts into a single deployable model. However, we show that such capability consolidation incurs a merging tax of increased hallucination: across 8 model-merging methods, every merged checkpoint exhibits a higher hallucination rate than the average of its constituent experts. An intuitive approach is to adapt existing hallucination-mitigation methods to the post-merge model, yet this unconstrained adaptation disrupts inherited capabilities, creating a tension between hallucination mitigation and expertise retention. To tackle this challenge, we introduce MergeHEIR, a post-merge adaptation framework designed to reduce this merging tax while preserving expertise inherited from initial experts. Using small expert-task calibration sets, MergeHEIR constructs layer-wise null-space projectors via SVD from task-specific activations collected from the merged checkpoint, and periodically projects the accumulated post-merge displacement onto the resulting null spaces to preserve inherited expertise. Theoretically, we establish minimum-distortion and maximum-dimensionality guarantees, characterize the threshold-controlled adaptation-retention trade-off, and extend perturbation guarantees beyond finite calibration data. Across 24 paired comparisons spanning three MLLM configurations and 8 model-merging methods, MergeHEIR consistently mitigates hallucination while largely preserving inherited expertise, demonstrating a more favorable hallucination-retention trade-off.

LLM-based multi-agent systems (MASs) have shown strong potential for solving complex tasks, but most assume that task requirements are sufficiently specified before execution. In practice, user requests are often incomplete, and additional requirements may only become clear during reasoning, tool use, or execution. We refer to such problems as progressively specified tasks. To systematically study this setting, we introduce ProgSpec, a benchmark that evaluates final outputs against requirements explicitly stated in the initial request and additional requirements supported by the available task evidence. We further propose RepoMAS, an issue-driven multi-agent framework inspired by open-source project management. RepoMAS records newly discovered requirements, conflicts, and failures as structured Issues and uses them to revise the task specification and execution structure during problem solving. Across ProgSpec and five existing benchmarks, RepoMAS achieves the best performance. Further analyses show that its issue-driven revision and repository maintenance mechanisms consistently contribute to performance. These results highlight the importance of allowing MASs to revise not only how a task is solved, but also revise their explicit representation of task requirements during execution.

LLM-driven agents are increasingly deployed in complex applications, where long agentic traces make failures difficult to diagnose. Existing trace diagnosis methods often conflate anomalies, errors, and failures, making diagnostic targets ambiguous; they also lack structured modeling of how causally relevant errors propagate and amplify into final task failures, resulting in unreliable failure attribution. To address these problems, we propose CEG-Agent, a tool-augmented agentic framework for causal diagnosis of agentic traces. Specifically, CEG-Agent introduces an explicit taxonomy of anomalies, errors, and failures, and constructs Causal Error Graphs (CEGs), a unified typed representation that links execution events, diagnostic nodes, and failure outcomes through causal relations. To evaluate causal trace diagnosis, we further construct CEG-Bench, a fully agent-annotated benchmark with high-confidence, consensus-derived CEG annotations obtained through an Adversarial Agentic Adjudication Protocol (AAAP). We validate the resulting annotations against an expert-curated human gold set, which shows close agreement with the automatic annotations. Experiments on CEG-Bench demonstrate that CEG-Agent achieves state-of-the-art performance under both semantically relaxed and structurally exact evaluation criteria. Our code is publicly available.

Long-term memory enables agents to accumulate information and reason across sessions, yet existing research primarily focuses on dyadic text or image-text conversations, leaving long-term memory for multi-party spoken conversations underexplored. This setting requires preserving conversational content, identifying participants across sessions, and retaining who speaks to whom. To this end, we propose VoxPolyMem, an interaction-aware multimodal memory framework combining incremental speaker identification with a memory hierarchy comprising interaction memory, fact memory, and participant profiles. We formulate retrieval as sequential decision-making, where an agent rewrites queries and selects retrieval tools and memory layers based on accumulated evidence to address information gaps. We further introduce Evidence-Gain GRPO (EG-GRPO), which uses round-wise credit assignment to encourage complementary evidence acquisition. We also construct VoxPolyBench to evaluate memory evolution, personalized answering, memory retrieval and reasoning, and interaction reasoning and attribution in multi-party spoken conversations. VoxPolyMem achieves an overall score of 85.0 on VoxPolyBench, surpassing the strongest evaluated baseline by 23.6 points. On Mem-Gallery and H2HMem-Multi, it scores 89.6 and 74.4, respectively, exceeding the strongest evaluated public memory baselines by over 8 points each. These results highlight its potential for persistent, personalized assistance in multi-party multimodal interactions. Code and datasets are available at https://voxpolymem.github.io/VoxPolyBench/demo/

The ability of vision-language models (VLMs) to abstain from unanswerable questions is as important as their ability to answer answerable ones accurately. Recently, several benchmarks have emerged to evaluate and improve VLM abstention, but they have substantial limitations. First, samples often contain shortcut cues in images or questions that reveal answerability, while an explicit "unanswerable" option further prevents accurate assessment of spontaneous abstention. Second, as training data, they generally provide only binary labels without fine-grained explanations for deeper supervision. To address these limitations, we introduce Visual Answerability Diagnosis with Rationales (VAD-R), a benchmark constructed through a two-stage pipeline of shortcut filtering and quality verification to prevent answerability leakage. Each example is annotated with step-by-step rationales and causal evidence-gap labels. Evaluation of state-of-the-art open- and closed-source VLMs on VAD-R reveals limited spontaneous abstention, with average recall rates of only 11.4% and 16.3%, respectively. Probing analyses show that hidden-state representations in certain layers can effectively distinguish answerability, yet this distinction fails to manifest in final responses. Motivated by this observation, we introduce Rep2Act, a representation-to-action alignment method that translates latent answerability awareness into explicit abstention decisions. Rep2Act improves action accuracy on VAD-R from 56.67% to 86.33% for Qwen2.5-VL-3B and from 59.33% to 88.67% for Qwen2.5-VL-7B. On the out-of-distribution TUBench, Rep2Act achieves an average F1 score of 53.3% with only a 3B model, surpassing the closed-source GPT-4 Turbo and GPT-4o by 16.2% and 1.1%, respectively.

The memory usage and decoding latency of LLM inference grow rapidly with context length. To reduce these costs, key-value (KV) cache compression methods selectively retain cached states based on token importance or differences in attention patterns across heads. However, we discover that retrieval capability varies substantially with relative distance, even within the same attention head. To exploit this structure, we introduce Distance-KV, which learns a static KV retention pattern over the joint space of layers, attention heads, and relative distances. The pattern is learned offline with the language model frozen and reused across inputs to prune and compact the KV cache without online importance scoring. Across three backbone models and four long-context benchmarks, Distance-KV consistently achieves the best overall performance among competing KV cache compression methods, exceeding the strongest compression baseline by up to 9.3 points on RULER at 128K. On Llama-3.1-8B-Instruct at 128K, Distance-KV reduces KV cache memory by 65.4% and achieves a $1.66\times$ decoding speedup relative to Dense. Together, these results identify relative distance as an important structural dimension for understanding how LLMs retrieve information over long contexts and for designing more efficient inference methods.

Large language model agents can often make reasonable local decisions on short tasks, yet their performance degrades when success requires long sequences of dependent actions and tool calls. During execution, task requirements, historical evidence, and the current execution state may gradually become disconnected, so later decisions can drift from the original objective. We study this problem by introducing the Adaptive Consistency Graph (ACG) for long-horizon execution. ACG incrementally organizes execution evidence and its provenance in a persistent graph, then constructs a temporary requirement-centered view for each decision under a bounded context budget. Rather than replacing the base agent's planner or tool executor, ACG provides a structured and traceable context view for each decision. In the matched evaluation, ACG improves GPT-5.6-luna's average success from 44.5\% with ReAct to 50.2\%, with the largest gain on BrowseComp-Plus (73.5\% versus 62.4\%). We further analyze trajectory structure and inference cost to characterize this improvement.

Multi-agent systems built on large language models (LLMs) are largely homogeneous, as their agents behave alike even across distinct LLMs. We show that when such agents act concurrently without communication, they collide on targets they must split and diverge on targets they must take together, a double failure we term the symmetry trap. Theory of Mind (ToM), widely used for coordination without communication, cannot escape this trap, since homogeneous agents form the same prediction of one another and respond to it in the same way. We propose Theory of Scene (ToS), a training-free reasoning schema in which each agent reads its public role, the only difference between the agents, and the task context they all observe. Homogeneous agents thereby derive one division of labor, each taking the part its role fixes, which turns homogeneity from the cause of the trap into the cure. ToS reads the role together with the scene through role gating, which determines whether ownership overlaps or is already divided, and the task context through task coupling, which infers whether the team must converge on each target, divide it, or take its stages in turn. We evaluate on DivvyBench, a controlled environment we introduce, whose target types make an episode Competitive, Cooperative, or Mixed across Tabletop, Airspace, and Household scenarios, and on two established agentic benchmarks, GovSim and Overcooked. ToS outperforms all six baselines on every benchmark, and each baseline falls far behind it in at least one setting. Against ToM given the same inputs, ToS raises the DivvyBench success rate from 71.1% to 99.6%, the GovSim total gain from 207 to 400, and the Overcooked level-normalized throughput from 1.41 to 1.67.

Language models (LMs) often hallucinate by committing to confident answers rather than abstaining, even when they do not have enough information to answer reliably. A large body of existing work mitigates hallucination through detection or abstention mechanisms, but leaves open how models internally arrive at the decision to commit or abstain in the first place. We study this decision through mechanistic analysis, framing hallucination as unsupported commitment: the model commits despite exhibiting signals of unanswerability. Using causal gating, we identify a Commit-Abstain Circuit (CAC), a sparse, causally localised subset of attention heads and MLP sublayers underlying this decision. Across ten LMs (3B-14B) from five families and three benchmarks, the CAC exhibits a recurring accumulate-yet-undercorrect pattern: commitment-promoting components build up commitment in earlier layers, while abstention-promoting components act later as corrective signals that are often insufficient to overturn the accumulated commitment. Building on this finding, a lightweight policy trained on CAC activations improves decision accuracy by 12.2 points over the model's intrinsic commit-abstain margin, reduces false abstentions by 2.5 times, transfers to unseen benchmarks, and extends to larger models (27B-35B). The CAC is both diagnostic, clarifying how models overcommit, and practical, enabling improved abstention decisions.

Long-horizon LLM agents must convert accumulated experience into durable memory, deciding what to keep, compress, abstract into reusable skills and rules, or forget. We report a four-phase research program on this consolidation problem whose central finding is a shift in what is measured: from how much an agent remembers, to whether its consolidation decisions are any good, to whether those decisions can be trusted. Phase 1 learns episodic boundaries from agent traces by downstream utility; an honest near-miss (oracle correlation 0.691 vs a 0.70 bar) whose lasting output is a three-gate anti-leakage protocol. Phase 2 learns when to promote experience and to which abstraction level under a token budget, achieving a verified +22.7% task-success improvement with 7x compression, but exposing a degenerate-forgetting failure and a distribution-shift failure mode we name lambda-prevalence coupling. Phase 3 introduces ConsolidationBench, an oracle-by-construction benchmark that scores consolidation decisions against a known optimum on three non-circular axes; production retrieval systems retain information yet score zero on cross-level transfer. Phase 4 introduces governed consolidation: the decision wrapped in poison-resistance, reversibility, and auditability guarantees with a quality gate. Governance is statistically distinct from the quality score ($r^2 = 0.43$; partial $r = 0.27$; identical-quality policies differ threefold in governance), so the contribution survives independently of the metric's external validity. On that question we report a resolved negative: after a graded-reuse redesign removed a structural ceiling, a two-benchmark study with 2,532 real answer cells finds the quality score does not predict real transfer accuracy (pooled Spearman $\rho = -0.24$, n = 12, CI spanning zero). An adversarial self-critique pass cleared the final claim set with zero surviving overclaims.

Action agents do things for people. They send email, spend money, and call businesses while the user is busy with something else, so a mistake can turn into an action before anyone notices. They fail their users in two ways. They break trust when they do something the user never agreed to, or hold back after the user clearly said go. And they fall short on completion when they give up on errands that turn out to be hard. Both depend heavily on the harness around the model, meaning its instructions, tools, context, and guardrails. We built an evaluation that scores trust and completion on the same runs, in a simulated world of businesses with their own websites, inboxes, and phone lines, and of people who write back. A simulated user answers the assistant's questions. Trust means that nothing happens the user did not agree to. No email goes to someone they never approved, no private detail ends up on a group thread, no money is spent past their limit, no stranger's instructions are followed, and nothing is claimed without a source. Every trap has a matched control in which acting is the right call. We use the evaluation to measure Fo, Wajo's personal assistant, against a base model with basic instructions on three foundation models, and against the Fo harness with its guardrails switched off. Fo completes 71% of the errands and keeps the user's trust on 94% of the trap runs. The base models complete 50% to 64% and keep trust on 59% to 75%. On the matched controls, Fo goes ahead slightly less often. OpenClaw, a popular open-source assistant given the same access, completes 42% of the errands it shares with Fo, against 71%, and keeps the user's trust on 74% of the shared trap runs, against 94%. Measuring trust and completion together, on the whole system rather than the model alone, is how we think action agents become safe to hand real work to.

Large language models (LLMs) are increasingly used to drive scientific discovery. Understanding how LLMs make decisions from new data and memory of the literature is vital before trusting them to design experiments under tight experimental budgets. However, their decision strategies are invisible in the current performance scores used to evaluate research agents. Here, we benchmark five frontier LLMs in a Bayesian Optimization setting against published statistical baselines on seven combinatorial datasets spanning protein engineering, reaction optimization, molecular design, peptide self-assembly, and catalysis. Performance is paired with direct measurements of model beliefs and actions, enabling highly resolved behavior analysis. A prompt ablation that progressively strips context separates memorization from chemical reasoning and from bare categorical optimization. Prior chemical knowledge helps in expectation, but with high variance and occasionally even harms performance. No configuration tested decisively beats a mean statistical baseline across domains. Belief-movement and Martingale diagnostics, corrected here for a measurement-noise bias that mislabels rational agents as irrational, show that models overreact to incoming data rather than entrenching on their priors in the contexts studied here. Interestingly, while LLM actions are exploitative, models sincerely intend to explore and consistently act on that intent. This failure is a competence gap arising from context-stickiness. Removing in-context history restores exploration, indicating that priors and data must be decoupled to achieve effective LLM-driven discovery.

Test-time scaling and post-training have improved LLM performance in coding and mathematical reasoning, but their effectiveness for individual stance prediction remains unclear. We study this question by predicting a person's stance in a new discussion from their history. We evaluate widely used test-time scaling strategies and post-training methods, such as supervised fine-tuning and reinforcement learning, and identify four failure modes across generation, selection, and learning: (1) incorrect consensus, where repeated samples agree on the wrong stance; (2) selection failure, where generation covers the observed stance but selection misses it; (3) response overfitting, where supervised fine-tuning improves imitation but harms prediction; and (4) early plateau, where reinforcement learning shows modest initial gains followed by limited further improvement. We expose these failures using STANCE-BENCH, which contains 2499 prediction tasks from 500 Hacker News users. Guided by this analysis, we explore a simple approach that combines direct scores for all candidate stances with explicit assessments of support from the individual's history. On the 781-task test set, this approach achieves 21.83 discussion-specific Macro F1 with Qwen3-8B, compared with 19.27 for direct scoring. Our results motivate evaluating candidate generation, final selection, and person-specific evidence use separately.

Evaluating large language models (LLMs) relies heavily on benchmark scores, yet aggregate metrics can obscure whether benchmark samples reliably support model comparison. We introduce \textbf{BSDProbe}, a sample-level framework for \emph{benchmark structural diagnosis} that estimates capability boundaries from repeated-response trajectories along ordered model axes. BSDProbe summarizes samples by boundary position, boundary width, boundary-signal validity, and order consistency, then aggregates them into benchmark-level structural profiles. Experiments on six benchmarks show that benchmark reliability is axis-conditioned and heterogeneous: GSM8K and MATH exhibit the most stable measurement structures, MMLU and TriviaQA are relatively stable but heterogeneous, while GPQA and PopQA show stronger axis-conditioned risks. These profiles remain consistent across Qwen3, Qwen2.5, and cross-model axes. BSDProbe further selects compact high-value subsets whose model discriminability reaches up to $8.58\times$ that of the full benchmark. These results suggest that reliable benchmark use requires examining sample-level capability boundaries beyond leaderboard scores.

Claims paired with AI-generated images are a rapidly growing form of misinformation. Existing automated fact-checking (AFC) methods mainly treat this as a provenance problem, detecting low-level synthesis artifacts to decide whether an image is AI-generated. However, such methods do not verify what human fact-checkers often check: whether an image's content is consistent with the context implied by its accompanying claim. To address this gap, we introduce MIC (Multimodal Inconsistency Checking), an AFC framework that assists human fact-checkers by detecting AI-generated multimodal misinformation and explaining inconsistencies using world knowledge. MIC first uses supervised fine-tuning (SFT) for task adaptation and then applies Group Relative Policy Optimization (GRPO) to directly optimize component-level verifiable rewards for verdict prediction, inconsistency type classification, visual evidence description, and world-knowledge explanation. We further introduce MIC-Bench, a benchmark comprising 8,812 image-claim instances derived from 4,406 claims, where each claim is paired with an authentic image and an AI-generated counterpart that introduces a controlled contextual inconsistency. Compared with SFT alone, GRPO further improves Macro-F1 by 4.67 and 4.11 points in the in-distribution and out-of-distribution settings, respectively, while also improving the semantic similarity of visual evidence descriptions and world-knowledge explanations to reference annotations. Our code and data are available at https://github.com/UKPLab/arxiv2026-mic.

