2026-10-09

Reward hacking, where a model exploits an evaluation signal without completing the intended task, threatens the reliability of deployed language model systems. Existing datasets use different labels, response formats, and metadata conventions, making detector results difficult to compare. We present RH-Detect, a benchmark that combines reward-hacking-relevant subsets from eleven public datasets, comprising 92,761 rows and six behavior categories, into a common schema. On 5,021 open-ended evaluation units, each comprising a task prompt and a free-form model continuation, including multi-turn tool-use trajectories, we evaluate six off-the-shelf language models from five families as reward hacking detectors without additional training. The best model achieves a pooled AUROC of 0.962, with accuracy above 93%. For the four strongest models, however, accuracy on the two multi-turn tool-use datasets, MALT and TRACE, is 10.7-15.9 percentage points lower than on the other sources at a common decision threshold, highlighting a key gap for deployment-time monitoring. We find that different input formats have different effects across models. Removing thinking raises Qwen3.5-4B AUROC from 0.779 to 0.849, but lowers Qwen Flash from 0.977 to 0.950. We also evaluate the benchmark as a training dataset for detectors. Holding out each source in turn, single-token SFT improves average AUROC on five of six held-out sources. A GRPO follow-up on that failure case yields a slight improvement in detection performance. Our results show that a single pooled score can conceal variation across data sources, detector inputs, and training procedures.

Large language models are increasingly used in legal research and drafting, but they can still produce claims that sound convincing without being supported by the cited source. We introduce PARCEL, a benchmark for checking whether a legal claim is supported by the underlying authority. Using recent New York State Court of Appeals decisions, we build a dataset of 3,396 parenthetical-style claims labeled as Supported, Refuted, or Not Found. We cast this task as a three-way natural language inference problem and evaluate several state-of-the-art LLMs in a zero-shot setting. Although the strongest models reach up to 0.97 accuracy, the results also show an important weakness: models still incorrectly mark unsupported claims as supported, even when the full opinion text is provided. Across models, missing support is harder to detect than direct contradiction, and fabricated but plausible citations cause the largest drop in performance. Overall, PARCEL provides a practical benchmark for testing claim-level groundedness in legal RAG systems.

Conversational task disambiguation over tabular data uses dialogue to resolve missing information about a user's intended task before producing a solution over tables or databases. Existing evaluation and training lack a leakage-aware foundation. Task success mixes the agent's disambiguation and solution-generation capabilities and can also reflect oracle leakage, that is, information that a user simulator reveals beyond what a real user would. Existing datasets also lack a shared representation of ambiguities and access boundaries. We introduce the notion of an ambiguous verifiable task, which formalizes ambiguities and resolutions, decomposing the agent into an asking policy and a solution policy, and the environment into an oracle and verifier. This framework provides baselines and metrics for evaluating task disambiguation separately from solution generation, formal definitions of oracle leakage, judge-free leakage diagnostics, and a training objective for the asking policy. We instantiate the framework in text-to-SQL with AmbiTab, a benchmark suite that unifies six ambiguous datasets under a common representation specifying what the agent, oracle, and verifier may access. We evaluate clarification strategies and oracle leakage, and train an asking policy with reinforcement learning. The trained asker improves our disambiguation metrics on all six datasets and task success on five, and our leakage diagnostics measure how training affects oracle leakage.

Reinforcement learning environments are now a primary lever for improving large language model (LLM) capabilities in post-training, yet most agentic benchmarks remain static: the world moves only when the agent acts, the reward is a terminal verdict, and the pass bar is set arbitrarily. We introduce StoreBench, a live-commerce environment in which an agent runs a mid-size online apparel store on a production-grade commerce backend, testing long-horizon planning and economic judgment under uncertainty. Customers order around the clock, suppliers reprice and fail, and market shocks arrive with partial or no warning. The agent acts through the same 29 merchant tools a human operator would use, under a windowed operation budget that makes simulated time a function of actions taken, so model latency cannot influence simulated time. Pass thresholds are calibrated against scripted anchor policies, the reward is hardened against a catalogue of reward hacks, and every episode replays identically given a sequence of actions. We evaluate seven frontier LLMs on 11 scenarios of 30 to 45 days and a full simulated year, over three world seeds at matched reasoning effort. No model matches the scripted smart-triage policy on average: the best, DeepSeek-V4-Pro, passes 49% of task-seed cells against the heuristic's 97%. Human experts working through the same tools and budgets outscore every model (mean composite 0.708 vs. 0.700). Over a full simulated year under the Claude Code harness, most models show dramatic performance improvement. In a GRPO post-training run, Qwen3.5-27B trained on only five disjoint tasks raises its mean composite on the held-out evaluation tasks from 0.136 to 0.373. We release five example training-split tasks, ten sample trajectories, and the scoring and verification tooling; the full environment and evaluation suite are withheld to keep the benchmark uncontaminated.

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

Computer-use agents are capable of completing complex tasks, increasing the use of automatic judges to determine success, either for training or for evaluation without human involvement. Despite their flexibility, their reliability on long tasks spanning multiple applications remains unclear. A trajectory, composed of long sequences of screenshots and actions, may appear complete, but in reality violates constraints from the instruction or introduces an unwanted side effect. To identify these errors, a judge needs to examine the trajectory with respect to the user's instruction. To this end, we introduce AgentHorizon, a benchmark of 1,373 computer-use tasks (instruction-trajectory pairs) drawn from 166 hours of human-recorded trajectories spanning three operating systems. By recording trajectories for closely related instructions, we can construct negative tasks by swapping the instructions. This paired design evaluates judges on their ability to distinguish a successful trajectory from one that completed a similar (but incompatible) request. We release the benchmark under three splits: a frontier split, AgentHorizon (AH), a simplified split, AgentHorizon-Simple (AH-S), and a development split, AgentHorizon-Development (AH-D). We further evaluate eleven judges by (1) directly passing the full trajectory (with up to 300 screenshots and actions), and (2) using them as coding agents across five agent harnesses. We find that our best agentic judge, GPT-5.5, achieves 80.9% balanced accuracy on the AH subset. We find that tool-use improves certain models but results in worse performance for open-weight models, and that judges differ drastically in their ability to accept a valid trajectory and reject failed ones. Our findings highlight the need for judges that are capable of locating and verifying often hidden evidence that a task was properly completed inside long interaction histories.

