2026-09-24

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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