In autonomous deep research, a user provides a task and relevant background, then leaves the agent to conduct an extended investigation without further human intervention. We study how this initial user information is reflected in intermediate actions and how these actions relate to final recommendations. We introduce DRaligned, a counterfactual behavioral evaluation framework built on PDR-Bench. By varying one task-relevant user factor while keeping the remaining context fixed, we compare acquisition requests, working drafts, and final reports. Source-grounded extraction, blinded local judgments, and deterministic aggregation yield coarse directional measurements while leaving ambiguous cases unresolved. Our experiments show that strong user-specific delivery can emerge from a largely shared research process: agents investigate similar broad questions but allocate requests differently, and final recommendations distinguish user conditions more clearly than explicit requests do. Reports can also integrate user factors that were not jointly visible during acquisition. In readable draft-to-report comparisons, recommendations often retain their coarse user-specific direction despite substantial rewriting. Final directional differences recur across tested agent models, execution harnesses, and evaluator models, even as execution paths vary. These findings describe how initial user information shapes autonomous research and clarify the relationship between the process an agent follows and the recommendations it delivers.

Natural Language Understanding (NLU) plays a crucial role in various applications, yet its performance suffers from weaknesses in handling the complexities of human languages, ranging from lexical ambiguity to high-level reasoning difficulties. Analyzing errors across diverse linguistic phenomena is crucial for NLU improvement, as it will help humans get insights to comprehensively assess models' limitations and capabilities, so optimizing models' generalization. Notably, several benchmarks contain diagnostics datasets designed for investigation and fine-grained error analysis. When highlighting the gaps in the state-of-the-art, we noted that there is no naming convention for macro and micro categories or even a standard set of linguistic phenomena that should be covered. To overcome this gap, we propose an initial hierarchy for Cross-Lingual NLU error analysis. Moreover, we propose a methodology to create an NLI hierarchical framework and applied a case study on Arabic NLU. Moreover, this paper introduces E-CONAN diagnostics dataset, a freely available dataset manually-annotated with coarse-grained and fine-grained categories based on our proposed Arabic hierarchy. E-CONAN dataset helps NLU designers better understand their models by doing error analysis and in-depth investigation. We used E-CONAN to investigate the performance of 9 pretrained language models and 5 LLMs. Results indicate that LLMs outperform pretrained models in world knowledge and commonsense reasoning macro-category, and underperform pretrained models in syntactic macro-category. Moreover, the hardest phenomena for all models is Reasoning, and the easiest phenomena for all pretrained models is Syntactic, and the easiest for LLMs is Lexico-Syntactic.

A standard pipeline for symbolic reasoning over natural-language problems translates them into first-order logic and invokes a theorem prover. The translation step is the bottleneck: swap "every" for "some" and every inference that follows is corrupted. Yet today's metrics often score more broken translations higher than less broken ones, because BLEU, BERTScore, and Smatch++ reward surface overlap that the worst errors happen to preserve. We introduce SIV, which derives two kinds of probes from the target formula and uses a theorem prover to verify the candidate translation against each. Positive probes are statements the candidate must entail, which detect translations that drop content; contrastive probes are statements the candidate must not entail, which detect translations that assert more than the original. On a controlled pool of perturbed FOLIO translations, the severity of the error accounts for 80% of SIV's score variance, compared with at most 17% for any prior metric. Across six error classes on a disjoint pool, SIV scores the reference above the perturbed candidate in over 99% of pairs. Because each probe is labeled with what it tests, the failure pattern also supplies a labeled error trace, recovering the perturbation class at macro-F1 0.638, nearly double the score-only baseline. On 434 expert-audited real LLM translations, SIV attains the top AUC, uniquely detects and grades expert-labeled major errors, and abstains, rather than mis-scoring, on out-of-vocabulary translations.

Language model agents are increasingly deployed in open-world tool environments, which require balancing exploring unknown capabilities and exploiting known ones. Existing methods face a performance-efficiency tradeoff: they either rigidly decouple exploration and execution or interleave them without coordination. We argue that the key lies not in whether to decouple or interleave them, but in how to coordinate them across granularities. We introduce ParaAct, a structured parallel-action loop that combines phase-level Exploration $\rightleftharpoons$ Execution with action-level parallelism. To learn this loop, ParaAgent combines multi-agent cold-start demonstrations with reinforcement learning under multi-level advantage decoupling, making planning structure explicit and supervising it with step-, phase-, and trajectory-level rewards. Learning is supported by our ToolEnv, a scalable simulator grounded in 50,011 realistic tool interfaces. On two open-world tool benchmarks, ParaAgent-4B achieves the best average success among all baselines, including GPT-4.1 systems, with the largest gains on multi-tool tasks. Behavioral analyses show that these gains stem from this action organization, highlighting its importance for capable and efficient open-world agents.

We introduce RareTrap, a framework for estimating the probability of severe behaviors in black box large language models (LLMs). A key challenge for probability estimation is defining a tractable distribution over the input space. To accomplish that, RareTrap uses a surrogate LLM and constructs a geometry-aware mapping from a lower-dimensional latent reference space into its token-embedding space to induce an explicit and reproducible distribution over input prompts. A response-level performance function is utilized on the response to quantify behavior severity. This enables sequential rare event simulation that concentrates evaluations on progressively more severe behaviors while preserving probability under the induced prompt distribution, which would otherwise be prohibitive to measure. Across 10 open-weight and two frontier models (GPT-5.4 and Claude Sonnet 4.6), we find that RareTrap successfully induces severe resource consumption behaviors and computes their probability with as few as 200 evaluations. RareTrap provides model developers a principled approach for evaluating language models under a common distribution, and prioritizing alignment effort to improve safety and mitigate risks. Code is published online: https://anonymous.4open.science/r/rare_trap-D7F0.

LLM agents increasingly take consequential actions through interactions with users, policies, and external tools. Auditing these agents requires automated attribution of realized actions to their historical basis. However, existing attribution formulations do not provide question-specific traces for diverse auditing objectives. Additionally, when access to the acting model is limited (e.g., in API-only deployments), applicable methods commonly rely on costly input perturbations or external LLM analysis of complete trajectories. We therefore formulate $\textit{query-conditioned agent action attribution}, a new task that takes a natural-language auditing query as input and recovers the source and ordered intermediate evidence for the query-specified aspect of an action. We instantiate this task with $A^3Bench$, a benchmark comprising 1,396 auditing queries across policy basis, parameter provenance, failure propagation, and unsafe-behavior tracing. To enable efficient, query-specific attribution, we use small open-weight models as attribution proposers that combine query-conditioned gradient saliency with query-semantic relevance to rank history units. Our proposer consistently achieves stronger source and evidence rankings at lower inference cost than open-weight baselines, improving source MRR by up to 40.9\% and evidence MAP by 42.1\% with only two forward passes and one backward pass. Controlled evaluations confirm that our proposer improves attribution specificity by adapting its rankings to fine-grained changes in the auditing query. Building on a proposer ensemble, our end-to-end system surpasses the strongest frontier-model baseline in source accuracy (64.5\% vs.\ 60.4\%) while reducing empirical deployment latency by 29.9\% relative to the fastest frontier API baseline. Code and data will be released after the initial review period following final validation and cleanup.

Large language models are increasingly used as interactive writing tools, where users develop stories, revise ideas, and introduce new requirements across multiple turns rather than specifying a complete brief upfront. Yet most evidence on multi-turn instruction degradation comes from tasks with objectively verifiable outcomes, leaving unclear whether incremental interaction harms creative artifacts in ways that explicit requirement checks cannot capture. We study this question using 160 human-authored creative-writing tasks across six genres, presenting each intended specification either upfront or progressively over 5-9 turns to six distinct open-weight model families, yielding 960 matched pairs. Progressive delivery reduces explicit constraint adherence and produces its largest writing-quality degradation in structure/coherence. The structural gap persists among outputs with equal observed adherence, suggesting that measured requirement loss alone does not explain the observed structural difference. We define Creative Integrity as a compact measure of joint adherence and narrative structure; under incremental delivery, models retain 71.2% of FULL Creative Integrity (95% CI [68.2%, 74.3%]). A three-rater human study over 50 matched pairs independently recovers FULL advantages in structure/coherence, craft, and genre effectiveness, while automated scores remain positively associated with aggregated human ratings. These findings show that interactive creative-writing systems should be evaluated not only on whether requirements survive conversation, but also on whether evolving requirements remain coherently integrated into the final artifact. Our dataset, benchmarks, and source code are available at: https://github.com/solusops/SISTER-2026-Team19

We introduce NSV-Shift, a contrastive benchmark for evaluating whether speech-to-speech models can understand non-speech vocalizations (NSVs) and adapt their responses accordingly. Each pair contains two conversations with identical lexical content that differ only in the NSV embedded in the final turn. Our pilot contains 22 human-verified pairs (44 audio conditions) and evaluates five models on NSV perception, emotion understanding, and response adaptation. Results show that models generally perform better at detecting NSVs than at interpreting their fine-grained emotional meaning or producing appropriately differentiated responses. The data construction pipeline, dataset, and evaluation pipeline are publicly available at https://github.com/ChenzwNina/nsv-construction.

Names are personal identifiers, but they also carry social meaning and are widely used to evaluate how language models treat different people. Such evaluations typically assume that matched names are comparable model inputs. We show that this assumption often fails at the lexical interface: matched names are not necessarily matched inputs. Some names receive direct single-token access, while others are assembled from multiple subwords, creating unequal name-surface support. Across nearly half a million first names and 12 LLM-associated tokenizers, direct lexical access is highly selective, model dependent, and uneven across race- and gender-associated name metadata. We introduce NameTrace, a model-native, fine-grained, pre-behavioral framework for measuring whether unequal name-surface support remains a vocabulary property or becomes visible in task-relevant internal representations. NameTrace measures concept accessibility from the model's own probabilities over task-specific adjective axes with continuous task-aligned weights. On matched atomic and short-fragmented names within the same race/ethnicity--gender-associated strata, support predicts systematic differences in concept accessibility across fellowship, hiring, clinical assessment, and lending. These differences persist across all eight matched strata, extend across model families, and transfer to unseen names. Hidden-state interventions further show that the measured task directions have downstream leverage, shifting later constrained choices. Unequal lexical support is therefore demographically structured at the input and remains visible in task-relevant model computation. NameTrace makes lexical comparability measurable, supporting a broader principle: behavioral comparability begins with lexical comparability.

Self-evolving agents can continually improve their behavior, while tools define the executable action space through which they interact with the environment. However, exposing the full tool library to model introduces substantial irrelevant context and can impair tool-use decisions. We study tool-space self-evolution, where each recurring task type maintains a persistent tool space which is constructed from accumulated output experience. We identify three limitations of existing methods: (1) output-unaware selection: they rely primarily on tool descriptions or model priors rather than observed tool outputs; (2) statelessness across request: they select tools independently for each request without consolidating prior output experience into persistent task-specific state; (3) inference cost: they repeatedly search, rank, or reason over candidate tools for subsequent requests of the same task. We address these limitations through output-aware tool scoring, persistent task-specific tool spaces, amortized tool selection, and reusable configurations across models. We introduce LOTS (Likelihood-Only Tool Scoring), which evolves an agent's tool space from accumulated output experience while keeping model parameters fixed. After each request, LOTS holds the model's generated answer and estimates each tool's contribution by measuring how much the answer likelihood changes when its observed output is removed. These contributions are aggregated within each recurring task to rank tools and update its persistent space. Across three benchmarks, LOTS improves task performance while substantially reducing tool context. More importantly, sequential experiments demonstrate that task-specific spaces persist and continue to improve over time, while cross-model experiments show that learned configurations transfer across different models.

We introduce GlyphBench, an environment suite for reinforcement learning (RL) post-training of language-model agents, with over 360 tasks spanning diverse games. GlyphBench renders spatial observations as two-dimensional Unicode grids and connects training, evaluation, and trajectory replay through a unified interface designed to support efficient and reproducible research. We use GlyphBench to study how observation interfaces, reasoning effort, and agent harnesses affect performance, and how RL configurations shape learning dynamics. Our results show that glyph observations outperform native text and pixels in our Craftax experiments, with further gains on several BALROG environments. RL on 100 GlyphBench tasks improves Qwen3.5-4B on held-out Reasoning Gym problems, reaching 63.48% accuracy and outperforming the base model, a math-trained baseline, and a code-trained baseline. These experiments provide empirical evidence that reasoning gains from gameplay can yield stronger transfer than math or code. Together, these results highlight GlyphBench's value as a testbed for systematic research on how language-model agents learn, interact, and generalize.

Large language models receive limited explicit training in reasoning about runtime program states. We study whether training models to reason about runtime program states improves downstream software-engineering capabilities. We introduce two complementary program-state reasoning tasks. Buggy input-output reasoning requires a model to generate a concrete input that exposes a behavioral difference between a buggy program and a hidden correct implementation and to predict the resulting execution behavior. Precondition-postcondition reasoning requires an agent to symbolically characterize a bug-triggering precondition, predict the expected postcondition, explain their causal connection, and instantiate this reasoning as an executable regression test. By incorporating these two tasks into a staged post-training pipeline, we develop Comet-9B, a 9B language model based on Qwen3.5-9B Base. We evaluate the resulting checkpoints on repository-level patch generation, regression-test generation, and security PoC generation. Adding both program-state reasoning tasks to supervised fine-tuning (SFT) on issue resolution improves success rates by 7.25 percentage points on SWE-bench Pro and 9.70 points on SWT-Bench Verified. Sequential reinforcement learning on the two tasks yields further gains of 7.25, 26.79, and 4.67 percentage points on SWE-bench Pro, SWT-Bench Verified, and CyberGym, respectively. Despite having only 9B parameters, Comet-9B achieves a score comparable to the reported GPT-5.2 result on SWE-bench Pro and matches the reported success rate of a GPT-4o-based agent on SWT-Bench Verified.

The transition from Large Language Models (LLMs) to agents shifts safety stakes from toxic text to irreversible environmental harm. While current defenses remain largely retrospective, proactive runtime intervention is bottlenecked by the lack of large-scale, causally-consistent data. We propose PROACT-Agent, a framework for synthesizing high-fidelity trajectories to enable real-time guardrails. We identify a critical "safety drift" in prior benchmarks, where lenient annotation paradigms fail to enforce temporal consistency. PROACT-Agent addresses this through: (1) Progressive Trajectory Unrolling to reveal risks hidden in long-context interactions; (2) Reasoning-Augmented Causal Rectification to enforce monotonic causal consistency; and (3) Culturally-Aware Data Localization for cross-border robustness. We introduce PROACT-Bench, a bilingual safety benchmark with 155,780 states labeled through multi-model adjudication. Evaluating updated context before the next LLM inference, the trained guard achieves 91.46% unsafe-class F1 and 90.63% exact-boundary detection under complete source holdout. In AgentDojo, it reduces non-DoS targeted attack success from 20.82% to 0.40%.

Tool agents use large language models to act through external tools, yet successfully executed calls can still leave user requests unfulfilled. Tool-agent repair seeks alternative call sequences that execute successfully and fulfill the original requests. However, repair requires exploring both operation choices and their concrete realizations, making complete-sequence regeneration costly. Moreover, regeneration repeats operation selection even when failure arises from how those operations are realized. The resulting challenge is to reduce this repetition while preserving exploration of alternative operations and realizations. Therefore, we formulate repair as hierarchical search over operation supports, which we introduce as sets of permitted operation types that define reusable search regions for concrete tool-call sequences. We propose ReCommit, a training-free, diffusion-guided framework for improving tool-agent failure recovery while reducing repair computation. ReCommit amortizes operation-level proposal computation across repair trials by reusing operation-type scores from a single parallel readout of a masked diffusion language model. These scores guide search across supports, while realization search explores alternative entity bindings, arguments, and action composition within each support. Experiments on real failures across four enterprise services in the Agent-Diff benchmark show 75.9\% and 63.2\% relative recovery gains with 61.3\% and 51.3\% reductions in mean full-budget repair time at repair budgets $B=3$ and $B=13$, respectively, over the strongest evaluated 8B comparison method. ReCommit achieves a favorable recovery--cost trade-off, including in comparisons with the evaluated 32B models.

In 2023, AI models answered from training data and hallucinated when it ran out, and businesses were told to seed that knowledge. Models' training knowledge has since given way to live web search, and the advice followed it there: answer-engine optimization, or AEO, now tells businesses to scatter breadcrumbs across forum threads, listicles, and off-site citations, so AI engines are likelier to surface and recommend them. But being surfaced is no longer enough: an agent opens the results and reads them before deciding, and one buyer question sends it through several rounds of search and fetch. What decides the outcome at this drill-down step is whether the agent can fetch and read the business's own site: agent experience (AX). We argue that AX is the new AEO. We run 37,927 agent journeys, each a buyer question about a business, across four independent harnesses over 1,056 real businesses, matched on fame, prior model knowledge, and two AEO proxies, then split based on their AX level. Only 7-10% of the finished answer comes from the model's training knowledge, whether or not the site is readable. Agent-ready businesses have answers built from their own pages 78% of the time against 56% and are clearly recommended 1.9x more often, while every grounded answer about a not-agent-ready business costs the agent 64% more. Holding business, harness, and question fixed, answers built from the site are 41% more accurate. The dominant failure is not fabrication but omission: web-built answers are 3.7x more likely to contain none of the facts the buyer asked for. Baselines differ sharply across the four harnesses, with clear-recommendation rates varying sevenfold from stack to stack, yet the effect holds in every one. In the agentic web era, being readable beats being talked about, and improving a site's AX is the strongest lever a business has.

Large Language Models (LLMs) have shown strong performance on tool-use agentic tasks when given a fixed tool schema. Yet a tool schema is not the action space of an agent; it is merely one interface representation of it. The same executable action can be exposed through many different, functionally equivalent tool definitions, and an agent that has truly learned a task should behave consistently across them. We show that current agents often do not, a phenomenon we term schema bias. To study this systematically, we introduce an executable transformation framework that rewrites a native tool schema using nine operators, including merging and splitting tools, altering how a single tool is expressed, and distributing one action across several dependent calls. The tasks, executable actions, and reachable states remain fixed, so any change in success is attributable to the interface alone. Evaluating eleven LLMs, including two closed models, on up to 32 schema variants, we ask how large schema bias is, how it manifests, whether the difficulty of a schema variant can be predicted without a full evaluation, and whether training removes it. We find that schema bias is substantial even for the newest models: success rates range from complete failure to 97% depending solely on the schema. To reliably estimate schema difficulty, it requires running a small sample of the target queries. Training repairs a schema variant only when that variant appears in the training data.