A language model (LM) can usually answer the same question in more than one way, but reinforcement learning with verifiable rewards (RLVR) is indifferent to which correct answer a model produces. A solution will earn the same reward whether it is the thousandth copy of a familiar answer or one the model has never produced before. Yet, there is potential value in having the model retain multiple correct solutions as it is trained. For instance, multiple modes may give users a choice and provide problem-solving strategies that improve overall model performance. Here, we introduce ModeBench, a benchmark of multi-solution tasks in which the verifier returns both correctness and mode discovered. We then use ModeBench to measure how solution diversity changes under RLVR post-training. We find that RLVR post-training concentrates probability onto fewer correct modes even as accuracy holds or improves, and moreover, that frontier models are already highly concentrated. We then introduce our solution, Re:Max, which stores one verified example per discovered mode in a replay buffer and trains on those stored modes uniformly. A solution found once is, therefore, practiced as often as one found repeatedly. Across three model scales, two RL objectives, and harder task constructions, replay improves both how often a policy succeeds and how many different ways it can succeed.

The rapid growth of scientific publishing has strained peer review, particularly in machine learning, raising concerns about declining review quality and increasing reviewer workload. Large language models (LLMs) have been proposed as automated review assistants, yet their evaluation has focused largely on imitating human-written reviews rather than supporting the core functions of peer review. Here, we introduce a verification-centric perspective on LLM-assisted peer review, emphasizing error detection as a critical and resource-intensive task. We present a scalable benchmark that evaluates review systems' ability to identify logical contradictions, constructed through synthetic insertion of errors into conference papers, yielding unambiguous evaluation targets and enabling systematic comparison. We further propose a Multi-Layered Review (MLR) framework that prioritizes detailed manuscript comprehension before review generation, aligning more closely with human reviewing practices while improving token efficiency. Across evaluations, our approach demonstrates strong alignment with human review scores, achieves high error detection performance, and provides complementary perspectives on reviewer focus. These improvements can be attributed to both the choice of the underlying LLM and the design of our system. At the same time, we corroborate persistent vulnerabilities to adversarial manipulation, underscoring the need for robustness in automated review systems. Our findings highlight the importance of rigorous, error-focused evaluation to guide responsible deployment of LLM-based tools in peer review and other critical scientific workflows.

The next frontier for artificial general intelligence is tackling unresolved scientific problems, calling for benchmarks that assess progress beyond established knowledge. We introduce OpenProblemBench, a benchmark of 82 unresolved problems drawn from the mathematics and theoretical physics literature. Each problem supplies the research context, assumptions, and prior progress needed to investigate the question. We select problems whose proposed solutions admit comparatively clear checks of their decisive mathematical or computational claims. Four evaluator models independently assess the correctness, completeness, and degree of progress of each submission without reference solutions. Across seven evaluated configurations, GPT-6-Astra achieves the highest mean judged solve rate of 14.0%, compared with 5.5-6.7% for the evaluated full-size open models and 2.4-3.7% for Flash models. Case comparisons connect stronger outcomes to changes in problem representation, general arguments that extend beyond finite evidence, and proofs of the steps needed to complete a solution. By grounding evaluation in questions arising from the research literature, OpenProblemBench provides a setting for investigating the capabilities and limitations of AI as a contributor to foundational theoretical science.

Most human-agent interaction today remains text-based. Natural language can impose cognitive overload, ambiguity, information chaos, and slow input for complex tasks; ephemeral generative UIs can present structured information and guide users toward task completion. We propose GenUI-Harness, a multi-agent harness pairing a Tool Agent for information retrieval and task execution with a GUI Coder Agent that identifies ambiguities and generates front-end code for structured interfaces. Training the coder with reinforcement learning is challenging: verifiable rewards for interactive UI generation require costly execution, while LLM-as-a-Judge rewards are prone to reward hacking. We address the first challenge with Dynamic UX, a lightweight package for dynamic interaction and reward collection in a single sandbox, and the second with Reward Auditor, a meta-reward mechanism that monitors reward distributions and distills diagnostic patterns into a shared rubric and scoring specification. We introduce UI-TAU Bench, a benchmark for active human-agent interaction through generated UI code, built on 10 real-world domain databases constructed from public data sources and based on Tau-Bench tool-use settings, with Lite (300 tasks) and Full (1,000 tasks) splits. GenUI-Harness achieves an average Pass@3 gain of 4.48 percentage points over smolagents on Lite. Training with GenUI-Harness improves a 4B backbone from 9.33% to 58.00% Pass@3, outperforming larger frontier models such as Claude Opus 5 (46.67%). GenUI-Harness also remains robust on ambiguous and non-ambiguous queries. In a reviewer survey comparing communication channels, generated UIs reduce average dialogue rounds from 3.4 to 1.2. These results show that data-aware generative interfaces can support effective task completion and reduce dialogue rounds in evaluated database-backed workflows.

Game commentary is an open-ended generation task requiring multimodal perception, strategic reasoning, and contextual knowledge. Existing AI-Generated Game Commentary (AI-GGC) studies remain fragmented across games, modalities, and evaluation protocols, while overlap-based or holistic evaluators fail to capture the functional heterogeneity of commentary. We introduce \textsc{GameCommBench}, a unified benchmark spanning board games, sports, and esports, with commentary aligned to heterogeneous game contexts and annotated by commentary type. We further propose Type-Aware Commentary Evaluation (TACE), a structured framework for evaluating different types of commentary. We then validate TACE for reliability and human agreement, and use it to benchmark representative AI commentators. Results reveal non-uniform capability profiles, with live observation and strategic analysis emerging as major bottlenecks. Together, \textsc{GameCommBench} and TACE provide a diagnostic foundation for comparable and interpretable AI-GGC evaluation.