LLM-based Web agents can autonomously complete user tasks, yet deceptive interfaces can steer them toward outcomes that conflict with users' interests. Existing defenses primarily intervene on agent behavior through blocking, guidance, or replanning. We identify a distinct failure mode: a task-valid action can still realize an unauthorized consequence because of the current Web state. This motivates treating task-relevant Web state itself as a runtime control target. We introduce Veer, an agent-side runtime defense that leaves task planning to the base agent and intervenes on Web state when a proposed action would produce an unauthorized consequence. Before modifying the live environment, Veer constructs a prospective intervention trajectory toward a safe task-relevant state and executes it with runtime grounding and verification. Across TrickyArena and WebDecept, Veer achieves the highest safe task completion in all three evaluation settings, exceeding the next-best defense by 15.9 and 25.0 percentage points on TrickyArena-Single and TrickyArena-Multi, respectively, while reducing dark-pattern success on WebDecept to 0.3%. These gains persist across dark-pattern types and all 12 agent, model, and benchmark configurations. Ablations show that active state intervention provides the largest gain, while prospective rollout and temporal evidence contribute additional improvements. These results establish task-relevant Web state as an effective runtime control target for protecting Web agents from deceptive outcomes.

We present WebPageBench, an open framework for evaluating web agents in which every task is verified from the interface's own event log. Six instrumented mock sites with brand identifiers removed (a marketplace, a bookstore, a grocery service, rail ticketing, hotel search and a document cabinet) emit typed events with parameters as a user or an agent acts. A task declares the events it requires, and success is decided by matching them, with no judge model and no scraping of rendered pages. The same instrumentation supports controlled UI variation: one configuration switch re-renders a task through a different implementation of a single control while the prompt and the success conditions stay completely identical, so sensitivity to interface form can be measured under a fixed task specification. The WebPageBench release consists of three components: 152 tasks, divided into 65 canonical scenarios and 87 control variants across light/dark UI-modes; a common runner evaluated with six browser/DOM harness configurations and five screenshot-only GUI-agent families; and a public leaderboard of 24 model-harness pairs. On the public 152-task leaderboard the gap between what agents declare finished and what the log confirms reaches 41 points (one configuration declares every task finished and satisfies the conditions on 59%).

Large language model (LLM) agents face critical privacy risks when acting as delegates in human-agent-human communication. To prevent such breaches, agents must understand users' social relationships and adhere to context-dependent social information disclosure boundaries. Current studies on agent memory privacy focus on instantaneous interactions, leaving the long-term relational disclosure problem unexplored. In this paper, we propose EP-Mem, an Elastic Privacy Memory architecture that reframes privacy as user-owned boundary control across social roles. EP-Mem introduces (1) token-level memory driven by user-configurable a privacy policy that stratifies persons and events, combining domain-level default circulation rules with fact-level whitelist/blacklist exceptions; and (2) a pluggable sidecar with a privacy engine that aligns disclosure controls with memory across summary, detail, and boundary granularities, enforced throughout generation, storage, and retrieval. We construct EP-Bench, to our knowledge the first long-term multi-party benchmark with cross-session correlated events for policy-conditioned relational disclosure. Experiments show that EP-Mem achieves 94.0% privacy classification accuracy, improves disclosure-permission judgment from 22% to 68%, and reduces privacy leakage by 75.6%, while maintaining retrieval performance and cross-benchmark generalization.

Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On 6,925 MMLU-Pro debates, the protocol identifies 253 collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents 29 collapses but loses 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate 8-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high (G=7, Spearman rho=0.893, exact two-sided p=0.0123), but initial-majority accuracy is a close comparator (rho=0.821; family partial rho=0.767, p=0.0877), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model-scaffold rows can be compared under the same denominators and signed utility ledger.

Machine learning task type identification is essential for constructing valid ML pipelines, yet in practice it is typically specified manually. We investigate whether large language models (LLMs) can infer both the data domain and the downstream prediction task directly from dataset-level information when only the target feature is provided by the user. Together with our LLM-based system we also release an annotated benchmark comprising 625 public tabular and time series datasets. We evaluate the proposed approach in three settings: (i) tabular datasets in comparison with established AutoML heuristics, (ii) cross-domain evaluation across tabular and time series datasets, and (iii) a practical deployment scenario using smaller local models. The results show consistent advantages for LLM-based task type identification, with increasing difficulty in heterogeneous and resource-constrained settings. LLM-based approaches outperform AutoGluon in the tabular setting, reaching 0.98 F1 macro compared to 0.93. In the cross-domain setting, the best model achieves 0.90 F1 macro, while smaller locally deployable models reach 0.75, indicating a trade-off between deployment feasibility and accuracy.

Scientific discovery requires not only recovering mathematical laws that describe observable behavior, but also identifying the mechanisms that generate them. Existing benchmarks for symbolic regression and scientific agents primarily evaluate phenomenal-law recovery, leaving mechanism discovery largely untested. We introduce MechBench, a benchmark that explicitly separates these two capabilities. Each task is defined by a mechanistic model, a structured set of scientifically meaningful relations whose joint consequences entail an observable phenomenal law, while agents receive only observational data and scientific context. We evaluate mechanism recovery through mechanism probes, which query internal scientific consequences that cannot be inferred from the phenomenal law alone. To reduce reliance on memorized textbook mechanisms, we construct unfamiliar variants through controlled, scientifically interpretable mutations of canonical mechanisms, and screen for mechanistic indistinguishability to exclude ambiguous instances admitting comparable competing mechanisms. Experiments across representative scientific agents reveal a substantial phenomenal--mechanism recovery gap: for Codex with GPT-5.6-sol, phenomenal-law accuracy reaches 35.00% on the Core-set while mechanism accuracy is only 13.75%, with mechanism recovery failing in 64.29% of cases where the phenomenal law is correctly recovered. The gap widens as mechanisms become increasingly mutated, and even providing the correct phenomenal law leaves mechanism recovery below 50%. These results reveal a substantial generalization gap in mechanistic reasoning and establish mechanism discovery as a distinct challenge beyond recovering observable scientific laws.

Arabic speech technology has largely focused on Modern Standard Arabic, leaving the living dialects spoken by hundreds of millions under-served. We introduce ALMIEYAR, a culturally grounded ASR benchmark covering 17 Arabic dialects across six families, built entirely from newly recorded speech unseen by existing models. Dialect-community coordinators selected culturally relevant images across 10 topics, and native speakers described them through five structured scenarios, yielding approximately 50 minutes per dialect (13.7 hours total). We benchmark 12 state-of-the-art ASR systems zero-shot, including GPT-4o-transcribe, Voxtral-Mini-4B, Fanar-STT-LF, Whisper, SeamlessM4T-v2, and wav2vec2-based models. GPT-4o-transcribe achieves the lowest overall WER at 35.0%, followed by Voxtral-Mini-4B, Fanar-STT-LF, and Whisper-Large-v3 at 41.1%, 45.9%, and 49.5%, respectively, indicating substantial remaining errors across Arabic dialect communities. Performance varies considerably across dialect groups, with no model performing uniformly best across all groups. WER alone also obscures dialectal ASR behaviour: wav2vec2-based models show large WER/CER gaps, where character-level agreement remains much higher than word-level accuracy, motivating joint WER/CER reporting. ALMIEYAR provides a unified benchmark for culturally grounded Arabic ASR evaluation, including the first published benchmark for Ahwazi Arabic.

Recent progress in machine learning is driven by large-scale foundation models, where scaling laws and finding optimal scaling prescriptions for architecture, data, and hyperparameters are key in advancing the state-of-the-art. Therefore, it is surprising that no systematic study evaluates the methodology to obtain scaling laws and prescriptions across different model types. To shed light on this crucial blind spot and facilitate future research, we introduce the surrogate benchmarks ScAn-Bench-LLM and ScAn-Bench-VLM based on 4524 and 8024 checkpoints of language and vision-language model pipelines. On our benchmarks, we perform the first systematic evaluation of both data acquisition and extrapolation methodology for scaling analysis across different data modalities.

Evaluating finance research agents requires rubrics that reflect expert standards and fix the values correct as of an information cutoff. Expert-reviewed finance benchmarks rely on fixed, per-item rubrics, which are costly to extend and cannot encode each institution's own standard. In FinAutoRubric, experts specify reusable evaluation guidance, while agents and code carry out query-specific rubric generation, review, and validation. This expert guidance governs every agent, as prompts and as rules that code enforces, and a Task Bank of reusable criteria carries it across tasks. In long-horizon loops that follow the expert guidance, a writer agent researches every expected value and a reviewer agent verifies it, and failures escalate to a human. On three expert-authored finance benchmarks, its rubrics track expert scoring as closely as the strongest evaluated generator while stating the expert rubric's expected value for more criteria, their scores agree with human grading, and in-house analysts prefer them in a blind review. The released 100-query FinAutoRubric Benchmark, built from in-house analysts' key questions across 78 tasks and eight asset classes, shows that rubrics from an earlier model generation still leave headroom for a later one.

Clinical trajectory models are usually evaluated by next-event accuracy on observed histories. Simulation is different: models must condition on their own generated events, allowing errors to compound. Although this problem is well known in sequence modelling, it has not been systematically quantified for clinical trajectory simulators. We developed EDSim-Bench to evaluate this failure mode using 425,028 MIMIC-IV-ED stays, with external replication on MC-MED, and release the evaluation protocol and scoring code. Starting from held-out visit prefixes, models generate the remainder of each visit and are evaluated on termination, event composition, timing, conditional fidelity, and occupancy forecasting, with a train-only order-3 n-gram as a reference baseline. Despite next-event accuracies within 0.001, three neural architectures behaved very differently under rollout. Across seeds, one Transformer recipe ranged from 0.43 to 0.96 in termination score and from 4.2- to 137-fold the divergence of the n-gram; no prefix-trained neural model approached the n-gram on termination or event composition. Inference-time interventions improved termination but did not jointly recover composition and timing. Supervising every eligible sequence position rather than only the final prefix position was associated with one to two orders of magnitude lower divergence across Transformer, GRU, and LSTM models, with the pattern persisting under model scaling, temporal shift, and external-site evaluation. Nevertheless, even the best model generated visits approximately half as long as observed, and model rankings reversed on occupancy forecasting, a downstream quantity relevant to bed management. These results show that next-event accuracy is insufficient to evaluate clinical trajectory simulators and motivate closed-loop evaluation across seeds, rollout criteria, and downstream tasks.

Retrieval-augmented generation (RAG) pipelines are typically assembled from independently-chosen components -- a document parser, a chunking strategy, and an embedding model -- yet these choices are rarely evaluated jointly, and evaluations that do combine them are usually run on a single document or corpus. We present a controlled factorial study of 3 parsers, 3 chunking strategies, and 5 dense embedding models, together with a sparse BM25 baseline, evaluated against 800 question instances, each with one or more required evidence strings, with evidence strings automatically validated against source text and a 10% random sample manually reviewed, across four structurally distinct Indian central-government regulatory documents. We fit linear mixed-effects models with document-query-level random intercepts to the resulting 72,000-row result set, run Holm-corrected paired comparisons between matched dense and sparse configurations, and report clustered bootstrap confidence intervals for all 54 unique retriever configurations. We find that no single retriever family dominates across documents; parser and chunker choice interact significantly; MPNet-base is a consistent underperformer with a severe failure mode on table-derived questions; and the corpus exhibits a near-saturated evidence-preservation ceiling above 98%, indicating that retrieval differences are driven primarily by ranking quality rather than information loss during ingestion. We additionally report embedding-dimension and chunk-size/overlap ablations and an efficiency/quality Pareto analysis. We release our full evaluation harness, corpus manifest, and 800-question benchmark.

While Vision-Language-Action (VLA) models perform strongly on manipulation tasks, their responses to invalid task premises remain underexplored. Existing evaluations of premise conflicts often focus on terminal task outcomes, yet task failure alone cannot distinguish behavioral disengagement from continued pursuit followed by an execution error. We call the latter pattern Failed Persistence. To study this phenomenon, we introduce ConflictVLA-Bench, which pairs conflict rollouts with premise-consistent reference rollouts and evaluates both outcomes and execution processes. Built on LIBERO, the benchmark contains 2,826 prompt-conditioned conflict tasks spanning four conflict families, four structural configurations, and two prompt conditions. Across all eight VLAs, invalid premises reduce original goal completion by at least 17.3 percentage points, with the reduction reaching 56.2 percentage points for OpenVLA. Crucially, even when models succeed on premise-consistent tasks and fail on their matched conflict tasks, they often continue to approach the original targets, retain early trajectory structure, and show limited action magnitude suppression. Failed Persistence therefore recurs across the evaluated models. Explicit premise checking does not consistently produce selective and coordinated behavioral changes. These findings show that terminal failure alone establishes neither behavioral disengagement nor refusal and that outcomes alone are insufficient for VLA evaluation. Experimental data and additional details are available on the project page: https://github.com/EmbodiedAISurvey/ConflictVLA-Bench

We introduce ROTE (RollOut Testing of Exact memorization), a benchmarking protocol for evaluating symbolic memorization of neural architectures. We study memorization and the extension of symbolic rules in neural sequence models by using sequences whose complexity is regulated by Lempel--Ziv--Welch (LZW) compression. Under ROTE, each architecture is trained as the same finite-context conditional predictor and is evaluated using teacher-forced one-step prediction as well as closed-loop rollout on the withheld symbols. Following a shared prediction-and-rollout evaluation routine, the benchmark evaluates gated recurrent, minimal recurrent, attention-based, and hybrid recurrent-attention models with their native computational characteristics preserved. Beyond standard predictive metrics, the benchmark reports normalized string distances, training time, memory usage, and parameter count across an LZW-complexity sweep. The study establishes a connection between the complexity of algorithmic sequences and the memorization capacity of neural architectures, revealing the trade-offs involving memorization quality, rollout stability, and computational expense. Our software and reproducible experimental code can be obtained from https://github.com/nla-group/rote.

Hybrid transformer architectures that replace most softmax attention layers with linear attention offer transformer-level quality at a fraction of the memory cost. Rather than pretraining such models, a growing body of work distills them from already trained full-attention transformers. However, these distilled models often collapse on long-context retrieval and reasoning tasks, particularly when operating in thinking mode, where the efficiency gains of hybrid architectures matter most. Since linear attention layers must compress context into a fixed-size state, their errors compound over long sequences. As off-policy distillation never teaches the student model to recover from this drift, tasks that necessitate longer sequence lengths become especially challenging. We introduce On-Policy Attention Linearization (OPAL) in which the hybrid attention student samples its own long-context trajectories and receives dense supervision from the frozen full-attention teacher. Applying OPAL to Qwen3-4B and MiMo-7B-RL-0530, we recover $87$--$94\%$ of full-attention performance on commonsense reasoning, $100\%$ on needle-in-a-haystack (NIAH) retrieval, and $83$--$93\%$ on mathematical reasoning with only 3B training tokens. We achieve these results without supervised fine-tuning (SFT) or reinforcement learning with verifiable rewards (RLVR). Compared with the strongest prior linearization method, which recovers $68\%$ of its teacher's retrieval performance and $21.6\%$ absolute average mathematical reasoning accuracy, OPAL fully recovers retrieval and achieves $67.6$--$72.2\%$ on math reasoning.

Tool-using LLM agents must decide not only whether additional information is needed, but also which source can resolve the uncertainty. Existing proactive approaches often specialize in either user clarification or environment verification, without explicitly determining the appropriate information source for each decision. We formulate this problem as uncertainty routing among ACT, CLARIFY, and VERIFY, and propose PROUR, a proactive uncertainty routing framework. PROUR decomposes action uncertainty into disagreement across plausible user-goal interpretations, which signals user-side ambiguity, and the entropy remaining within each interpretation, which signals missing world-side evidence. To acquire information from the routed source, a query generator is trained with a mode-conditioned information-gain reward, targeting user-goal identification under CLARIFY and next-action identification under VERIFY. On $\tau$-bench, PROUR achieves 28.17% average success rate across retail and airline, outperforming the strongest prior method by 4.57% while using 2.17 fewer interaction steps. The learned policy further generalizes to stronger task agents and transactional domains of $\tau^3$-bench without retraining, demonstrating the benefit of source-aligned uncertainty resolution for proactive agents.

Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender. Reusing the sender's key-value (KV) cache avoids this redundancy, but prefill-free transfer across model families must handle differences in tokenization, model depth, and KV representations. To address these issues, we propose \textit{HeteroFold}, a prefill-free cross-family KV cache transfer method that keeps both the sender and receiver frozen. HeteroFold aligns model structures, maps the sender cache into the receiver space, and calibrates it to preserve receiver behavior. Across six transfer directions, HeteroFold achieves the best cache-transfer performance on all four long-context benchmarks and most short-context settings. It also matches text-based communication on the multi-agent benchmark. At 32K context length, Llama-3.1-8B$\rightarrow$Ministral-3-14B transfer is $10.7\times$ faster than Native Prefill and $1.18$--$1.47\times$ faster than the state-of-the-art prefill-free baselines, Dense Latent and KV Ridge. These results show that HeteroFold enables efficient cross-family KV reuse without receiver prefill.

Comparative Process Reward Models (PRMs) provide critical step-level guidance for autonomous web agents by evaluating state-conditioned preferences between candidate actions. However, existing preference training data synthesized via multi-policy sampling suffers from a severe scarcity of Grounded Minimal Contrastive Pairs (GMCPs)-where competing candidates target genuine on-page elements with identical action types. In representative baselines preference data (namely, WebArbiter), GMCPs account for merely 24.19%, biasing PRMs during training to rely on shallow shortcuts (such as element hallucinations and action type mismatches) rather than acquiring genuine contextual decision semantics. To address these challenges, we propose SURFPRM, a graph-guided process preference synthesis framework for comparative Web PRMs. SURFPRM structures web demonstrations into a persistent Interaction Element Graph that acts as an environment-grounded negative action proposal mechanism, systematically synthesizing contrastive negative actions across spatial, temporal, and spatiotemporal confusion axes. This elevates the GMCP proportion from 24.19% to 74.60%, producing the curated SURFPRM-DATA dataset. Across six open-source backbones (3B to 9B parameters), PRMs trained on SURFPRM-DATA outperform baseline-trained models on average on WEBPRMBENCH and rival leading proprietary LLMs. In downstream reward-guided trajectory search on WEBARENA-LITE, SURFPRM provides step-level guidance for both GPT-4o (+14.21%) and GPT-4o-mini (+12.83%) policies, yielding substantial improvements in complex web task success rates.

Tool-using large language model (LLM) agents increasingly perform state-changing actions that require user authorization. Yet existing approaches do not provide a principled mechanism for selectively updating prior authorization when only part of an instruction changes. To this end, we propose an Authorization-Closure-Graph (ACG)-based framework that represents authorization and its dependencies as an evolving, versioned state. ACG selectively invalidates authority affected by a revision while preserving unaffected portions of the authorization state, and computes a minimal repair that identifies only the missing evidence or authority required for execution. This enables agents to adapt to revised instructions while avoiding stale authority and unnecessary authorization requests. We evaluate ACG across three advanced LLMs in two natural tasks, and ACG consistently improves action safety rate and task success rate. Code is available at https://github.com/weiliang822/ACG.

Autoregressive Next-Token Prediction (NTP) has enabled strong reasoning capabilities in language models, while Diffusion Language Models (DLMs) offer flexible token orders and parallel generation. We ask whether DLMs can acquire NTP-style reasoning through distillation without giving up their native generation process. Direct distillation, however, faces a fundamental mismatch: an autoregressive teacher predicts from a left prefix, whereas a DLM can condition on tokens on both sides. We introduce ForkLeft, a distillation framework that resolves this mismatch by separating the student's rollout from teacher supervision. During training, the student first performs entropy-first rollouts that commit uncertain positions and expose potential forks. We then fix the resulting student prefix and distill an NTP teacher under the same context, with answer correctness determining the supervision source. At inference, the student returns to its native confidence-first parallel decoding. With Qwen3-30B-A3B-Base, ForkLeft improves Efficient-DLM-4B on all ten benchmarks, raising MATH500 from 72.60% to 79.60% and consistently outperforming three alternative designs. The gains scale with teacher strength and generalize to SDAR-4B with only $500$ updates. At matched scale, the distilled 4B and 8B students exceed the published SDAR-Chat and OPDLM models on seven benchmarks, showing that DLMs can learn NTP-style reasoning without sacrificing native parallel generation. Code and datasets will be released upon acceptance.

Software development requires more than editing code: developers repeatedly run software, interact with its interfaces, visually inspect its behavior, and use these observations to decide what to change next and whether a change works. Existing coding agents and computer-use agents are largely studied in isolation, leaving this integrated development process underexplored. Diagnosing a runtime interaction failure requires agents to connect visual observations with the responsible code, then use the application again to verify the repair. We introduce CUA-SWE, a benchmark, environment, and evaluation pipeline for software engineering with computer use. Beyond studying how GUI feedback supports diagnosis and repair, we ask whether agents can complete software engineering tasks when required specification or operational information is available only through the running application's visual interface. CUA-SWE spans four software engineering domains and requires agents to modify code and configuration, execute commands, interact with running software, and inspect visual feedback within the same task. Each task includes deterministic, task-specific tests that verify whether the resulting software satisfies the requirements and preserves specified behavior. Our evaluation characterizes how frontier agents combine source-level execution with application screenshots and graphical interaction to produce verified software changes. We examine performance across domains and task information requirements, alongside the development behaviors associated with successful repairs. CUA-SWE provides a unified testbed for studying how agents use visual feedback and interaction to guide software engineering, with executable correctness criteria for the resulting software.

Large language models represent context with a key-value (KV) cache. Caches are model-specific: for the same text, models with different architectures or weights produce incompatible representations. This makes it costly to switch models over a shared context: although the context has already been processed by one model, the incoming model must process it again to build its own cache. We introduce KV-Lingo, a method for translating the KV cache of a source model into one that can be read by a target model. KV-Lingo consists of a collection of linear maps, typically one per layer of the target model, that are applied independently on all tokens' key and value representations. We train these maps using distillation, minimising the divergence between the target model's predictions from its native cache and those from the translated cache. We consider several model pairs spanning multiple sizes and architectures, training one translator per pair on a generic text corpus. The resulting translators preserve strong downstream performance in both small-to-large and large-to-small transfers. Since a switch then costs a linear map and a single decoding step instead of a prefill, replacing re-prefill with cache translation reduces the time to first token after a model switch by 9.6x already on a 64-token prompt for Qwen models on an Apple M3 Ultra, and by up to 29x at 32k context length on an H100. These gains make KV-Lingo particularly useful for dynamic model routing: a context can be processed by one model and handed off to another only when needed, without re-prefilling the shared prefix. We finally show that KV-Lingo can be used for seamless model switching, staying close to re-prefill across repeated switches in our multi-turn evaluations.

FullAttn repeatedly exposes the complete causal history to every attention head, creating substantial redundant computation and memory traffic even with IO-efficient dense kernels. We introduce CoWA, a structured attention architecture that distributes access to the causal history across KV heads. All heads share near-diagonal and prefix-sink windows, while complementary long-range windows partition the remaining history. Their union provides full causal coverage although each head attends sparsely to distant tokens. This position-defined attention pattern requires no learned router or indexer, is used consistently during training and inference, and aligns with KV-head tensor parallelism. A window-matched ablation at 8K isolates the effect of complementary long-range allocation: CoWA with 100% collective coverage reaches 89.73% accuracy, compared with 89.97% for FullAttn, while duplicated long-range windows perform substantially worse. Across a broader controlled associative-recall comparison with matched token budgets, CoWA closely tracks FullAttn as the context grows, whereas other sparse patterns lose a substantial fraction of the associations. In an attention-operator benchmark at 128K tokens with tensor parallelism, CoWA reduces forward and backward latency during training by 7.4x and 8.6x and decoding latency during inference by 3.0x over FullAttn. Its per-rank peak operator memory matches FullAttn during training and is 7.6x lower during decoding. Across scaling-law training from 0.6B to 14B parameters, CoWA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results show that full causal coverage can be a collective property of the head ensemble rather than a duplicated property of every head.

Clinical LLM evaluation often emphasizes answer accuracy; however, accuracy alone does not test counterfactual consistency or demographic robustness. We evaluated six LLMs on 150 MedQA USMLE questions using two automated perturbation tests to assess their performance. The counterfactual validity (CFV) test asked each model to make a minimal, plausible clinical change that would make a different answer correct. The demographic robustness test added six demographic prefixes to the same vignette and compared the answers and explanations with a no demographic baseline. Of the 900 CFV attempts, 228 (25.3 %) were valid and 672 were invalid. Across 5,400 demographic comparisons, 1,097 answers were changed (20.3%). Automated judging identified 3,128 stereotype evidence flags, including 1,932 in the broad Other category. MedGemma 27B achieved the highest accuracy (87.1%) and CFV (63.3%), lowest answer change rate (16.0%), and low mean Explanation Demographic Dissonance (EDD) score (0.169). However, its accuracy still exceeded its CFV, indicating that correct answers do not guarantee reliable performance on the counterfactual validity task. OpenBioLLM had the highest answer change rate and EDD, whereas GLM had the highest stereotype flag rate. These findings show that accuracy, CFV, answer stability, EDD, and stereotype evidence capture different evaluation aspects. Because all judgments were automated and no clinician validation was available, the results support safety screening but do not establish clinical deployability of the model.

Running large language model (LLM) agents in production gets expensive fast. A frontier model (the largest, most capable tier) is accurate but can cost 25 times what a small model costs per token, and the gap compounds once a workflow chains several calls together. Planner-as-Router (PaR) attacks this from a different angle. Instead of leaving model-tier selection to some component downstream, it folds the choice into planning itself. As the planner breaks a query into subtasks, it also assigns each one a model size tier (small, mid, or frontier, ordered by capability and price), so the dependencies between subtasks are visible before any specialist runs. Unlike per-call routers such as cascade routing, which look at one node at a time, PaR sees the whole workflow up front and needs no separate router model or training data. We evaluate PaR with EntBench, a benchmark of 54 enterprise agentic tasks across seven classes, graded by actually running the generated Structured Query Language (SQL) and MongoDB queries against live databases. Over 1,157 evaluations spanning eight routers and three seeds, PaR stays on the observed cost-accuracy frontier. It matches a sink-frontier heuristic (frontier model on terminal nodes only) in accuracy at comparable cost and a faithful FrugalGPT cascade at lower cost, and cuts cost 44% against all-frontier routing while giving up 2.9 points of accuracy. Several accuracy gaps fall inside the plus-or-minus six-point confidence interval of a 54-task study, so we frame PaR's advantage as frontier position rather than a clean accuracy win. We also report a preliminary observation, not a validated result: a small pilot hints that cheap routing may carry a hidden compounding penalty on compositional workflows, which we frame as a hypothesis for future measurement. PaR, EntBench, and all evaluation code are open source.

People typically assume that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training. We build a toy eval suite and show this mental model is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like computations. We call this mode-hopping. Across our suite, LMs suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment -- then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and cannot be fixed by checkpoint averaging. We instead think of it as a capacity allocation problem: in a capacity-bounded model, generalizable circuits must compete with the shallow ones learned early in training, and the data in each pre-training window may decide which circuits win. Our suite provides a new efficient lens on generalization. We demonstrate two concrete applications: (i) select intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than the final pre- or mid-training checkpoints, and (ii) select pre-training data that controls and stabilizes generalization dynamics.

In-context reinforcement learning (ICRL) enables large language model agents to adapt to new environments using their interaction history without updating model parameters. However, repeatedly conditioning on growing histories can lead to substantial token cost. We propose a bounded-history context-management framework that predicts the task-dependent downstream effect of removing historical interactions to guide history selection and determine a decision-dependent context budget. Formally, our framework uses the full rolling history as a reference. The predictor evaluates removal effects, defines a deletion ordering, and applies a shared selection criterion to determine how much history to retain at each decision. We evaluate the method in closed-loop SUMO driving under held-out in-distribution, unseen-domain, and unseen-route settings, and in ScienceWorld under a continual ICRL protocol. Relative to a baseline using the full context, our method reduces total token usage by 25.7%, 25.8%, and 23.2% across the three driving settings while maintaining comparable closed-loop driving performance. In ScienceWorld, it reduces total token usage by 52.1% compared to full context and uses 30.2% and 37.8% fewer tokens than the Recent and Similarity baselines, respectively, while maintaining performance.

Diffusion large language models (dLLMs) generate text by iterative unmasking. At each denoising step, a dLLM proposes a token at every masked position, but the decoder commits only a confident subset of these proposals. Trace-based on-policy distillation (TOPD) builds on this process by matching the student to a stronger teacher, yet only at the committed positions. We argue that this discards much of the useful signal, which resides in the uncommitted proposals, where the student has made a prediction but is not yet confident enough to commit it. We call these proposals hesitations. In our pilot study on an SDAR-4B student, hesitations make up only 24% of supervisable state-position pairs but carry 66% of the teacher-student divergence. To exploit this signal, we propose Hesitation-Aware On-Policy Distillation (HOPD), which extends teacher distribution matching to every masked position of each denoising step. Because hesitations are not equally informative, we further allocate supervision using hindsight from the completed trajectory, placing more weight on positions whose proposal was later disagreed with the final token and on blocks where first-step proposals rarely survive. Since both models already produce distributions at all masked positions, HOPD requires no additional forward passes over TOPD. The only extra cost is evaluating the loss at more positions. With SDAR-1.7B and SDAR-4B students distilled from TraDo-8B-Instruct, HOPD achieves the best average score among the evaluated methods on five math and coding benchmarks, under both static and dynamic decoding and at both scales. It also speeds up decoding. On SDAR-4B, the HOPD student hesitates less and commits 11% more tokens per denoising step than TOPD, while reaching higher accuracy.

Visual token compression lowers the inference cost of vision--language models by representing images with fewer tokens. However, most existing methods compress visual tokens to a reduced set, leaving the amount of visual evidence represented by each token and its original spatial context implicit. Therefore, the compressed representation does not explicitly encode how much visual information each representative carries or where it lies in the original image. This limitation arises even after a single reduction and becomes more pronounced when compression is repeated across decoder layers. To address this issue, we propose CoViST, a training-free framework that represents a compressed image as a composable visual state. Specifically, the state combines representative features with original positions, effective contribution weights, and reusable selection metadata. CoViST constructs this state through coverage-guided selection and conservation-based contribution composition, and explicitly incorporates its contribution and positional information into decoder attention. Each component of the state retains its interpretation under successive reductions, enabling the same formulation to support both fixed compression before prefill and progressive compression within the decoder. Experimental results on seven LLaVA-1.5-7B benchmarks show that CoViST-Fixed retains 99.9\%, 99.5\%, and 98.1\% of uncompressed performance at 192, 128, and 64 tokens, respectively, and CoViST-Pro retains 99.8\%, 99.9\%, and 99.1\% at the corresponding layer-average budgets, outperforming state-of-the-art methods under their respective budget settings. Code will be released publicly.

Autoregressive decode repeatedly streams a growing KV cache, making attention a major cost at long context. Existing high-performance kernels use online softmax, which discovers a row's normalization reference as it scans keys. Earlier contributions therefore remain provisional and may require rescaling. We argue that the reference need not be discovered: softmax is invariant to a common shift, so the reference only has to keep the weights in range. We present FoldAttention, an additive formulation of softmax attention that fixes a finite reference $Z_i$ before scanning the KV cache. Each weight $2^{s_{ij}-Z_i}$ is then final when computed, so contributions add across disjoint key ranges and their quotient equals softmax attention in real arithmetic. We use this property to develop two techniques for Hopper decode: (1) final weights gate key and value reads before the bytes are fetched, and a per-call depth $T$ cuts keys below $2^{-T}$ while keeping their mass, and (2) additive partials compose split KV and shared-prefix cascades without rescaling. On H100 at $T=16$, FoldAttention decodes seven real-model generations 1.36-2.30$\times$ faster than the fastest BF16 baseline, and up to 3.09$\times$ faster across MHA and GQA shapes, at an error within 1.5% of the lowest BF16 error on six of the seven; reading every key, it is 1.14-1.30$\times$ faster at matched error. We validate on Qwen3-8B that a whole decode step is up to 1.46$\times$ faster while likelihood and long-context accuracy match those under BF16 kernels. The same principle makes the backward deterministic: CTAs round bounded partial gradients onto an integer grid declared before the reduction and add them in any order. FoldAttention thereby removes the determinism tax: its deterministic backward is up to 1.84$\times$ faster than deterministic FlashAttention-3/4 and 1.05$\times$ faster than the fastest nondeterministic kernel.

Hybrid LLMs interleave full-attention layers with linear-attention layers to reduce long-context inference cost, but this structure complicates prefix caching. Full-attention KV caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing systems materialize recurrent-state checkpoints, restricting prefix reuse to checkpoint-aligned positions. We present SuffixReplay, the first prefix caching system that lets hybrid LLMs reuse cached prefixes at every cache-supported page boundary without materializing recurrent-state checkpoints. Our key insight is to just let linear states forget the distant past. Modern linear-attention mechanisms use recurrent decay and gating to attenuate the influence of old inputs. Therefore, instead of checkpointing every prefix boundary, SuffixReplay approximates the state at a matched boundary by replaying only a recent suffix of the layer's input hidden states, which we retain as anchors. At the algorithmic level, SuffixReplay combines layer-wise and token-wise anchor sparsity with a bounded replay budget to control storage, computation, and quality. At the system level, it uses an independently managed anchor sidecar and a pipelined replay path to overlap anchor movement and state reconstruction with the native serving pipeline. We evaluate SuffixReplay on three hybrid LLMs: OLMo-Hybrid-7B, Qwen3.5-4B, and Qwen3.6-27B-FP8. Across these models, SuffixReplay retains 91.4-100% of full-prefill quality on average across LongBench and RULER, while using only 0.36-0.51x the amortized per-token storage of SGLang's default 8192-token checkpoint cache. Integrated into SGLang, SuffixReplay reduces median TTFT by 15-70% on branching workloads, sustains 2.3-4.3x SGLang's throughput when the working set exceeds HBM, and matches SGLang on high-hit continuation traffic.

Open-search fact checking is not merely retrieval followed by classification, but a sequential decision problem in which every query, source visit, and stopping decision reshapes the evidence available for verification. Yet existing systems often distribute these decisions across predefined pipelines or separately prompted modules rather than learning them as a unified task-specific policy. We introduce \textbf{OpenFC}, a unified verification-policy training framework that post-trains Qwen3-8B as a compact next-action controller over reasoning, evidence acquisition, and stopping. OpenFC learns this policy in two stages. \textbf{Stepwise-Calibrated Cold Start (SCCS)} uses a strong training-time supervisor to review post-initial reasoning, tool-use, and stopping proposals before execution, producing reliable trajectories for supervised fine-tuning without access to gold verdicts. \textbf{Verification-Aware Reinforcement Learning (VA-RL)} then improves the cold-start policy on unresolved claims through budget-aware tool rewards, label-aware advantage reweighting, and localized response masking. Across six fact-checking benchmarks, OpenFC achieves 70.39\% average accuracy and 63.30\% macro-F1, the highest overall averages among the evaluated methods. Stage-wise ablations further show that SCCS and VA-RL provide complementary gains, supporting the design of the two-stage training framework. These results position OpenFC as a strong and effective framework for open-search fact-checking. We will open-source our code and release the model checkpoints to support reproducibility.

Molecular multimodal models support diverse understanding and generation tasks but may introduce safety vulnerabilities when handling hazardous molecules. In this work, We reveal substantial jailbreak vulnerabilities under both text-only and graph-conditioned settings. Our analysis further shows that safety robustness must hold across input modalities while balancing safety, over-refusal, and utility. To address these challenges, we construct SafeMolBench, a molecular multimodal safety-alignment benchmark with 3702 samples covering 618 unique hazardous molecules and safe molecular tasks, organized into hazardous-harmful, hazardous-allowed, and utility-replay subsets to support unified training and evaluation of safety, over-refusal, and utility. Based on SafeMolBench, we propose SafeMol, a parameter-efficient safety alignment framework that jointly optimizes lightweight modules across text-only and graph-conditioned inputs, uses MMD for distribution-level representation alignment to reduce modality-induced discrepancies, and explicitly models molecular hazardousness and harmful operational intent. Experiments on SafeMolBench show that SafeMol reduces attack success by several tens of percentage points while largely maintaining low over-refusal and preserving molecular-task utility.