Advances in foundation models are driving efforts to introduce agents to assist people in the physical world. Such agents require agentic spatial intelligence: exploring unfamiliar environments, updating spatial understanding through interaction, and adapting actions based on feedback to sustain progress toward a sequence of goals. Existing benchmarks cover only a limited range of spatial layouts, scales, and traversal requirements. We introduce Mine Odyssey, a benchmark for evaluating agentic spatial intelligence using Minecraft reconstructions of real-world locations. It comprises 180 tasks covering 30 such locations across 20 countries and regions on five continents, including 20 outdoor and 10 indoor settings. These settings span diverse spatial scales, layouts, terrains, and connectivity patterns, from Midtown Manhattan and rural Entrup to Santa Luc\'ia Hill and Buckingham Palace. We select meaningful waypoints, such as landmarks, buildings, and rooms, and manually verify their accessibility. Each task provides a natural-language instruction specifying which waypoints to visit and in what order. Completing these tasks requires agents to find accessible routes and entrances, open doors, and move between levels using stairs and ladders, while monitoring their progress and recovering from navigation errors. Across eight evaluated state-of-the-art models, GPT-6 Astra achieves the highest success rate of 85.6%. However, the second-best model, Claude Opus 5.5, completes 73.9% of tasks, while the strongest evaluated open-weight model, DeepSeek-V4.1-Flash, reaches 23.9%, highlighting substantial room for improvement in the agentic spatial intelligence of current models. Comprehensive analyses and ablation studies on Mine Odyssey reveal current models' limitations and provide insights for advancing agentic spatial intelligence.

System One models output calibrated probabilities over typed answers such as categorical choices, ordinal levels, or binary outcomes, via a non-generative interface. Software can act on these probabilities through thresholds, cost-weighted choices, and escalation rules. Consequently, if these probabilities are miscalibrated or wording-sensitive, the software ma take unintended actions leaving human operators with no textual rationale to inspect. Current evaluations largely report accuracy and calibration on public classification datasets without a clear reference. We present TypedBench, a benchmark for typed decision models like Jev built from seven policy-labelled generators and nine evaluation suites. We report accuracy as median and range across paraphrases, and calibration error relative to the finite-sample noise floor of a matched, perfectly calibrated predictor. We assess probability quality through selective prediction, ordinal proper scoring rules, and realised cost under asymmetric cost matrices. We evaluate a hosted model, an open encoder, and a family of open decoders spanning 0.8B-9B parameters on identical items. The hosted model follows the stated policy but is wording-sensitive and systematically underconfident; under asymmetric costs, using its probabilities can be worse than taking its top answer. Decoders route exactly and are slow as options or questions are added. The decoder is least accurate on policy questions and degrades with more options. Overall, typed decision models must be evaluated jointly on policy adherence, wording robustness, probability quality, and induced decision outcomes.

Coding assistants such as Claude Code and Codex have become a major application of LLM agents, yet existing benchmarks remain far from real-world use, particularly in task horizon and interaction length. Code assistants require completing long chains of development work in continuously evolving repositories, while repeatedly clarifying requirements and adapting implementations through multi-turn interaction. To address these gaps, we introduce SWE-Journey, a benchmark for more realistic evaluation of coding assistants. To address the task-horizon gap, we propose a weak-to-strong synthesis pipeline that automatically constructs long-horizon coding tasks. To address the interaction gap, we mine four representative user personas from real interaction data and build a user-simulation agent to reproduce realistic code-assistance interactions. On average, models pass over 75% of tests for requested functionality with software architects, but fewer than 25% with non-coders. These results show that current coding assistants still fall short of enabling reliable coding for non-coders. We further analyze the reasons for this gap and identify asking right, finding right, and fixing right as key capabilities during interaction.

LLM-based agents now interact with their environments continuously, shaped by such organizational channels as user instructions, peer messages, and long-term memory. Existing safety research has examined these influences, but largely as separate agent components. How such factors jointly shape an agent's safety behavior from a unified perspective remains unmeasured. To bridge this gap, we advocate organizational behavior (OB) as a framework for studying the safety of advanced agents, reorganizing the objects of study, theoretical foundations, and experimental design around the relational structure in which agents operate. We present the first systematic formalization of counterproductive work behavior (CWB), a canonical safety-relevant subfield of OB, as Agentic Counterproductive Behavior (ACB). ACB specifies three organizational antecedents (vertical supervisor relations, horizontal peer norms, and internal cognitive structures) and maps them onto three counterproductive outcome dimensions (unauthorized disclosure, destructive operations, and production deviation). To operationalize ACB, we introduce Workerville, a controlled benchmark that manipulates organizational conditions over shared tasks, applying 16 organizational configurations to 210 tasks to yield 3,360 challenges, evaluated by human-validated agentic judges. Benchmarking 6 frontier LLMs, we find that (I) negative organizational antecedents exhibit non-monotonic amplification when combined, with the unauthorized-disclosure rate rising from 16.5% under no negative antecedent to 60.1% under two and falling back to 50.3% under three; (II) agents reproduce typical behavioral patterns predicted by human CWB research; (III) these results establish OB as a systematic framework for agent safety research, pointing toward a new research agenda.