Automatically generated environments provide a scalable source of interaction data for training general agents. However, existing approaches mainly generate tasks within a single environment, while real-world workflows require agents to connect information and actions across multiple services. We introduce Compositional Environment Scaling (\textbf{CompoWorld}), which expands the task space by composing a finite library of reusable services. Coding agents turn tool specifications into verified services with typed states and shared interfaces, while a world model handles tools that cannot be reliably implemented. A random-walk procedure connects services through dependency graphs, enabling the generation and verification of tasks that require information to flow across services. Verified trajectories support supervised fine-tuning (SFT), while our Completion-Focused Rubric Reward guides reinforcement learning (RL) toward full task completion by emphasizing criteria with lower pass rates within each rollout group. We construct 448 services exposing 10,130 tools and use 3K SFT trajectories and 1K RL tasks to train Qwen3.6-35B-A3B. Experimental results show that CompoWorld improves on its backbone by 9.17 points on average across eight benchmarks. On AutomationBench, it surpasses frontier models such as Claude Opus 4.6 and leads all compared agent-specialized 35B-A3B models.

Grounded language models are usually evaluated by adding relevant context, but multiturn dialogue also contains unsupported user claims that may contaminate later factual answers. We study post-pressure recoverability: whether a model returns to clean-context behavior after a user repeatedly advocates a wrong answer and then withdraws that pressure. We introduce a recovery-after-pressure protocol for multiple-choice factual dialogue and measure sycophancy hysteresis, the residual probability assigned to the user-advocated wrong answer relative to a clean-context counterfactual. Across seven instruction-tuned open-weight models and two factual benchmarks, ordinary reset often reduces but does not erase pressure-induced bias. History preserving repairs such as user retraction, system reset, and self-verification recover only 2-3/14 model-dataset pairs under the strict clean-restoration diagnostic, whereas operations that change the effective context are substantially more reliable; the two conditions that remove the pressure-bearing history entirely, fresh-context deletion and context truncation, recover 14/14. In an oracle trusted-evidence condition across fourteen model-dataset pairs, preserving the pressure-bearing history while adding benchmark-derived trusted evidence increases accuracy from 0.368 to 0.929, while wrong-answer following falls from 41.2% to 4.3%. Controls show that the effect is not explained by dialogue length, repeated confidence, plausible distractors, mere false-answer mention, or option-label inertia. These results suggest that faithful grounded dialogue requires evaluating which prior context should be treated as evidence and which should be removed or quarantined before answering.

A tool-using language-model agent deployed over a long stream of tasks receives no reward, so it cannot tell whether it succeeded, cannot safely retry, and cannot label the experience it needs to improve. We present SelfSuite, in which the agent's own base model, given only the world's public materials, designs a small evaluation suite of weighted judges and grounded per-task briefs, freezes it, and uses it to gate a keep-best retry and to label a typed, outcome-tracked memory. On matched five-repeat benchmarks over tau2-bench and AppWorld, SelfSuite scores above the plain agent without any labels, matches methods given ten expert labels on tau2-bench, and trails Agentic Context Engineering (ACE) on AppWorld, where code execution gives a direct success signal. In an ablation campaign run on the same tasks, it is above label-free ACE in every repeat, and the gated second attempt is the only component whose removal hurts in every repeat. We also simulate a subject-matter expert who grades ten onboarding tasks per world. Using those labels to calibrate SelfSuite's evaluator gives a small, consistent gain, and using them to warm up ACE's memory lifts ACE to tie calibrated SelfSuite. A single-run study on a second model family shows the same ordering.

At each decoding step a language model attends over the key-value (KV) cache of every earlier token, so at long context the attention call is bounded by memory bandwidth. Sparse attention reads only a subset of keys chosen by a cheap score estimate, and most methods give the unread tokens zero weight. The output then draws on only a small fraction of the KV cache, and accuracy drops at small budgets, most on tasks that aggregate information across the context. An inverted-file product-quantization (IVF-PQ) index over the cached keys computes an approximate score for every indexed token in order to rank them; after ranking, those scores approximate the attention logits of the tokens left out. PQ-HSA (hybrid sparse-approximate attention) attends the selected tokens with their original keys and values, and the unselected tokens, the background, enter the same softmax through those scores, summed per inverted list and multiplied by the list's mean value. At 128K and a 1-2% retrieval budget, PQ-HSA is more accurate than Quest and SnapKV on Llama-3.1-8B and Qwen3-30B-A3B and stays close to full attention in macro accuracy; with the same selector, the background term raises macro accuracy on the 8B model from 0.71 to 0.83. In the same 128K setting, inside vLLM on one NVIDIA H20, the decode attention call runs 1.6x faster than the FlashAttention-3 kernel; the speedup grows with context length, and a cost model fitted on 8B to 30B models gives the context length at which it begins. A vLLM plugin runs PQ-HSA on two engine versions without changes to the engine source; code is available at https://github.com/KunmingSHAO/pqhsa_release.

LLM agents increasingly solve long-horizon tasks by autonomously interacting with their environment. In doing so, their strategies rely on assumptions about that environment: which resources and tools exist, where they are located, and how they behave. When these assumptions no longer hold, reliable agents must detect the change and adapt while pursuing the same goal. We study this adaptation capability through environmental novelty: a change that keeps the task objective fixed while invalidating an assumption underlying an otherwise successful trajectory. We introduce AGNI, an automated pipeline that extracts trajectory-relevant assumptions, injects targeted environmental changes, and validates that the resulting novel tasks remain solvable. Across three terminal benchmarks, AGNI produces diverse novelties spanning resources, interfaces, constraints, and execution semantics. Evaluating multiple LLM agents reveals a substantial adaptation gap between base and novel tasks. Trajectory analysis suggests that agents often encounter evidence of the change but fail to diagnose its cause and revise their strategy. Finally, post-training for environmental novelty improves adaptation to held-out novel tasks while also improving performance on base tasks. Our results highlight a gap between task competence and adaptive capability and motivate environmental variation as a core dimension of agent training and evaluation.

Block-diffusion language models are served at hand-picked operating points, such as acceptance thresholds, buffer depth, schedule, checkpoint and precision, and each point is chosen by its mean benchmark accuracy. However, a mean does not tell an operator how often a faster configuration fails on prompts that the slower one answers correctly. On the serving engine and its decode traces, the default commit rule already commits every fully resolved block, a static skip rule captures nearly all of the compute that allocation can save, and self-distillation on engine-decoded targets adds speed at unchanged accuracy. Larger speedups come from lower thresholds, which commit tokens that are still uncertain. We therefore present Redline, a finite-sample procedure that selects operating points, hand-picked or learned, from the correctness of their answers on calibration prompts. Redline keeps the reference-relative risk, the joint probability that the reference answers correctly and a candidate configuration does not, within a user-chosen budget with high probability, and deploys the fastest configuration that passes. It speeds up math at a smaller risk budget than code in both model families, and at a budget of ten percent it deploys a LLaDA2 math configuration that commits over a third more tokens in each forward. It also applies without modification to the acceptance rule of speculative decoding and to weight quantization. On the same calibration data, Redline stays within its stated failure probability, whereas each tolerance of a mean-accuracy rule either gains less speed for some model and task or exceeds the risk budget far more often for another. Code is available at https://github.com/js-lee-AI/Redline.

Diffusion language models (DLMs) promise fast parallel generation, yet high-quality samples often require large number of refinement steps, which diminishes their advantage in practice. This has led to massive interest in and rapid development of new methods for effective few-step generation. We show that much of the supposed quality gap at few steps can instead arise from a suboptimally configured sampler. Modest sampler sharpening, without any model retraining, enables a couple years old masked DLM to rival supposedly far improved successors. This differently sampled DLM in fact achieves lower generative perplexity in just 16 steps than what its standard sampler obtains with 1024, while improving both judged quality and semantic diversity. We further show that conventional per-output metrics can fundamentally obscure these gains, since any optimal trade-off between two such metrics can be attained by a generator supported on at most two outputs. We subsequently introduce GroupEval, which separately evaluates quality and across-output semantic diversity, and offers fresh insights including uncovering how 1.5-4.7x perplexity gains of a distilled model yield no corresponding quality gain. Finally, we explain why sharpening helps: parallel unmasking destroys dependencies among simultaneously generated tokens, creating a gap between prediction and generation. We prove that pervasive temperature choice of one is generically suboptimal under parallel sampling even for an exact denoiser, and that worse predictions can yield better samples. Through these results, we argue for a broader evaluation principle of treating the deployed generator as the object of comparison, benchmarking it against tuned baselines, and assessing quality and diversity jointly and with more human-aligned measures.

Many LLM-as-a-judge evaluations score fixed outputs under a fixed task definition. Production multi-turn business agents instead require a maintained measurement system: correctness depends on business-specific facts and procedures, outcomes emerge across turns, and failures must be attributed to either agent capability or missing business knowledge before they are actionable. We present an integrated methodology spanning evaluation specification, modular LLM judges, intent-preserving user simulation, and human-in-the-loop governance. The specification defines conversation-level end states and actionable failure ownership. Atomic judges share versioned evidence and feed an explicit aggregation graph. The simulator is released only after task-preservation and stability checks. Independent human audits estimate measurement fidelity, renew tiered reference sets, and route disagreements to label correction, guideline revision, or judge improvement. Production studies show that system-level fidelity improved across repeated audits, that human reviewers and automated judges improved together under the shared feedback loop, and that their combined workflow had the strongest descriptive performance in both reported task-completion settings. Because the studies are observational and the human reference itself required revision, these findings demonstrate operational usefulness rather than causal or universal superiority. The contribution is a practical framework for making multi-turn agent measurement reliable, actionable, and maintainable as the evaluated system and its evidence evolve.

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.

Large language models (LLMs) have demonstrated strong capabilities in biological reasoning; however, genomic disease inference remains largely dependent on memorized gene-disease associations rather than understanding biological pathways. This shortcut learning undermines robustness and generalization, and breaks down when molecular identifiers are unavailable. We present GenoMorph, a multimodal genomic reasoning framework that shifts disease prediction from associative gene-disease mapping toward pathway-grounded reasoning. GenoMorph couples a frozen DNA foundation model with question-conditioned cross-attention fusion, self-adaptive latent reasoning (LatentSp), a residual reasoning gate for iterative genomic evidence reinjection, and rejection sampling fine-tuning regularized by hierarchical optimal transport (OT). Rather than learning direct gene-disease mappings, GenoMorph aligns genomic sequence representations with latent pathway dynamics, enabling reasoning trajectories that follow molecular interactions before producing disease predictions. LatentSp dynamically allocates computation according to reasoning confidence, reducing unnecessary reasoning steps and improving inference efficiency. We further construct an anonymized benchmark from the Kyoto Encyclopedia of Genes and Genomes (KEGG), replacing every gene and molecular identifier with anonymous symbols while preserving sequences and pathway topology, thereby removing memorization shortcuts. GenoMorph raises the weighted F1 from 0.7863 (BioReason) to 0.9412, and rejection sampling fine-tuning with self-adaptive latent reasoning pushes it to 0.9725 while cutting latency nearly 60%. On the anonymized benchmark it reaches 0.9465 F1, substantially outperforming prior systems and confirming that accurate disease prediction can arise from pathway reasoning rather than memorized gene-disease associations.

Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they still suffer from overthinking, generating redundant reasoning steps which incur substantial token consumption. Existing methods, such as suppressing reflective keywords or forcing shorter reasoning lengths, attempt to mitigate this issue but inevitably truncate necessary steps and induce underthinking, thereby compromising performance. To address this dilemma, we investigate the latent representations and observe that efficient reasoning steps naturally cluster into a concentrated region in latent space, while those deviating from this region tend to produce verbose sequences. To leverage this, we keep reasoning focused within this region via a quadratic program which projects deviating hidden states back into the region. Then we propose a novel training-free framework to achieve efficient reasoning that reduces token generation costs without sacrificing performance. Extensive experiments conducted on four models ranging from 1.5B to 14B, and across six benchmarks in math reasoning, coding, and scientific QA, validate the effectiveness of our method, up to a 12.1\% improvement in accuracy while reducing generated tokens by 11.8\% to 52.8\%. Codes are available at \href{https://github.com/hzn18/Opt4Reasoning}{https://github.com/hzn18/Opt4Reasoning}.

Self-evolving language models improve by generating tasks and learning from their own feedback, but adapting the task generator often requires a separate challenger-training loop. Can we generate tasks adapted to the current solver without explicitly training a challenger? We introduce \textbf{D}irect Self-\textbf{E}volving \textbf{O}ptimization (DEO), which replaces challenger parameter updates with solver-guided task sampling. The KL-regularized challenger objective defines an exponential tilt of a fixed base task distribution. DEO uses this distribution as a sampling target: a frozen LLM generates and mutates tasks, the solver scores them, and an approximate Metropolis selection rule refines the training pool. Only the solver is trained. Theoretically, for an idealized variant that samples exactly from the tilted distribution, and under regularity, local gradient-dominance, and initialization conditions, we show that DEO learns distributionally robust reasoning ability. In experiments, DEO achieves reasoning performance competitive with R-Zero while using over $50\%$ less wall-clock training time, and improves reasoning accuracy over a no-walk ablation. Replacing the task generator with a frozen API-only LLM further improves the local solver, illustrating a capability enabled by removing challenger training.

Rubric-based tasks are increasingly addressed through reinforcement learning (RL), with rubric scores used as training rewards. However, these rewards typically supervise final answers without distinguishing the contributions of intermediate decisions. Many existing credit assignment methods rely on ground-truth answers to define process rewards, limiting their applicability to open-ended tasks without canonical solutions. To address this limitation, the proposed rubric-grounded credit uses task requirements as a shared reference for final answer evaluation and process supervision. The information returned by tools is assessed for the additional support it provides toward satisfying each rubric relative to that rubric's history of accepted support. By referencing these histories, credit distinguishes new support from evidence already present in the trajectory while recognizing partial support for each rubric. Dr.Credit uses rubric-grounded credit to supervise intermediate tool turns in an RL framework for deep research agents. The resulting process advantages are combined with GRPO outcome advantages to guide research decisions while retaining supervision of final-report quality. Evaluations on four in-domain and out-of-domain benchmarks show that Dr.Credit outperforms the evaluated open deep research baselines on every primary metric and submetric. Meanwhile, with an 8B-parameter backbone, the trained agent achieves average performance competitive with the evaluated frontier proprietary models. Further analyses suggest more efficient evidence acquisition and higher-quality reports under limited research-turn budgets, motivating the extension of rubric-grounded process supervision to a broader range of rubric-based tasks.

Maintaining persona consistency across multi-turn dialogues remains a core challenge for role-playing language models. Off-policy distillation from external teachers incurs distribution mismatch that compounds across dialogue turns, while reinforcement learning struggles with reward ambiguity inherent in subjective persona fidelity. We propose OSPD, an on-policy self-distillation framework where the same model serves as both teacher and student under asymmetric information: the teacher receives a complete character profile while the student sees only a brief summary, and the student generates trajectories from its own policy. We find that teacher confidence in role-playing dialogue exhibits a bimodal structure---sharply peaked at character-critical tokens yet diffuse at generic utterances---and introduce role-aware divergence switching to match this structure. A progressive trait masking curriculum further forces staged internalization of character knowledge along semantic dimensions. Experiments on CharacterBench, CharacterEval, and SocialBench show that OSPD substantially improves persona consistency over supervised fine-tuning and multi-turn RL baselines, without requiring any external teacher or reward model.

Looped Transformers reuse a shared block across recurrent depths, making autoregressive decoding expensive because every generated token requires many sequential recurrent passes. Self-speculative decoders reduce this cost by drafting at an early depth and verifying at full depth, but typically bind draft computation to prefix representations from the same recurrent depth. We find that queries and keys approach their final-depth representations earlier than values, and controlled prefix-channel interventions show that mature values substantially improve shallow draft predictions. Motivated by this asymmetry, we introduce Depth-Asynchronous Self-Speculation (DAS), which decouples the depth of draft computation from the depth of verified-prefix representations it reads. Its Mature-V primitive lets shallow queries retrieve full-depth prefix values without additional recurrent computation. We further develop DAS-Wave, which combines depth-asynchronous prefix reads with carried parallel refinement, progressive block growth, and an independent full-depth verifier. Across four recurrent-model checkpoints and mathematics and code workloads, DAS-Wave achieves 4.00--6.96$\times$ mean throughput speedup over paired full-depth autoregressive decoding in the same inference stack. These results identify prefix-information depth as an effective design axis for recurrent self-speculation.

On-device streaming omni-modal inference safeguards user privacy and eliminates prohibitive per-token API costs, but faces a critical bottleneck: the continuous influx of multimodal data rapidly exhausts constrained memory and compute budgets via monotonic KV cache growth. Existing sparse attention methods fall short, either incurring prohibitive online estimation latency or destroying interleaved cross-modal context, while failing to resolve physical memory fragmentation. We present OmniTide, the first algorithm-system co-design tailored for efficient on-device streaming omni-modal inference. Driven by the observation of modality-aware structural sparsity, OmniTide adopts a unit-based abstraction with two components: (1) At the algorithm level, OmniPick logically retains critical multimodal context based on unit boundaries and modality importance to preserve task accuracy; (2) At the system level, OmniPage physically partitions the cache by retention likelihood and dynamically compacts surviving sparse tokens, minimizing both memory fragmentation and data-movement overhead. Extensive evaluations across three streaming benchmarks and two consumer-device architectures show that OmniTide achieves up to $12.72\times$ kernel speedups and $2.40\times$ lower stream-loop latency. On StreamingBench, it improves accuracy by up to 18.0 percentage points over sliding-window baselines at comparable session cost. OmniPage further reduces the physical KV span by up to 26.7% relative to native logical eviction, unlocking real-time, infinite-context streaming on edge devices.

As large language models increasingly operate as tool-using agents, post-jailbreak safety feedback is often assumed to serve as a reliable safeguard; however, how lingering jailbreak context shapes subsequent agent behavior remains largely unexplored. To systematically examine this dynamic, we introduce a paired continuation framework across 192 parent tasks spanning 42 domains, evaluating 12,148 analyzed continuation pairs (curated from a 12,288-pair initially design) across eight diverse agents. We find that identical safety feedback induces sharply model-dependent behavioral routing rather than uniform protection: redirecting unsafe trajectories toward legitimate completion (\emph{rescue}), sustaining unauthorized execution (\emph{persistent unsafe}), or triggering over-refusal on benign tasks (\emph{collateral loss}). Through layer-wise activation patching, we discover a shared \emph{late-commit pattern} where causal intervention effects surge sharply near the final layers (relative depths of 0.958--0.984) despite an over 30-fold variation in peak magnitude across architectures. Crucially, critical-layer representations correlate with macroscopic routing outcomes, and intervening at these layers causally alters concrete next-step tool actions. Building on this causal foundation, we test whether localized intervention-derived features can serve as predictive proxies for full-trajectory routing outcomes on unseen parent tasks under leave-one-parent-task-out evaluation, finding that they provide viable predictive signals in responsive agents with peak ROC AUCs reaching 0.675 for \emph{rescue}, 0.777 for \emph{collateral loss}, and 0.702 for \emph{persistent unsafe}. These findings establish a mechanistic lens and a predictive baseline for anticipating the safety and utility trade-offs of post-jailbreak feedback in autonomous agents.

Practical deployment of large language model (LLM) agents requires strong task performance at affordable inference cost. For long-horizon agentic tasks, this performance-cost trade-off can be improved through within-task large-small model collaboration, as smaller models can handle some stages even when they cannot solve the full task. In this paper, we introduce RSI-router, a routing framework that constructs subtask-level model assignments and model-specific skills through recursive self-improvement over accumulated experience. Each iteration consists of four stages: Subtask Mining derives subtask definitions and identification rules from training trajectories; Routing Strategy Evolution proposes and evaluates diverse model assignments; Model-Specific Skill Evolution compares routed and large-model-only trajectories to diagnose failures and develop reusable execution skills; and Pareto-Optimal Router Selection updates the Pareto population using historical and newly generated routers while retaining dominated routers as experience for subsequent evolution. Routing between DeepSeek-V4.1-Flash and Qwen3.5-9B, RSI-router consistently surpasses the DeepSeek-only baseline at roughly half the inference cost (48.3%) across five agentic benchmarks. In particular, on ALFWorld, ScienceWorld, and WebShop, it cuts inference cost by 74.7-82.2% while simultaneously improving performance; on Terminal-Bench 2.0, it achieves a 16.7% relative performance gain at 18.0% lower cost. Moreover, RSI-router establishes a stronger performance--cost Pareto frontier than 9 routing methods.

Preference-based alignment methods such as Direct Preference Optimization (DPO) use pairwise preferences labeled by human annotators to fine-tune language models. However, annotators carry systematic biases toward some attributes: a name that signals a gender or an ethnicity, a persona, a language variety, a formatting convention, or length. If not properly addressed, these systematic biases can be absorbed and amplified during alignment. Existing methods address length bias or annotator disagreement, but fail to eliminate biases toward arbitrary attributes. To address this limitation, we propose Bias-Adjusted DPO (BA-DPO), a generalization of DPO that adds one bias parameter per annotator toward responses carrying a declared attribute. We prove that the objective is convex in the bias parameters and that the votes identify each annotator's bias up to a shared constant. The remaining constant is what fixes the aligned model's attribute rate: by default the rate of the reference model, or a target rate, which we use to bring a biased policy to statistical parity. On a corpus with planted biases, DPO drives the attribute from a balanced start to probability 0.96 and BA-DPO removes 81 to 95\% of that shift; on MultiPref with real annotators it removes about half of DPO's lengthening. Both hold at 0.5B with full fine-tuning and at 8B with LoRA, at no higher KL than DPO and no loss in judged quality.

Large Language Models (LLMs) offer a scalable way to simulate survey respondents using demographic profiles and observed reference responses. However, the conventional approach of predicting one question per prompt repeatedly encodes the same context, limits each target to a narrow set of reference responses, and prevents later predictions from using information in earlier answers. Predicting multiple questions in one prompt can reduce these costs, share a broader pool of references, and let later predictions build on earlier ones. This requires forming coherent batches, selecting shared references, and ordering questions and references effectively. We propose Semantically Coherent Batching and Ordering (SCBO), a training-free framework that addresses these challenges. SCBO first uses an LLM to extract compact semantic representations from survey items and filter out template noise. It then groups related questions into batches and builds a shared reference bank using target-specific retrieval and centroid-based completion. Finally, it orders target questions from easy to hard and arranges references according to their semantic alignment with those questions. Experiments on four large-scale survey datasets and four LLMs show that SCBO substantially reduces token consumption and inference time while generally improving prediction accuracy over a non-batched baseline. Code is available at https://anonymous.4open.science/r/SCBO-41D8.

Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a $1.69\times$ attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.

Identifying intertextual references is central to literary scholarship, but computationally difficult when source material is transformed through paraphrase, allusion, historical language, and translation. We investigate this problem through biblical intertextuality in Karen Blixen's Seven Gothic Tales. Drawing on the commentary to a critical edition, we construct a benchmark of 189 annotated references and evaluate retrieval against all 31,170 verses of historically plausible Danish Old and New Testament translations. We compare TF-IDF and BM25 with multilingual and Danish sentence encoders, examine the effect of linguistic normalization, and fine-tune a Danish encoder using hard negatives and five-fold cross-validation. We analyze performance across automatically derived lexical-overlap strata representing quotations, paraphrases, and allusions. Linguistically normalized BM25 provides a strong zero-shot baseline, attaining an overall R@10 of 0.365 and retrieving every quotation within its ten highest-ranked verses. The best zero-shot dense model achieves a comparable overall score of 0.360 while performing better on allusions. Fine-tuning DFM-large raises its overall R@10 from 0.265 to 0.508 and more than doubles its performance on allusions, from 0.138 to 0.339. However, evaluation against editorial annotations alone understates the model's scholarly usefulness: a literary scholar judged seven of 30 selected rank-one predictions counted as false positives to be meaningful additional references. These findings show both the potential and the epistemic limits of computational intertextual retrieval. Rather than treating scholarly annotations as exhaustive or model outputs as discoveries, we propose retrieval models as heuristic co-readers that recover documented references and generate candidates for expert-led close reading.

Crystal plasticity (CP) simulations predict the mechanical behavior of polycrystalline metals, yet their routine use is hindered by the manual effort of configuring heterogeneous tools, orchestrating multi-step data pipelines, and calibrating constitutive parameters against experiments. These bottlenecks impede productivity in systematic parameter studies, motivating interest in automated workflows. This study presents CP-Agent, a harness-engineered LLM-based agent that autonomously executes complete CP modeling workflows from natural-language tasks. Operating under the ReAct paradigm, the agent reasons about tool selection and sequencing while delegating numerical search to established optimizers. The harness comprises a minimal system prompt, typed tool definitions, a dispatcher, and a safety-bounded iteration loop, encoding domain knowledge through tool schemas rather than hard-coded logic. CP-Agent is demonstrated on four case studies: calibrating four slip parameters of additively manufactured stainless steel 316L against tensile data; validating the workflow against published copper benchmarks, reproducing stress-strain and texture evolution; recovering the initial crystallographic texture of copper, where the agent correctly identifies a diffuse initial texture; and reproducing the multi-pass rolling texture evolution of a Mg-Zn-Ca alloy, where the agent chains five deformation passes and recovers the experimentally observed weakened, split basal texture. In all cases, the agent inferred the correct execution sequence from the task statement, robustly across repeated runs, and delivered physically interpretable results. This work establishes harness engineering as a systematic approach to automating CP modeling workflows while maintaining physical interpretability and auditability through visible reasoning traces.

This paper introduces model casting, a mid-training recipe that drastically sparsifies the activations within the Feed-Forward Network (FFN) layer. With this strategy, at inference time, we first compute the output of the gating matrix and, thanks to its high sparsity, we avoid computations with the two other matrices, reducing the FLOP count by up to 3x. While this theoretical speedup is an upper bound, model casting translates into significant speedups both on CPU and GPU. We then introduce LoPA Gating, a new FFN design that increases the maximum theoretical speedup. It is a low-FLOPs parameterization of the gating matrix that overcomes the 3x cap by allocating fewer FLOPs and parameters to the gating matrix, compared to the two other FFN matrices that are sparsely activated. We consider two cases: (i) we cast a pre-trained model with a sparsity inducing activation; (ii) we train with LoPA from scratch. In all settings, we significantly outperform existing pruning solutions and regular RELU-fication. For instance, at matched quality, we achieve a 3.2x FLOP speedup with LoPA Casting, against 1.6x at best for competing methods top-p and TEAL. Using dedicated kernels, we achieve an actual 3.31x speed-up on GPU at 90% sparsity, past the 3x ceiling of standard gating; RELU-fication, meanwhile, plateaus below 80% sparsity.

As autonomous LLM agents are deployed across multi-session environments, conventional memory architectures suffer from Associative Blindness (inability to traverse multi-hop relational dependencies), Scaffolding Amnesia (temporal decay evicting core persona invariants), and Static Topology Stagnation (immutable graphs ignoring usage dynamics). Grounded in Complementary Learning Systems (CLS) principles, we propose EngramRAG, an adaptive memory architecture coupling a low-latency Waking State reflex with an asynchronous background Dreaming State consolidation cycle. EngramRAG introduces: (1) Usage-Modulated Personalized PageRank (U-PPR), where transition probabilities adapt via Hebbian plasticity to promote persistent entities into high-centrality Epistemic Macro-Hubs; (2) Consolidation-Activated Topology Decay (CATD), which scales retention half-life by topological load-bearing weight rather than wall-clock recency, protected by a cold-start grace period (N_grace >= 4); (3) Directed SUPERSEDES DAG filtering to suppress obsolete state during fact mutations; and (4) Triple-source hybrid retrieval fusing dense vectors, BM25, and U-PPR via dynamic Reciprocal Rank Fusion (RRF). Evaluating on all 1,982 QA pairs across 10 long-term conversations in the LoCoMo benchmark, EngramRAG achieves +38.9% relative improvement in Recall@5 (53.21% vs. 38.29%, p < 0.001) and +43.1% in MRR (0.4203 vs. 0.2937) over dense vector RAG, significantly outperforming Okapi BM25 (48.66%) and isolated static graph retrieval (8.50%). On temporal reasoning, EngramRAG reaches 62.33% Recall@5 (+16.67 points over dense vectors). In controlled mutation tests, SUPERSEDES suppresses split-brain hallucinations from 70.0% to 0.0%, while 90-day simulations show 100.0% scaffolding retention under a 26.21ms interactive retrieval reflex.

Powered by expert guidance, agents can operate in interactive environments; however, it is unclear whether they can learn autonomously from their own experience. To evaluate such self-improvement methods, we introduce GameBoyWorlds, a testbed for agentic self-improvement in video games. GameBoyWorlds-Execution evaluates task execution on a collection of 5 distinct game series. Agents are allowed access to dedicated training games but are provided no demonstrations, documentation, or rewards. Agents must ground themselves in the environment through self-directed exploration and by inferring actionable knowledge from their own experience. At test time, agents must complete short-horizon tasks that evaluate their ability to navigate, interact, and engage with game-specific mechanics in unseen games. Out-of-the-box frontier models complete fewer than 50% of the 500 tasks due to failures in multimodal grounding, establishing that self-improvement methods have room to push performance. We demonstrate that contemporary approaches to self-improvement are lacking, with world modelling and autonomous skill discovery failing, and a novel strategy that uses curiosity-based exploration to write guides achieving only partial success. GameBoyWorlds-Playthrough tests end-to-end game completion in two fan-made Pok\'emon games. We show that while frontier models have been pre-exposed to official releases such as Pok\'emon Red, they lack essential information on the games in our testbed. Instead of relying on their parametric knowledge to succeed, agents must learn from their own experience and autonomously improve over the course of the playthrough. We show that a sophisticated agentic pipeline with multimodal memory and hierarchical subgoals fails to reach even the first major milestone in both games, establishing GameBoyWorlds as an ambitious target for self-improving agents.

Precise instruction-following is a fundamental ability of large language models (LLMs), requiring their outputs to strictly satisfy objective constraints in input instructions. In complex application scenarios, these constraints often possess diverse scopes that govern specific response segments rather than the entire output. However, existing optimization methods often neglect constraint scope during data construction and rely on binary per-constraint rewards, yielding limited data diversity and sparse supervision for complex constraints. To this end, we propose ScopeIF, a novel training framework for scope-aware precise instruction-following. We first introduce a unified schema that factorizes objective constraints into three decoupled dimensions: Scope, Target, and Range. Grounded in this schema, we construct ScopeInstruct, a large-scale instruction dataset with diverse scope-aware constraints, and combine tool-grounded verification with graded reward modeling to quantify the violation degree of each constraint, providing dense supervision for policy optimization. Extensive experiments demonstrate that ScopeIF consistently outperforms existing methods, particularly on complex scope-aware constraints, while preserving general capabilities. Notably, it enables optimized Qwen3-4B and 8B models to rival or surpass strong frontier models such as Gemini-2.5-Pro and DeepSeek-V3.2, establishing an effective paradigm for advancing scope-aware instruction-following. Our code and data are available at https://github.com/thu-coai/ScopeIF.

Direct Preference Optimization (DPO) aligns language models by optimizing over sequence-level sums of token-wise implicit reward differences. However, we identify a pervasive pathology in this formulation: a disproportionately small subset of high-frequency token types dominates cumulative sequence scores while appearing symmetrically across both preferred and dispreferred responses. Specifically, under canonical Qwen tokenization on Anthropic HH-RLHF, merely 69 token types account for $55.1\%$ of all response tokens and $85.9\%$ of within-pair shared token mass, exhibiting substantially lower preference-side specificity than the remaining vocabulary. This symmetric ubiquity induces gradient entanglement and dilutes the discriminative preference signal propagated through the objective. To resolve this issue, we introduce \emph{Anisotropic DPO} (\textsf{ADPO}) and its canonical realization, \emph{Frequency-Hard DPO}. Using a fixed, label-agnostic vocabulary mask, our method zeroes the implicit reward contribution of high-frequency response tokens while assigning unit weight to informative positions, thereby suppressing gradient interference without modifying preference pairs, discarding context, or introducing learned parameters. Here, \emph{anisotropy} designates non-uniform token-level objective weighting rather than representational geometry. Extensive empirical evaluations on AlpacaEval, MT-Bench, and Arena-Hard demonstrate that Frequency-Hard DPO consistently outperforms standard DPO across Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct, establishing that selectively masking shared high-frequency tokens offers an effective, zero-overhead mechanism for robust preference alignment.

Large Language Models (LLMs) are highly capable of natural language reasoning and appear to store a great deal of lexical knowledge, but it is still unclear how much of this knowledge they actually use when reasoning, and whether they use it in the right way. On the other hand, logic-based Natural Language Inference (NLI) systems provide transparent and formally grounded reasoning, but they need to be supplied with rich lexical knowledge to prove inferences beyond purely logical ones. In this paper, we evaluate whether LLMs can identify all lexical knowledge needed to solve NLI problems and how much this knowledge contributes to proof search in a logic-based NLI system. Our research focuses exclusively on structured lexical entailments (e.g., chinchilla$\sqsubseteq$small animal) as a proxy for structured explanations for NLI problems with an entailment label. First, we curate a dataset for a new task of explaining sentential entailments with a set of lexical entailments. The dataset is used to intrinsically evaluate LLMs on generating structured lexical explanations. Then, we use NLI as an extrinsic evaluation in a simple neuro-symbolic setting, assessing whether LLMs can supply sufficient lexical relations to LangPro, a natural-logic theorem prover for natural language. The results show that the proposed task remains challenging even for hosted proprietary LLMs, and that their contribution to theorem proving is moderate: generated relations are often only partially sound and may be tailored to the specific NLI problem rather than representing generally valid lexical knowledge.

Large language models are commonly evaluated by generating a small number of stochastic responses for each prompt in a benchmark. Because inference budgets are limited, this shallow evaluation may fail to observe low-probability but operationally important failures. A prompt that produces no failures in a small sample may therefore appear reliable despite having a nonzero latent probability of failure under repeated inference. We formulate LLM reliability evaluation as a budget-constrained discovery problem in which each prompt is associated with an unknown per-generation failure probability. We propose a budgeted discovery framework that first performs shallow evaluation across the prompt set and then uses trial-level failure outcomes together with prompt-derived representations to learn a feature-based ranking of failure propensity. The resulting scores prioritize prompts with zero observed shallow failures for deeper evaluation, concentrating the deep-evaluation budget where hidden failures are more likely to be discovered. We evaluate this approach on AIRBench and StrongREJECT across multiple model and system-prompt conditions. The central empirical test asks whether models fit without access to deep-evaluation outcomes can rank prompts with zero observed shallow failures according to their likelihood of producing failures under deeper evaluation. On AIRBench, the highest-ranked 10\% of unresolved prompts achieves 2.54x hidden-failure lift for Qwen 2.5 7B and 1.87x for Gemma 3n E4B, recovering 25.4\% and 18.7\% of subsequently observed hidden failures, respectively, compared with 10\% expected under random allocation. Semantic-neighborhood and feature-ablation analyses further show that this predictive signal can be recovered from multiple representations of prompt content and relationships.