Existing evaluation methods for healthcare LLMs primarily assess factual correctness,safety, and fluency, while providing limited insight into whether generated interactions reflect culturally situated healthcare reasoning. This limitation is particularly important in maternal health, where care decisions are shaped by social and relational norms. We introduce MH-INDIC, a culturally grounded evaluation framework for maternal-health interactions in urban and semi-urban North Indian contexts that operationalises cultural behaviour through ten dimensions of maternal-health reasoning. Using a 26-item survey administered to 102 pregnant and postpartum women from urban and semi-urban North India, we evaluate ten LLMs. We distinguish population level cultural alignment from profile-level behavioural variation. Although several models approximate the human population-level distribution, all evaluated systems exhibit substantially lower variation across demographic and household profiles than the human cohort, revealing a gap between aggregate alignment and profile-conditioned sensitivity. As a downstream application of MH-INDIC, we use the strongest-aligned proprietary and open-source models to generate culturally conditioned maternal-health dialogues under zero-shot, self-conditioned, and human-grounded prompting. Human-grounded conditioning produces stronger profile alignment and dialogue quality ratings, suggesting that measured cultural profiles can improve the cultural grounding of generated interactions

Evaluating user experience (UX) with automated computational methods has gained increasing attention, supported by empirical evidence from UXBench. However, binary preference prediction provides limited insight, while relying on a single user-agnostic reward model overlooks the inherent heterogeneity of users, whose expectations can differ substantially. In this paper, we present UXBench Pro, comprising 1{,}000 test instances derived from real user interactions across 12 task scenarios and 82 domains. Each instance is paired with a FACTORS user profile that characterizes the user through seven interpretable behavioral facets, differentiating user groups. To provide richer evaluation insights, we introduce a dual-perspective paradigm that combines a personalized User Reward Model (URM) for third-person judgment with Sim4Eval, a user simulator that enables multi-turn interactions and provides first-person evaluation across four cognitive state dimensions. To assess the reliability of these based evaluators, we further introduce two meta-benchmarks, URMBench and USimBench, that evaluate how faithfully they reproduce real human preferences and behaviors. Extensive experiments reveal seven key findings that highlight the importance of user modeling and multi-perspective evaluation, offering a fresh perspective on user-centric benchmarking and motivating personalized model optimization.

Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.

Reliable multimodal interaction depends on selective use of conversational history: an earlier question may remain relevant while its previous answer is outdated, whereas a current request may depend on historical evidence despite conflicting new observations. Existing multi-turn evaluations rarely separate these history-use demands from underlying question difficulty. To address this gap, we introduce ReTurn, a benchmark of 7,000 base tasks spanning visual and audio evidence for evaluating selective history use. For task-carrying history, Reconfirm/Reground require applying a historical question to current media while varying historical agreement; for evidence-carrying history, Retrieve/Rebind require answering a current question using historical media while varying current-media competition. Each pair preserves the target question, media, and answer. Tasks support open-ended and multiple-choice evaluation, with matched single-turn counterparts serving as answerability references. Across 13 omni-modal, vision-language, and audio-language models, median model-level open-ended accuracy falls from 93.7% with direct input to 72.3% in conversation. Behavioral probes show that high question recall can coexist with weaker task application, while competing media can redirect answers away from historical targets. Supervised adaptation yields only partial gains. ReTurn provides a controlled framework for assessing whether multimodal models select and use the historical information required by each request.

Large language models (LLMs) are increasingly deployed in sociotechnical systems where social attribution, the reasoning process attributing external events to the causes and reasons of agents' social behaviors, plays a critical role. These processes involve judgments of social cause, responsibility, and blame/credit to agents. Although attributional models are well-studied in social psychology and cognition through Attribution Theory, social attribution remains underexplored in AI, particularly LLM social reasoning. This paper provides the first systematic exploration of LLM social attribution. Our work focuses on responsibility and blame attributions, examining current LLMs' judgments and their underlying internal mechanisms. Guided by attribution theory, we construct a social attribution benchmark consisting of a Vignette subset based on classic scenarios from attribution theory research and a Reality subset based on real-world social narratives, yielding 7,639 responsibility/blame judgment questions. On this basis, we evaluate 32 representative LLMs and 5 basic non-LLM baselines. To further explore the internal mechanisms underlying the LLM judgment process, we develop a probing-based methodology to investigate the latent-space representations of 5 key attribution dimensions and the consistency of their influences on LLM judgments compared to those in human social attribution. Our research findings reveal that current LLMs exhibit measurable but incomplete agreement with human responsibility and blame judgments, and meanwhile, this agreement is positively correlated with model size. Some attribution dimensions are systematically decodable from specific positions in LLM hidden states, and their influences on the final judgment are consistent with those indicated by human Attribution Theory. The dataset and associated code are available at https://github.com/Yuzhaoxin946/SAB-Bench.

Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited. Recent large language model (LLM) agents offer a new way to approach BBO by combining task semantics, computation, optimization tools, and feedback-driven decision making, showing great potential due to the integration with mathematically rigorous tools. However, existing agentic BBO studies use different task domains and system configurations, making their results difficult to compare and the effects of individual design choices hard to isolate. We therefore introduce AgenticBBO-Bench, a cross-domain benchmark for agentic BBO spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design under a unified finite-budget evaluation protocol. In our experiments, agentic BBO achieves higher family-averaged scores than direct LLM-based methods in all five domains and outperforms the best numerical optimizers in four. We further study three factors shaping agent performance: optimization tools, task information and prior knowledge, and the role of the LLM during search. Our results show that additional numerical tools do not consistently improve performance, task semantics are broadly useful while more specific priors are less reliable, and numerical optimizers can effectively absorb gains from search trajectories established by the agent. Finally, we introduce a five-task frontier challenge within AgenticBBO-Bench and evaluate seven LLMs under the Codex agent harness, where GPT-6 Astra and DeepSeek-V4.1-Flash lie on the Pareto frontier of performance and cost among the evaluated models. Our code is available at https://github.com/lamda-bbo/agentic-bbo.

As claims about recursive self-improvement (RSI) and artificial general intelligence (AGI) proliferate, we ask a simple question: do frontier AI models have a sense of data, i.e., can they reliably select the right data for training? We introduce DataSense-Bench to study this capability through the fundamental problem of data selection and performance forecasting in machine learning. We ask AI agents to select and rank candidate training subsets that can be used to fine-tune a small LLM model. Agents are allowed to inspect the data, write and execute analysis code, and run model forward passes, but can not train the model or access the actual evaluation tasks. We then fine-tune the base model on each selected subset and evaluate its post-training performance under a standardized protocol. We instantiate the benchmark in terminal problem solving and tool use, selecting trajectories from OpenThoughts-Agent and EnvScaler and evaluating on TBLite and BFCL, respectively. We then evaluate the agents along two complementary dimensions: the post-training performance of the top-ranked subset, reflecting the ability to identify high-value training data, and ranking accuracy, reflecting the ability to predict the relative performance of the selected subsets. In our experiments, selection gains over random selection are limited; agents do not reliably rank their selected groups, and ranking ability does not hold consistently across tasks: Astra identifies the best group in all three tool-use runs but in only one of three terminal runs. Analysis of execution traces on both tasks shows that agents often use similar data signals while interpreting their training value differently.

Pre-trained transformer models are increasingly being used to study scientific and technological progress. Encoders tuned to paper or patent text outperform general-purpose models on downstream classification, regression, and proximity tasks within science and technology. However, the applicability of these models for studying time-dependent or archival properties of science, technology, and their interface is limited due to lookahead and domain biases inherent to these pre-trained models. These limitations arise from training on corpora with unconstrained chronological and text source distributions. We introduce SciTBERT: a family of chronologically consistent BERT-derived language models trained on text from scientific papers, patents, and high-quality educational web text with training data cutoff dates spanning each year between 2013 and 2025. We also post-train these models in a chronologically-consistent manner using paper and patent citations, creating SciTBERT-CI model family. We find that these models generally outperform predecessor domain-specific encoder models even when training data is limited by early year restrictions in the corpus. To further investigate the extent to which this class of models can learn representations that bridge science and technology, we introduce the PatRepEval benchmark, a suite of patent-related text embedding tasks at the science-technology interface. Performance in a variety of classification, regression, and retrieval tasks spanning papers and patents highlights the importance of aligning encoder model representations with the domain distributions of their downstream tasks, and chronologically consistent encoders can match or exceed models trained without temporal constraints.

Tool-using agents repeatedly carry observations whose useful content can be much smaller than their original payload. We study agent-controlled forgetting: the acting model selects previously observed tool results, replaces each with a short note at its original position, and retains the exact original in a recoverable archive. A Python harness exposes batch archival and explicit recovery without task-specific model training, while protecting user instructions and assistant messages from these operations. In an exploratory OpenTelemetry debugging case followed by an unrelated implementation task, the method ended with 231,951 provider-reported prompt tokens versus 912,492 under retained history, used 50% fewer cumulative input tokens, and had an estimated API cost of USD 1.28-1.44 versus approximately USD 4.38. Both arms passed the two-case primary behavioral oracle; neither fully satisfied the follow-up evaluation. The method made more requests and took 17% longer. A contrasting application-development pair produced no context or cost saving, and an earlier continuation exhibited lower manually assessed quality despite reduced context. These observations demonstrate substantial resource savings in noisy tool-use trajectories and identify workload dependence as a central consideration for reversible context management.

Corrective feedback is among the best-evidenced drivers of second-language acquisition, yet corrections delivered during lessons rarely accumulate into an actionable view of grammar mastery. Prompted frontier models can provide such a view from learner--tutor lesson transcripts, but they are costly at scale. We close this gap by fine-tuning Qwen3.5 small language models (SLMs) on filtered and rebalanced teacher-generated supervision, then deploying an efficient 0.8B model in an end-to-end grammar mastery tracker for all English learners on our platform. Internalizing the annotation contract into adapter weights enables pairing the 0.8B model with a compact matched prompt rather than verbose instructions. On two human-curated benchmarks, both the deployed 0.8B model and a 4B reference comparator outperform prompted GPT-5.4 and GPT-5.6 Sol in precision and recall under nested matching criteria of increasing strictness: concept, evidence span, and correctness. The deployed 0.8B SLM reduces serving cost by approximately 16$\times$. A feature-level online experiment shows significant gains in learner engagement ($+15.8\%$) and key business metrics, including scheduled hours ($+2.1\%$) and GMV from new lessons ($+13.2\%$).

Large language models (LLMs) are entering wildfire management, where overstated evaluations can cost property and lives. How do they perform on wildfire tasks, with and without grounding? Bare means a model receives the task input alone. Grounded means it also receives one task-specific addition: for smoke detection, a smoke-free reference frame from the same camera. AI4Fire runs six core models bare and grounded on five wildfire tasks, zero-shot; a sweep adds 29 more. Our literature search on fire tasks found 138 works; none combines this roster, task coverage, and paired bare and grounded runs. We report three findings. (1) Grounding helped most where the addition carried the answer: a read-only SQL tool lifted every core model's database accuracy from at most 16 to at least 88 percent. (2) Simple rules were hard to beat: no core model outperformed repeating today's staffing count, and two open-weight models mostly copied the median of similar earlier fire-days, a worse forecast. (3) Public releases carry hazards: a fire-danger column separates the holdout perfectly, and 67 aerial fire frames carry smoldering or fire-free labels read from a clipped thermal maximum. We release prompts, responses, scores, code, and the survey record.

Following retrieved evidence does not guarantee factual correctness: misleading evidence can induce a model to replace an answer it previously gave correctly. Standard accuracy measures obscure this behavior by combining answer replacement with preexisting errors. We introduce RAG-Stress, a controlled diagnostic protocol for examining the limits of evidence reliance in retrieval-augmented generation. The protocol holds the question and reference answer fixed, edits one assertion to support a designated incorrect answer, and crosses two source priority policies with three positions of the answer span within the evidence text. We measure misleading rate (MR) on each model's subset of questions answered correctly without retrieval, alongside clean accuracy on the full evaluation set. We evaluate fifteen systems spanning API models, open models, and search agents trained with reinforcement learning on TriviaQA-RC, HotpotQA, and SearchQA, with additional English and Chinese MedQA evaluations. Instructions that prioritize documents consistently produce higher MR than those permitting reliance on prior knowledge. Averaged over models and positions, the gap ranges from 10.9 to 13.5 percentage points across the three QA datasets. Mean MR follows End $>$ Beginning $>$ Middle under both policies, although individual models do not uniformly follow this ordering. A separate paired audit of 500 questions and two checkpoints supports increased harmful override without establishing a corresponding improvement in beneficial correction. These findings distinguish evidence adherence from factual reliability and motivate evaluating whether retrieved evidence preserves, replaces, or corrects a model's answers.