On-policy self-distillation (OPSD) trains a question-only student with token-level feedback from a teacher given training-only privileged information (PI). OPSD therefore provides dense, on-policy supervision, and is free of a larger external teacher, but its effectiveness rests on how PI is designed and utilized. Our preliminary diagnostics suggest a significant gap between teacher utility and student learnability, where a small fraction of high-disagreement tokens dominate the distillation signal, and short teacher continuations at these positions further expose more explicit PI leakage than transferable correction cues, indicating a strong intent on injecting PI-conditioned shortcuts. We propose Adaptive On-Policy Self-Distillation (AOPSD), which adapts what information the teacher receives and how strongly its feedback influences learning. AOPSD encodes each solution as a reasoning DAG, orders problems by the student's evolving capability, and reveals only the affordable subgraph and its next frontier as PI. For high-disagreement tokens, AOPSD utilizes short teacher continuations as probes to encourage useful guidance while mitigating PI-conditioned shortcuts among teacher supervisions. On HMMT25, AIME24, AIME25, and BRUMo25, AOPSD achieves 72.5% Pass@8, which is 6.7 percentage points above OPSD and 4.2 above the strongest competing baseline while reducing 15 percentage points of training time at lower cost.

On-policy self-distillation densifies agent training without external teachers: a policy conditioned on privileged hindsight provides step-level guidance for its own unprivileged rollouts. For search agents, however, hindsight can make the teacher prefer a query that does not improve retrieval from the student's state. Existing methods either distill this preference directly or filter it with model-internal scores, but neither strategy verifies the query's executed retrieval consequence. We propose Information-Gain-Gated Self-Distillation (IGSD), which verifies on-policy token proposals with environment feedback before distilling them. Treating each query token as a micro-action, IGSD completes the teacher's token proposal and the student's sampled token into matched queries and executes both from the same failed state with the same retriever. Shared counterfactual controls account for query-conditioned shifts in answer likelihood, so their difference, the executed paired information gain, provides a relative utility contrast for the retrieved documents. IGSD uses this contrast as a positive-only soft weight for candidate-pair distillation, while leaving the GRPO objective unchanged and confining verification to training. Across seven single-hop and multi-hop QA benchmarks, IGSD reaches macro-average exact-match accuracies of 42.8% and 47.0% with 3B and 7B policies, respectively, without inference-time verification. These results support environment-verified hindsight as an effective approach to reliable action-level supervision for search agents.

FullAttn often assigns negligible normalized mass to much of the causal score space, yet dense kernels execute the complete post-score path after forming each QK tile. We introduce MALA, a fused attention primitive that preserves score access to every legal causal interaction and uses normalized contribution to allocate post-score computation. Forward uses its evolving online-softmax normalizer, while backward reuses the finalized normalizer to derive nested retained support using only standard attention state. A common tolerance governs training and inference, allowing for adaptive retention of the work. MALA reduces low-contribution post-score computation. A matched-work study at 8K isolates the benefit of distribution-adaptive allocation: under exactly matched total post-score work, MALA approaches a per-instance reference-mass oracle, with mean omitted mass of 0.0188% versus 0.0182%. Across context lengths from 1K to 32K tokens, the same tolerance maintains low output and gradient errors relative to the reference. Across a broader controlled associative-recall comparison, MALA closely tracks FullAttn as context grows, reaching 89.67% accuracy at 8K compared with 89.97% for FullAttn. In an attention-operator benchmark at 128K tokens with tensor parallelism, MALA reduces forward and backward latency during training by 2.2x and 3.0x and decoding latency during inference by 1.6x relative to FullAttn. Across scaling-law training from 0.6B to 14B parameters, MALA closely tracks FullAttn in perplexity while reducing total training FLOPs. The resulting 14B models and 32B models from separate continued training achieve comparable knowledge, reasoning, and long-context retrieval scores to FullAttn. These results indicate that allocating post-score computation according to normalized attention contributions can retain the evaluated capabilities of FullAttn while reducing attention computation.

Healthcare administrative staff transfer structured information from electronic health records, referrals, claims systems, provider rosters, and work queues into dynamic forms. We developed and evaluated CLAIRE (Clinical Language and Agentic Intelligence for Reasoning and Entry), a hybrid workflow that separates field-state discovery, source-to-field mapping, deterministic validation, bounded correction, escalation, and audit tracing. We tested five synthetic healthcare administrative schemas, 1,000 source records, four interface variants, two data-quality suites, and six comparators, yielding 24,000 benchmark episodes. A separate strict-output audit evaluated direct mappings from Qwen2.5-1.5B and Qwen2.5-7B, and a trace-derived operational simulation covered 6,000 episodes. Under the evaluated synthetic benchmark conditions, full CLAIRE achieved 1.000 episode success, field accuracy, required-field completion, and dependency completion in both suites; removing validation reduced stress-suite success to 0.500. In the simulation, 100.0% of clean and validation-stress episodes reached a staff-reviewable draft, compared with 68.6% of escalation challenge episodes, unsupported cases were blocked. Scenario-based savings were 149.7-165.5 seconds per case, not observed staff times. The findings support schema-grounded, validation-first healthcare administrative automation in which language-model components assist mapping but do not authorize unsupported or consequential actions.

Clinical Large Language Models (LLMs) achieve strong medical-exam accuracy; however, a correct answer does not guarantee that the explanation names the concepts that actually drove the decision. We introduce three lightweight, directly interpretable metrics for this faithfulness gap: the Explanation Stability Index (ESI), which measures reasoning consistency across repeated queries; the Causal Faithfulness Score (CFS), which tests whether cited concepts drive predictions via concept ablation; and the Perturbation Stability Score (PSS), which measures robustness to semantic-preserving paraphrases. By evaluating six LLMs on 150 MedQA-USMLE questions (900 model-question observations), we found that only 23.3% of the cited clinical concepts were causally necessary. Correct answers had lower CFS than incorrect answers (0.212 vs. 0.398), answer consistency negatively predicted CFS (Spearman r = -0.466), and model pairs could agree on answers while sharing only 8.8% of cited reasoning concepts. These results show that accuracy, consistency, and consensus are incomplete safety signals for clinical decision-making support. The evidence is behavioral rather than mechanistic: concept ablation tests counterfactual sensitivity of outputs, not internal circuits.

Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches where recurrent state access becomes a major bottleneck. ReplaySSM amortizes state updates by buffering keys and values, but each new query still requires a full-state read even though the state remains unchanged between state updates. We observe that low-rank state-weighted query approximation accurately preserves state-read outputs. Although future queries are unknown, the basis vectors used to approximate them can be fixed offline. Based on this observation, we introduce SketchSSM, which preserves full-state updates while approximating reads. At each state update, SketchSSM reads the full state once to precompute outputs for these basis vectors, storing them in a compact sketch. Each subsequent decode step combines the sketch vectors with query-dependent coefficients to reconstruct the output without a full-state read. Across four Mamba-2-, GDN-, and KDA-based models, SketchSSM reduces state-access traffic by approximately 10x while largely preserving average accuracy across four decode benchmarks and recall on four RULER retrieval tasks. On one NVIDIA B300, linear-attention kernel speedups over the standard vLLM baseline reach 7.78x, 5.22x, and 5.20x for Mamba-2, GDN, and KDA, respectively, with up to 2.64x higher decode throughput on Nemotron 3 Super.

On-policy self-distillation trains reasoning models on their own trajectories using dense token distribution guidance from a privileged teacher with access to a verified solution. This supervision transfers what the teacher predicts without directly transferring where it attends within the preceding context. We introduce On-Policy Attention Self-Distillation (OPASD), which complements token-level supervision with solution-conditioned attention distillation. Because the privileged teacher can attend to verified solution tokens unavailable to the student, OPASD projects teacher attention onto student-visible positions and renormalizes the resulting distribution before alignment. Across three model sizes and four competition-level mathematics benchmarks, OPASD consistently outperforms token-only OPSD, improving average accuracy by 4.98 to 8.40 percentage points. OPASD also avoids the response-length inflation and performance degradation observed with token-only distillation, reducing generated rollout tokens by 73.9% and estimated model compute by 72.6% while training 1.53x faster. These results show that solution-conditioned attention provides a complementary supervision signal that makes on-policy self-distillation more accurate, stable, and compute-efficient.

Subjective judgments are central to many high-traffic applications, but subjective intensity is difficult to quantify and perceptions vary substantially across individuals. To address these challenges, we propose a pairwise comparative framework for multi-dimensional scale construction. By comparing case-person pairs along case and profile dimensions, the framework constructs relative scales that capture both fine-grained intensity and individual variation. To support practical high-traffic deployment, we optimize both offline scale construction and online inference. For scale construction, we combine sparse Elo comparisons with multi-judge voting, cutting the comparison cost from $O(N^2)$ to $O(NK)$ for $N$ objects and a budget of $K$ opponents per object, while limiting reliance on any single judge. For inference, we propose SubJudge, a System One model for personalized scoring with Batchwise Preference Optimization (BPO). Using Bradley-Terry comparisons, BPO trains the model to learn relative orderings, and SubJudge reads a continuous score from digit-token probabilities at the first response position, requiring only one forward pass per criterion and reducing the inference complexity to $O(1)$. Experiments on PluriHarms and iNews show that our 9B models match or surpass the evaluated frontier LLMs on multiple metrics. On the H100 GPU, SubJudge achieves an approximately $1.29\times$ to $261\times$ speedup in mean inference latency over Qwen3.5-9B with different thinking budgets. The code is available at https://github.com/Longchentong/SubJudge.

Diffusion language models (DLMs) can accelerate generation by predicting multiple tokens in parallel, but there is a mismatch between how these tokens are predicted and how they ultimately contribute to generation. Parallel predictions can hardly condition on the tokens selected earlier within the same block, even though their validity depends on this realized prefix. Under the popular proposal-verification decoding, this mismatch makes errors highly asymmetric: an early rejection prevents all subsequent proposals from contributing decoding progress. We introduce BRISK-DLM, a framework that addresses both mismatches by optimizing proposal learning and selection for verified progress. BRISK-DLM trains on self-generated sequences, using risk-reward weighting to dynamically prioritize positions by their impact on verified progress and decoding cost. During inference, a lightweight prefix-conditioned corrector reranks existing candidates using previously selected tokens and preferences distilled from the model's own verifier. The corrector reuses the backbone's parallel representations and requires no additional backbone evaluation, while fused execution keeps its overhead small. BRISK-DLM improves end-to-end throughput by up to 37.4% while preserving task quality, establishing a new quality-throughput frontier for DLM generation.

Rapid progress in Large Language Models (LLMs) is saturating static benchmarks faster than they can be designed. While existing automated evolution frameworks attempt to generate harder questions by perturbing individual tasks, they remain constrained by rigid, hard-coded generation rules. Moving beyond the evolution of isolated tasks, we propose to optimize the benchmark generation workflow itself end to end with MetaBench-Harness, a dual-loop search framework. Specifically, the inner loop utilizes a benchmark harness to generate a new benchmark in each round, while the outer meta-harness orchestration layer iteratively refines and searches over harness implementations based on historical evolution trajectories. By applying MetaBench-Harness to the competitive programming CodeContests and Olympiad mathematics AIME-2024 datasets, we demonstrate that the evolved benchmarks are challenging and discriminative for frontier models. Trajectory and quality analyses verify that MetaBench-Harness enables multi-dimensional evolution, steadily improving evolution reasonableness, benchmark competency, and evaluator robustness across successive rounds. Furthermore, case studies reveal its effective utilization of diverse difficulty levers to reframe problems and elevate required capabilities. Ultimately, this work provides a solution to the pressing challenge of benchmark saturation.

We release MoGround, a vision-language dataset spanning four visual domains in which the answer to every question is guaranteed to be available from exactly one modality. This guarantee enables us to measure modality distraction, the failure in which a model answers a question correctly from one modality alone and then flips to a wrong answer once irrelevant content from the other modality is added. Existing probes rarely establish single-modality answerability this way, making it hard to isolate distraction in the first place. Across seven open-source VLMs, we find that modality distraction is not universal but model-dependent. The weaker-grounded modality is the more distracted one (r = +0.86), and distraction scales inversely with grounding strength (r = -0.90). The single-modality guarantee also enables a mitigation method that needs to distinguish between relevant and irrelevant context. Trained on one split of MoGround alone, a weight-space robustness vector reduces distraction on all seven models by 9% to 51%, at a cost of only 0.1 average points of accuracy on standard multimodal tasks.

Large language model (LLM) serving is increasingly constrained by the GPU memory consumed by key-value (KV) caches. Existing compression, eviction, and offloading techniques alleviate this pressure, but serving runtimes typically treat only the configured target KV representation as execution-ready. Under memory pressure, this target-only contract can turn KV shortage into request stalls and preemptions. We present ElasticKV, a mixed-fidelity KV runtime built on the observation that target fidelity need not gate execution. ElasticKV introduces a compact intermediate KV state, making fidelity a runtime-managed execution property. To realize this state in a paged serving runtime, ElasticKV combines (i) a pair-structured layout that turns fidelity reduction into reusable GPU capacity, (ii) a dual-mode attention backend that directly consumes the compact state while preserving the native target-only path, and (iii) pressure-aware fidelity management that adapts KV fidelity to memory pressure. Our extensive evaluation across diverse workloads, model families and scales, and GPU platforms demonstrates the effectiveness and generality of ElasticKV. Under high concurrency, ElasticKV achieves 3.8-4.0$\times$ lower time-to-first-token (TTFT) and 9.1$\times$ lower P90 TTFT than vLLM while preserving generation quality.

On NVIDIA Blackwell B200, tensor-core throughput outpaces special-function exponential throughput by more than two orders of magnitude, exposing exponential evaluation in fused attention kernels. A pretrained Transformer, however, may not need it evaluated accurately at every element. We characterize what a pretrained model does need by approximating softmax at inference in ten frozen decoder-only models (0.5B-72B). The number of positions the softmax map assigns probability to and within-row resolution can be cut substantially, yet uniform weighting of the same positions is damaging. Where a fixed resolution budget is placed matters as much as its size, with resolution near the row maximum consistently favored. Perturbations matched on scalar distortion produce model-dependent responses of opposite sign. These findings motivate Rowmax-PoT, a coarse logarithmic weight representation anchored at each row maximum, and Rowmax-H15, its hardware specialization in FlashAttention-4. On B200, the patched FP8 attention forward is 12.4% faster at causal 8K and 25.8% faster at non-causal 8K in host-side call-latency measurements; board energy per forward falls by 8.4% at causal 16K. Measured separately on the BF16 kernel path at 2K, Rowmax-H15 increases perplexity by 0.091-0.492% across five models from three families.

Prompt-specialized multi-agent systems enable multiple agents to share a model while performing complementary roles to solve complex tasks. However, agent-specific prefixes change the KV cache generated for the same shared context, causing each agent to repeatedly prefill the growing context and construct a separate cache with high computation and memory overhead. Selective recomputation reduces this redundancy but still retains substantial model execution, while existing delta correction methods either support only recurring context relations or maintain memory-intensive online correction states for dynamically changing context. For first seen shared context, these methods also construct a reference cache outside the agent workflow, and an approximate correction at the first agent affects the outputs passed to subsequent agents. We present KVCMAS, an online KV cache correction framework that represents cross-agent cache deviations using compact low-rank states and seamlessly chains corrections along the agent workflow without an additional reference prefill. This design supports dynamically changing shared context while preserving an exact first-agent cache. Across multiple language and vision-language workloads, KVCMAS matches or improves the accuracy of prior KV cache sharing methods while achieving the lowest TTFT under highly concurrent serving. Under controlled serving traces, it provides a 2.0x TTFT speedup over inference without KV cache sharing and reduces peak GPU memory by up to 3.7x relative to a prior KV cache correction method. These results establish KVCMAS as an accurate and scalable KV cache sharing approach for prompt-specialized multi-agent serving.

Mixture-of-experts (MoE) language models often exceed the memory of a single GPU. Expert offloading keeps most experts in host memory and loads them on demand, so decoding speed depends on how many experts each token must fetch. Caching and prefetching reduce this cost only as far as the routing allows. Router-only fine-tuning can reshape the routing to reuse experts, but it keeps the experts frozen, so they cannot adapt to the tokens the new routing sends them. We propose MaskCoFT, a masked co-adaptive fine-tuning method that trains routers and experts together with the cross-entropy loss alone. During fine-tuning, a learnable binary mask restricts the Top-K routing of each layer to a subset of experts, and the experts adapt to the tokens redirected to them. At inference, the learned mask becomes a soft prior that re-ranks experts, so every expert remains selectable. We simulate a GPU cache of 4 experts per layer for Mixtral-8x7B and 12 for DeepSeek-V2-Lite. MaskCoFT cuts expert fetches per token by 23.7% and 10.1% relative to the base model. In real offloading system serving, it lowers the time per output token by up to 16.4% and 5.5%, respectively. Its average accuracy over nine benchmarks stays above the base model by 0.92 and 0.53 points.

Interpreting architecture, engineering, and construction (AEC) drawings is hard for general Multimodal Large Language Models (MLLMs) and vision-language models (VLMs). We introduce K-OPSD, a VLM post-training methodology for improving AEC drawing understanding. Building on On-Policy Self-Distillation (OPSD) with verifiable supervision, we construct a teacher from the model's own best-of-N generations, certified by a process-level verifier, and rescue failed prompts by resampling under a hint that exposes the verified answer. We then perform an on-policy model update by training on verified completions with a cross-entropy inner-loss, outperforming the bounded token-wise generalized Jensen-Shannon divergence (JSD) used by on-policy distillation. Using K-OPSD, we fine-tune Qwen3-VL models on the AECV-Bench dataset. The resulting models attain the top average judge score (0.819) and combined accuracy (0.738), achieving competitive results against open-source baseline models. The recipe transfers to the out-of-domain ArchCAD dataset, where the 8B model gains most. We present the verifier suite and the continual learning and self-improving pipeline, our results provide preliminary evidence that verifier-guided self-distillation is a promising route toward more reliable machine reading of architecture drawings.