Agents combine reasoning with tools to interact with external systems and complete real-world tasks. Early agents typically interleave reasoning and actions along a single execution chain. On complex tasks, this chain becomes unreliable because growing histories obscure intermediate dependencies and allow early planning errors to propagate. Recursively decomposing a complex task into smaller subtasks offers a natural solution, yet effective decomposition must account for the system's capabilities so that each subtask can be executed by the available tools. In realistic systems, however, tool libraries can be too large to expose in full. Injecting every tool description consumes substantial context while making relevant tools harder to retrieve and useful task boundaries harder to identify. We propose tool-aware recursive decomposition, which organizes tools by functional relationships into a hierarchy of capabilities. During execution, the agent discovers tools on demand and uses the hierarchy to recursively decompose a complex task into a subtask tree whose levels are aligned with the capabilities required at each stage. Experiments on complex real-world tasks show that the proposed method improves end-to-end task success rate by up to 40 percentage points over the compared baselines. The implementation of TaReD is available on GitHub: https://github.com/WeiXiang-Mao/TaReD.

A language model may revise the same proposition after receiving genuinely relevant evidence or after receiving directional user pressure that adds no relevant fact. The observable response shift alone therefore does not reveal which source drove the change. We introduce BeliefScope, a controlled black-box framework for separating these two sources of influence around a fixed target proposition. BeliefScope crosses Evidence and Pressure with factor-specific local controls and measures response changes through probability reports, categorical judgments, and action recommendations on channel-appropriate scales. To determine when these observable contrasts support reliable attribution, we evaluate the observation design under controlled synthetic conditions. Known-truth recovery and targeted ablations establish where Evidence- and Pressure-related effects can be separated, while semi-synthetic stress tests map how that recoverability changes as the observation process becomes noisier and more heterogeneous. Across a 36-family Qwen/Llama study, with targeted 12-family checks that also include Gemma3-12B, the resulting profiles show substantial evaluation-context dependence: broad model-level differences can change under matched controls, decoding, or response interfaces, while some narrower within-model patterns remain stable. Instruction interventions further show that reduced target-aligned Pressure following can reflect either stable resistance or movement in the opposite direction. BeliefScope summarizes these measurements as a conditional belief-response profile that keeps diagnostic effects tied to the evaluation conditions under which they are observed, together with explicit validity boundaries for that diagnosis.

Affective computing has progressed from categorical emotion recognition to open-ended affective analysis with large multimodal models. Yet affective science describes emotion as an unfolding process shaped by appraisal, regulation, and social interpretation, which remains underexplored computationally. We propose TRACE, a cognition-oriented framework that formalizes an affective episode through three interrelated stages: Condition, Affect, and Effect, integrating observable cues with cognitive factors such as internal stance and regulation of emotional display. Based on this formulation, TRACE-Bench evaluates multimodal models in real-world social scenes through five tasks spanning grounded affect recognition, regulation decoding, cause reasoning, effect reasoning, and full-chain reconstruction, with 3,746 structured question-answer pairs over 646 videos. A matched human-model comparison reveals a substantial performance gap, while affect-specialized models also generally lag behind general-purpose MLLMs. Model outputs show recurring failures, including treating displayed behavior as genuine feeling and fabricating unsupported events during long-chain generation. We further propose TRACER, a cognition-grounded structured reasoning method that couples each inference with explicit premises from factual observations, cognitive appraisals, and established upstream conclusions, forming a traceable graph of intermediate and target conclusions. TRACER outperforms all evaluated model baselines on each of the five tasks. Project page: https://cogaffc.github.io/TRACE

Current frontier LLMs can theoretically process long contexts with 1M tokens or more. But to what extent can they go beyond simple retrieval and perform deeper reasoning over such long contexts? We empirically investigate long-horizon reasoning capabilities of LLMs, focusing on deductive logic expressed in Prolog. We construct ProloNg, a synthetic testbed to probe Prolog Long Reasoning, which systematically varies the complexity (reasoning depth) of problems, where the hardest case has a reasoning depth of 22 and 62k context length. We study 8 reasoning models across 5 families of frontier LLMs, and find that performance degrades substantially as reasoning depth grows, with the majority of models approaching chance beyond depth 10.

A large body of research measures model coherence based on output variance without adequately considering competing causes. We identify two such causes, ambiguity and indifference, and we introduce a set of 175 questions where contradicting answers cannot easily be explained by either. We then measure incoherence in terms of contradictions when resampling answers to the same question. In contrast to other methods our metric has high specificity, and only ranks models as incoherent when the issues are glaring. Even so, we find narrow finetunes score poorly. Inspecting inconsistencies flagged by our method, we find that model organisms from the literature display severe issues such as identity conflation, introspection failures and rationalizations. These findings suggest that the pathologies induced by narrow finetuning may limit what these models can tell us about coherent misaligned behaviour.

Adversarial games have driven advances from heuristic search to reinforcement learning, yet learning and adapting strategies from limited samples remain challenging. AI agents offer an alternative by turning game experience into revisions of executable policies. Building on heuristic learning (HL), we formalize Adversarial Heuristic Learning (AHL), a paradigm that uses AI agents as learning engines to refine game policies and supporting software while keeping model weights fixed. We introduce AAArena, a benchmark comprising 12 authentic adversarial games and 1,920 archived human programs, with an evaluation protocol modeled on real-world game competitions. Agents interpret rules, choose opponents, analyze replays, and revise game agents to achieve their highest ranking within fixed match and evaluation budgets. We evaluate \val{completedmodels} model and harness configurations: Opus5.5 with Claude Code earns 6 gold medals, while no evaluated configuration tops the remaining 6 human ladders. Performance is generally weaker in games with more complex rule specifications. Further experiments show that opponent selection and dense feedback support policy improvement, and that agents learn from both on-policy replays of their own matches and off-policy replays of other players' matches. These results highlight HL's potential in adversarial games and identify persistent challenges in game understanding, strategy implementation, and long-horizon policy development.

We propose BrickBench, a benchmark for agentic text-conditioned LEGO-set design. Given a prompt, an agent is tasked with producing an assembly that not only satisfies semantic and design criteria, but that can also be physically built. To do so, it must select parts from a discrete library and reason jointly about local and global constraints. We score validity, alignment, and design across three settings that vary in scale and part availability. We provide BrickAgent, an environment for coding agents to construct, inspect, and validate their designs. We find that leading agents largely satisfy verifiable physical and semantic requirements, but fall short of human designs. We release our benchmark and environment at http://www.brickben.ch

Looped Language Models (LoopLMs) offer a parameter efficient approach to scaling reasoning by reusing shared parameters across recurrent computation steps. Despite their promise, effective post-training of LoopLMs remains challenging. Existing approaches either provide reward based supervision that is sparse or costly to extend across loops, or rely on external teachers or privileged information, leading to limited teacher availability or teacher-student context mismatch. To address these limitations, we introduce LoopOPD, a cross-loop on-policy distillation framework that uses additional recurrent computation within a LoopLM as its own source of supervision. LoopOPD uses a frozen terminal loop policy as a compute privileged teacher for an intermediate loop student on student generated rollouts, providing dense supervision without an external teacher or privileged information. We further propose Dynamic LoopOPD (D-LoopOPD), which continually refreshes the terminal loop teacher as the shared model parameters are updated, enabling recurrent self-improvement. We characterize how distillation updates propagate across loop depths and derive sufficient conditions under which a single update yields simultaneous local improvement at both loop depths. Experiments on Ouro-Thinking models show that LoopOPD improves mathematical reasoning, while D-LoopOPD yields further gains through dynamic teacher updates. Despite being trained only on mathematical data, the resulting models also improve on general reasoning and code generation benchmarks, demonstrating that recurrent computation can serve as an effective source of supervision for LoopLMs. Our code and model checkpoints will be released upon acceptance.

Sparse attention reduces the cost of long-context attention, but existing kernels typically process queries independently, repeatedly loading and dequantizing KV entries shared across queries. We observe substantial overlap in the KV entries selected by neighboring queries, creating opportunities for cross-query reuse. We present QUILT, a workload-aware sparse-attention execution mechanism that jointly processes neighboring queries and reuses shared KV entries to reduce redundant memory traffic and computation. QUILT introduces Shift-and-Compare Set Decomposition (SCSD), which transforms irregular set operations into regular data-parallel primitives suitable for modern accelerators, and pipelines SCSD with attention computation to hide its overhead. Cascaded sharing captures reuse hierarchically at multiple granularities. A tile-aware execution strategy balances sharing granularity with hardware tile utilization and selectively removes low-importance query-specific tails to eliminate underutilized tiles. We evaluate QUILT on LongBench using GLM-5.3 and DeepSeek-3.2 under both tensor and sequence parallelism. Compared with the state-of-the-art sparse-attention kernel, QUILT reduces average kernel latency by up to 55.1% and processed KV data by up to 55.9%, while reducing time-to-first-token (TTFT) latency by up to 36.8% with negligible accuracy degradation.

Cross-model KV-cache reuse remains a key challenge in modern LLM serving. Coding agents and multi-model systems increasingly route a shared context across models: a user may switch models mid-session, or a cascade may escalate a difficult query. Because KV caches contain model-specific representations, each switch typically forces the receiving model to prefill the entire context from scratch. Recent work shows that closed-form linear maps can translate KV caches between models in the same family, but transfer accuracy degrades as the model-size gap widens. In this paper, we establish that these transfer failures are concentrated in a small subset of information-dense tokens. To bridge this gap, we introduce RaReCache, a framework that enables a large target model to decode accurately from a cache prefilled by a much smaller source via selective recomputation. RaReCache identifies these critical positions using a novel rank disagreement metric, scoring each token by the energy of its mapped KV in output directions weakly supported by the calibration data. Across two model families and five benchmarks, on a 23x parameter gap (Qwen3-0.6B to 14B) recomputing just 30% of positions retains 95-99% of the target accuracy, whereas on a 8.8x gap (Llama3-8B to 70B), recomputing 40% retains 96.5% of the target accuracy. RaReCache largely removes sensitivity to source-model size, and achieves up to a 3.04x prefill speedup. For online serving, it handles 1.8x the request throughput of target prefill on a single GPU, and at the target's saturation load, reduces median and 99th-percentile time-to-first-token (TTFT) by 5.0x and 6.4x respectively, with a 30% recompute budget. RaReCache establishes an efficient serving paradigm where small models prefill on behalf of massive targets, enabling large models to recompute only critical tokens, drastically reducing prefill latency.

Neural Machine Translation (NMT) models, while capable of producing highly fluent outputs, remain vulnerable to hallucinations, which are translations that are natural yet semantically unrelated to the source. This vulnerability is acute in low-resource settings like Sinhala-to-English, where weak cross-lingual alignment leads to hallucinations. This paper introduces a framework for reference-free hallucination detection in this language pair. We present a 45,000-sample synthetic dataset generated through a probabilistic chain of five linguistically motivated corruption strategies, with a semantic rescue mechanism that uses character-level similarity to distinguish hallucinations from morphological variants. We fine-tune mDeBERTa-v3 for token-level sequence labelling, reaching a token-level F1 of 0.841 +/- 0.001 over three seeds on a source-disjoint test set, and study a three-signal ensemble integrating neural risk scores, sequence log-probabilities, and cross-lingual semantic embeddings (LaBSE). A source-ablation control shows that the detector relies on the Sinhala source rather than on surface artefacts of the corruption process: shuffling or removing the source reduces sentence-level AUROC from 0.970 to chance. We benchmark eight NMT systems spanning five model families and find that detector firings vary by an order of magnitude across architectures.