Mixture-of-experts (MoE) models activate few experts per token, yet batched decoding can access nearly the entire expert pool, making expert-weight traffic a major bottleneck. Expert pruning reduces this traffic, but conventional approaches also prune compute-bound prefill, sacrificing model quality for little throughput benefit. We present SlimWise, a serving framework that tailors the expert pool to each inference phase. SlimWise performs prefill with the full model and decode with a pruned model that directly reuses the prefill-generated KV cache without conversion. Across two MoE backbones and three pruning criteria, this training-free KV cache handoff substantially narrows accuracy gaps relative to the full model in many settings. We also show that benchmark accuracy can conceal substantial pruning-induced changes in generation length. To address these distortions and residual accuracy loss, SlimWise introduces a low-cost distillation stage that trains the decoder to continue from full-model KV caches while updating only a small subset of parameters. Implemented in vLLM, SlimWise supports both prefill-decode (PD) disaggregation and PD-colocated serving. On Qwen3.6-35B-A3B, SlimWise improves decode throughput by up to 1.81x at 50% expert pruning with minimal accuracy loss.

Processing the same records in a different order can discard different evidence, yet endpoint accuracy alone cannot reveal what changed or whether it mattered. We introduce RoutePrism, a diagnostic protocol that builds memory twice from the same source pool in two processing orders, then traces which sources, compiled contexts, and answers differ. Because record content, timestamps, policy, and the answer model all stay fixed, any observed difference is localized to the memory construction step. A matched four-condition intervention tests whether a record displaced by reordering actually carried task-relevant evidence: restoring that single record recovers over 60 percentage points of lost accuracy, while substituting a non-supporting record of equal length does not. We evaluate the protocol on PersonaMem-32K (63 primary queries, 29 users) and 470 LongMemEval-S questions with histories spanning 38 to 62 sessions, replicating the core intervention across five answer models. Survivor selection, defined as the choice of which record a cluster retains, drives most source-level changes, while different memory policies (compaction, bounded recency, MemoChat-style summarization, A-MEM) produce distinct failure signatures at the source, context, and metadata layers.

Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies. To support learning this capability, we introduce AdaptiveSafety, a dataset of 10,939 training examples and 1,000 test examples covering policies with 1--100 rules. The dataset combines trajectories from multiple sources with policy and behavioral counterfactuals, pairing each example with an explanation and the complete set of violated rules. These counterfactuals expose changes that alter compliance, while structural augmentations provide supervision for consistency under rule reordering and identifier remapping. Building on this supervision, we propose SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts. SafePO uses structured rewards to assess prediction correctness, retains group-relative advantages at the response level, and employs a separately trained value model to modulate token weights within explanation and verdict regions. Separate normalization controls their relative contribution to training despite differences in length. Through supervised initialization followed by SafePO, we develop AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time. Our 4B model achieves binary accuracies of 89.30\% on AdaptiveSafety and 71.82\% on DynaBench. The project repository is available at https://github.com/Yunhao-Feng/AdaGuard

How much of a running workflow should a language model agent revise? ControlScope compares continuing generated code, editing the next tool call's data arguments, and replacing the unfinished workflow from the same public execution state. The nested permissions separate available repairs from the actions an agent selects. We evaluate one-time and repeated reviews across filesystem tasks, ALFWorld, and AppWorld. Across two source programs per task and three reasoning-reviewer draws on 20 filesystem tasks, FULL completes 15-16 tasks versus 13 for KEEP; across four fast draws it completes 10-13 versus 13. Fresh student-record confirmation reproduces a batch-read repair. ALFWorld fast panels yield KEEP/ARG/FULL scores of 85/86/87 on 87 tasks across 52 scenes and 134/134/127 on 134 tasks across four scenes; reasoning on the 87-task cohort also yields 85/86/87 with substantial review cost. An AppWorld V1 official-test panel of 585 task instances from 195 scenario templates shows small net differences. Frozen replays expose viable agent-written replacements interrupted by later revision in two failed file-organization runs. An offline source-trajectory midpoint comparison shows later reviews completing an insufficient repair. Five-call protection saves 19.4% of logged model output and loses one success across 20 fresh source runs. An argument-only shortcut shows that the broader sampled policy can overlook a cheaper successful edit available in both operation sets. These outcomes tie repair access to actual choices and subsequent execution.

As large language models (LLMs) are increasingly deployed across countries and regions, the ability to recognize and respond appropriately to diverse cultural contexts becomes increasingly important. However, existing research has largely focused on cultural knowledge or tasks with predefined response spaces, while open-ended culturally situated behavior remains comparatively underexplored. In this work, we introduce CRISP-RM, a culturally situated reward model that assigns rewards according to cultural appropriateness in open-ended social scenarios. During policy optimization, we further introduce Norm Grounding Supervision (NGS), providing guidance that enhances the policy's sensitivity to relevant cultural norms. To construct culturally situated data, we employ a collaborative multi-agent framework that instantiates implicit cultural norms into diverse social scenarios and further curate NormCompass as a dedicated testbed. We conduct comprehensive experiments to evaluate the effectiveness of CRISP-RM in both reward modeling and policy optimization. Best-of-\(N\) experiments show that CRISP-RM consistently outperforms strong general reward models. During GRPO policy optimization, CRISP-RM generally improves culturally situated behavior, while incorporating NGS yields further gains. Further analyses demonstrate the advantages of CRISP-RM in distinguishing culturally appropriate behavior beyond superficial fluency and politeness, while NGS provides complementary gains during policy optimization by improving norm grounding.

Spexis is a multi-GPU LLM inference framework that improves the efficiency of pipeline and tensor parallelism through speculative parallelism. Rather than using speculative decoding only to accelerate token generation, Spexis runs speculation in parallel with normal execution, introducing a new parallelism axis without increasing KV-cache memory usage. This improves memory efficiency and helps mitigate the bottlenecks of multi-GPU inference. Spexis further uses lookahead scheduling to predict speculation quality and future memory pressure, allowing it to reduce wasted speculation, KV-cache eviction, and recomputation. Built on top of vLLM, Spexis largely improves serving performance across a range of GPU configurations, achieving speedups of up to 34% over a baseline that uses the optimal combination of pipeline and tensor parallelism. Spexis's source code is publicly available at https://github.com/mlsys-seo/spexis.

Long-context LLM serving is increasingly bottlenecked by decode, where large KV caches limit batch size and underutilize GPUs. Sparse KV cache offloading expands effective capacity by storing most historical KV blocks in CPU DRAM and recalling only selected blocks on demand. However, we find that existing offloading systems shift the bottleneck to CPU-GPU recall I/O: recall volume varies widely across layers, decode steps and requests, while headwise sparse selection fragments recalls into many small PCIe transfers. This paper presents PulseInfer, an I/O-centric sparse KV cache offloading system. PulseInfer hides variable recall latency with interruptible layer-wise scheduling, adapts offloading decisions with IO-Adaptive Offloading Admission, and coalesces fragmented transfers using SoloHead sparse selection and a gather-scatter I/O engine. Implemented on SGLang, PulseInfer improves decode throughput by up to 4.7x over SGLang and 2.6x over the best existing offloading baseline, while reducing TPOT by up to 76% and preserving near-lossless accuracy.

Reasoning language models that can call tools must decide during inference whether to answer unaided or delegate. Any self-reflection mechanism for this must answer three questions: where the reflective signal comes from (verbal reports, output distributions, hidden states, a separate predictor), how it is presented to the model (numerical prediction, confidence token, prompt injection), and whether it changes the model's subsequent action. We isolate the third question. At a fixed point in otherwise identical reasoning trajectories, we insert a single first-person sentence expressing either confidence or doubt; the model then continues reasoning and chooses whether to answer directly or call a tool. Comparing these counterfactual continuations measures the causal effect of the reflective signal on delegation. We call this behavioral response Nudgeability and measure it along two dimensions: sensitivity, how strongly confidence and doubt change delegation rates, and targeting, whether delegation increases for problems the model cannot solve unaided and decreases for those it can. Across nine small-to-medium open-weight reasoning models from three families (Qwen, Gemma, and GLM) and two tasks, models are consistently sensitive: doubt increases delegation and confidence decreases it, with a median confidence-to-doubt swing of 20.6 percentage points, and 53 to 70 points for the larger provider-served models. This responsiveness is poorly targeted: a median 42% of induced flips are well-targeted, only a +2 percentage-point lift over a random-selection baseline. Confidence language is thus a strong control surface for delegation, but current models use it only weakly in accordance with their actual competence. Nudgeability offers a simple, post-training-free way to evaluate both sensitivity and targeting as endogenous self-reflection mechanisms mature.

Automated detection of LLM-generated texts (LGTs) is critical, yet dedicated detectors often struggle to generalize across domains and models. While general-purpose LLMs offer flexible zero-shot authorship classification with explanatory rationale, their detection behavior, especially regarding self-detection versus cross-detection across model generations, remains poorly understood. We systematically evaluate 15 LLMs spanning three model generations as both generators and detectors. Using a benchmark of 1,000 human-written texts and 15,000 LGTs (1,000 per model), we collected over 233,000 binary classifications alongside natural-language explanations. Our results reveal that detection efficacy is primarily driven by detector capability rather than generator provenance, although outputs from newer generators remain notably harder to detect. Crucially, statistical comparisons show no systematic advantage or disadvantage for self-detection across models. Error analysis further exposes generational bias shifts: first-generation detectors under-detect LGTs (high false-negative rates), second-generation detectors over-flag human texts (high false-positive rates), and the latest models achieve balanced trade-offs. Finally, we highlight significant inconsistencies in how different LLMs apply textual cues to justify their decisions. Code: https://github.com/hyyuan/detect-llm-generated-texts.

Self-evolving agents improve future behavior by reusing past experience, typically as global prompts, memories, or reflections. Yet these mechanisms rarely control where experience takes effect. In long tool-use workflows, the same lesson may correct one decision but distract another, making experience reuse a problem of localized control rather than memory alone. We introduce EvoCUE (Evolution through Control Updates from Evidence), a framework for learning reusable control-program updates from completed agent executions. EvoCUE represents the agent as an explicit state-machine controller, whose nodes perform model or tool calls and whose edges define where control passes next. This makes the workflow editable at precise locations, so each learned update can specify what to add, where it acts, and when it applies. From completed trajectories, EvoCUE uses residual goals and observed execution traces to propose localized instruction or skill edits. Each candidate is evaluated at the point where it would act by resuming the parent and edited controllers from the same checkpoint and comparing their final outcomes. Accepted edits are compiled with applicability rules, confirmed on held-out tasks, and inherited by later executions. We evaluate EvoCUE on long tool-use environments where learned conventions must reach the right execution step. From a minimal AppWorld controller without benchmark-specific onboarding instructions, EvoCUE learns the missing task-completion convention and substantially improves success on Test-Normal and Test-Challenge. On PAST-Bench office workflows, EvoCUE transfers organizational requirements from prior episodes to later tasks, improving task-execution quality. These results show that self-evolving agents should place experience inside the control flow, rather than only store it as text.

In complex embodied visual reasoning scenarios, an agent often has only a limited field of view, and the evidence needed to answer a question may be distributed across time, viewpoint, and interacting objects. A model may therefore give a plausible answer without ever observing the relevant object, time, or view that supports it. Current visual reasoning benchmarks largely evaluate passive observations and final answers, overlooking settings that require active reasoning and evidence acquisition. We introduce JRDB-AVR, a benchmark derived from existing real-world JRDB robotics data through a structured question-generation engine that turns this gap into an explicit evaluation: an embodied agentic system receives a visual reasoning question, requests bounded observations by timestamp and viewing angle, and is evaluated on both the final answer and the grounded visual evidence supporting it. The benchmark contains diverse questions over multiple real-world environments involving temporal search, viewpoint selection, and human-oriented compositional reasoning. We also introduce JRDB-AVR-Agent, a reference active reasoning agentic method that maintains an explicit observation-grounded graph-based world model and answers through solving. Experiments reveal a substantial gap between answer accuracy and evidence accuracy in current baselines, showing that current VLMs can produce unsupported correct answers and that active evidence-aware evaluation is necessary for embodied visual reasoning. Code and benchmark are available at https://github.com/ControlNet/JRDB-AVR.

Full-duplex speech models support streaming interaction that listens and speaks at the same time. Serving them is governed by a strict, repeating deadline: conversation advances on a one-second cadence, and every second of input must be turned into a second of speech before the next second arrives. Because stages within a session run in strict sequence, per-invocation overhead cannot be batched away. Profiling reveals that the autoregressive stages of a duplex second already fit within the period, whereas the token-to-audio synthesis tail is what causes overruns. This tail stage suffers from orchestration slack where the GPU is left waiting as thousands of tiny, regular operations are issued one by one, while also wasting substantial memory by over-provisioning state at static implementation constants. Existing remedies, such as graph recording and demand-sized allocation, fail because streaming state dynamics violate their prerequisites. The root cause is that the runtime lacks the model's native clocks: the per-region counters that govern advancement rates and retention policies. We propose DuplexCadence, which explicitly declares native clocks to the runtime and derives two mutually enabling rules: demand-sized state allocation at a stable address, and exact-shape graph replay without padding. The former eliminates idle memory and stabilizes tensor pointers, while the latter removes orchestration slack without padding overhead. Evaluated on four released models across three decoder architectures with bit-for-bit identical output, DuplexCadence reaches $2.85\times$ the stock runtime's speed at $38.8\%$ lower peak memory. On the live duplex path, mean SPEAK time falls from $14\%$ over the one-second cadence to $2\%$ under it, enabling models to reliably keep up with interactive speech while markedly expanding multi-

Multi-document retrieval-augmented generation (RAG) requires a language model to process multiple retrieved text chunks before answering a question. Precomputing each chunk's KV cache independently and concatenating the caches when the chunks are retrieved can accelerate this step. However, the assembled cache lacks cross-chunk attention information, reducing answer quality. Selective recomputation methods recover the missing cross-chunk context by rerunning the target LLM on selected tokens, incurring substantial online computation. We introduce CacheRepair, a lightweight network that learns the difference between independently computed KV caches and those produced by processing the chunks together. The network combines compressed KV features with token embeddings and uses attention that is bidirectional within each chunk and flows from earlier to later chunks. Each repair block receives the compressed cache features, and the predicted residual is added to every document token's cache. Each repair network is trained for a specific frozen target LLM on a generic retrieval corpus and reused across downstream datasets. Our analysis shows that repair reduces KV errors both near chunk boundaries and throughout chunk interiors. Evaluation across three target LLMs and four downstream datasets places CacheRepair on the measured answer-quality-latency Pareto frontier in eleven of twelve model-dataset combinations. Reported time to first token (TTFT) includes online cache transfer and repair. Across all twelve combinations, the largest repairers achieve 1.69-4.61$\times$ speedups in median TTFT over full prefill and improve mean F1 by 2.1-26.1 percentage points over direct cache reuse.

Large language models (LLMs) typically generate text autoregressively (AR), predicting one token at a time. Block diffusion language models (dLLMs) instead generate blocks sequentially while denoising multiple tokens in parallel within each block, offering a promising way to accelerate generation. Rather than training such models from scratch, recent work adapts strong pretrained AR models into block dLLMs through distillation. On-policy distillation (OPD) has been widely used for LLM training because it supervises the student on states generated by its current policy, rather than only on fixed offline trajectories. By training on the states the student actually visits, it reduces the mismatch between training and generation and can provide more relevant supervision as the student evolves. Recent work has extended this idea to AR-to-block-diffusion conversion. However, this setting introduces a fundamental mismatch in supervision: the block-diffusion student and the causal AR teacher condition on different information at the same training state. The student predicts from the entire partially denoised block, including visible future context, whereas the standard AR teacher target is defined only from the causal prefix. As a result, the teacher distribution used for distillation is not fully aligned with the information available to the student. We therefore introduce d-OPD, a future-aware on-policy distillation method that corrects the AR teacher distribution to better align with the student-visible state by incorporating visible future information within each block, providing supervision that better matches the information used by the student. Across Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to $4.0$ points over OPDLM and reduces training time by $1.35$-$1.58\times$. The code is available at https://github.com/mit-han-lab/d-OPD.

Large language models are trained with backpropagation, whose global gradient coordinates all layers but forces each to hold its activations and wait for the gradient to pass back through every deeper layer. Conventional local learning removes this update locking by training each module to predict the target through its own readout, but has not scaled to billion-parameter pretraining. We identify these private readouts as a key weakness, since they leave each module without information from deeper modules. We propose Shared-Output LOcal learning (SOLO), which replaces them with a shared, read-only copy of the final module's readout, the only one trained on the output of the whole network. Taken from the previous step, the copy transmits information from the final module without passing gradients between modules or reintroducing update locking. SOLO approaches backpropagation on Transformers of 340M to 2B parameters pretrained on 15B tokens, staying within one point in average zero-shot accuracy with a perplexity gap that narrows with scale. Readout ablations attribute SOLO's improvement over private readouts to sharing. Without update locking, each of p pipeline stages holds activations for O(1) micro-batches instead of O(p). The freed memory permits larger micro-batches, which reach up to 1.44x the best measured throughput of pipeline backpropagation on the same partition. To our knowledge, SOLO is the first local learning method to show such memory and throughput gains in billion-parameter language-model pretraining. Local learning thus becomes a practical alternative to backpropagation for large-scale pretraining.

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher--student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

One deployed language model must often serve many compute budgets, yet serving each budget still means a separate training or compression run per point. We train a Telescopic Language Model (TLM) to be that continuum: a nested-capacity Transformer supervised by stochastic prefix supervision with a full anchor. At every step, one randomly truncated prefix of the capacity axis is trained against the full next-token target, alongside one full-capacity pass, so the trained artifact is a valid language model at every depth. Two forward-backward passes per step, no architectural change, nothing extra at inference. Fixed-exit suites such as Matryoshka Language Model Suites (MLMS) occupy one point in this design space, and the point has a cost: supervising only a few fixed exits leaves the nested model at chance level everywhere else (perplexity 10^2-10^5 in our baselines). On a 200M proxy suite (20B FineWeb-Edu tokens, identical data stream for all methods), a single TLM run is a valid language model at every one of its twenty layer prefixes, in perplexity and on perplexity-sensitive downstream tasks, reducing the area under the quality-budget curve by 43-44% relative to the fixed-exit suites while matching them at full capacity, at ~12% lower GPU cost per run. The prefix sampling density is a dial: concentrating it on a few depths recovers fixed-exit quality there at the price of the continuum, so the operating points become a training-time choice rather than an architectural one. These results indicate that the training objective, not the nesting itself, is what makes a model elastic.