While domain-specific Large Language Models (LLMs) have encoded vast biomedical knowledge, their limited context windows often hinder a deep understanding of nuanced relationships within and across texts. To address this limitation, we introduce BioBigBird, a bidirectional language model pre-trained on extensive biomedical literature and clinical data, specifically designed to handle long-range dependencies. BioBigBird leverages a sparse attention mechanism to process sequences up to 4096 tokens, and its training incorporates a multi-stage process to mitigate noise from the large-scale pre-training corpus. We further enhance its performance by employing a multi-task learning (MTL) framework that jointly optimizes for Named Entity Recognition and Relation Extraction. Comprehensive evaluations on the BLURB benchmark reveal that our MTL-enhanced BioBigBird achieves highly competitive results against state-of-the-art models. Our work contributes an effective methodology for developing powerful, long-context language models for specialized domains, demonstrating the value of extended sequence processing for complex text analysis. Our models are publicly available at https://huggingface.co/collections/bisectgroup/biobigbird.

The genome holds the blueprint that governs the biological properties of the cell. Consequently, advancing our knowledge of genomic function is crucial both for a broader understanding of biology and for continued biomedical advances. The success of large language models on natural language and protein sequences has motivated similar efforts on genomic data. However, standard genomic language models (gLMs) often require extremely large computational resources and still fall behind traditional methods on some downstream tasks. Recently, MSA-based pretraining has been proposed as an efficient alternative, but existing models are limited to short input contexts, restricting their use to short-range tasks, such as variant effect prediction. In this work, we present RAGenome, the first retrieval-based gLM that scales pretraining to longer contexts (100$\times$ longer than existing MSA-based gLMs), allowing it to capture both across-species evolutionary relationships and within-species longer-range interactions. Trained on whole-genome alignments from 100 vertebrates, RAGenome substantially improves the long-range capabilities of MSA-based gLMs, raising gene finding performance from 0.45 to 0.60, while remaining competitive on purely evolutionary-based tasks like prioritizing pathogenic variants. RAGenome provides competitive gLM performance at a fraction of the training cost, unifying evolutionary modeling and long-range capabilities within a single, flexible, scalable framework. Code is available at https://github.com/PanosAntoniadis/RAGenome.

Tabular in-context learning (ICL) has emerged as a training-free and accurate paradigm for tabular prediction, but current approaches to compressing its in-context examples face an accuracy-throughput tradeoff: fixed subsets can sacrifice accuracy, while query-specific retrieval limits cache reuse and batching across queries, reducing throughput. We propose QCOC (Query-Calibrated Operator Compression), which exploits the exchangeability and repeated use of in-context examples by compiling their full KV cache once into compact memory shared across subsequent queries. Instead of retaining raw examples, QCOC clusters their states into joint-KV prototypes, preserves per-cluster multiplicities and the original example count, and calibrates prototype values against attention query vectors produced by the in-context examples through an anchored closed-form solution. Prototype compression drives the speedup, while value fitting helps preserve accuracy. On 64 held-out OpenML-CC18 datasets, QCOC achieves the highest mean accuracy among the compared compression and retrieval methods at both retained counts. Across 12 configurations on seven long tables, it ranks first among compressed methods in ten and averages 0.23 percentage points below full context. Compressing 8,192 in-context examples to 512 memory slots yields a 10.5x cache compression ratio; excluding one-time compilation, in a single-core CPU online-serving comparison over 1,000 queries, QCOC is up to 508x faster than dynamic retrieval baselines and 1.98x faster than full-context inference. These results show that QCOC enables compact-memory reuse and efficient inference across queries while retaining accuracy close to full context.

Increasingly, AI interviewers are being developed to elicit open-ended responses in applications like market research, public polling, preference elicitation, and social science research. However, evaluating AI interviewers is challenging because they function in extended, multi-turn interactions where they must adapt to participant behaviors. To address this need, we develop InterviewPlayground, a simulation environment for evaluating AI interviewers using simulated study participants whose behaviors are grounded in social theory. Simulated studies in InterviewPlayground produce an InterviewReportCard, which assesses the performance of AI interviewers using a suite of validated measures. To test whether our simulation-based evaluations predict performance with human participants, we conduct 15 real qualitative studies with five AI interviewers, three interview topics, and 450 human participants and compare them to simulated studies in InterviewPlayground. We find that AI interviewer performance in InterviewPlayground predicts performance in human studies with an average Pearson correlation of 0.86 across 12 measures, and the simulated interactions from InterviewPlayground reproduce key findings from behavioral analysis of AI interviewers in the human studies. Together, these findings support the validity of InterviewPlayground in assessing AI interviewer performance and examining potential failure modes. Our work contributes a simulation environment for AI interviewers supported with empirical validation, and more broadly, a roadmap for future work to develop validated, simulation-based evaluations of conversational AI systems.

Instance-specific rubrics are common in AI benchmarks where reliable evaluation requires specific expert knowledge. This approach is difficult to scale, prompting research into the generation of rubrics with large language models (LLMs). However, even when expert rubrics are available as references, it is unclear how to productively evaluate the quality of generated rubrics at scale. In this paper, we introduce a method for the evaluation of rubric generation, which we call Interventional Transfer (IT), based on the idea that two rubrics are similar if they move together when a response is perturbed to pass/fail one of them. In contrast to existing approaches for evaluation of rubric generation, we argue that different forms of interventional transfer can be used to evaluate the utility of generated rubrics for different tasks. For instance, we apply this approach in a case study on HealthBench, where we demonstrate an asymmetry in rubrics generated by Qwen3.8-27B, Deepseek-V4-Flash, and Opus-5, used to evaluate responses from GPT-5.6-Terra. Perturbations that degrade responses according to the generated/expert rubric transfer into lower scores on the corresponding expert/generated rubric, but perturbations that improve on one rubric do not reliably transfer into higher scores on the other. We argue that this finding has implications for the usage of LLM-generated rubrics for performance monitoring and hill-climbing. We contrast our approach with existing approaches for rubric evaluation, which do not surface the same asymmetry that we observe.

Accurate evaluation is central to selecting which LLM to deploy, yet testing a candidate on live traffic exposes real users to an unvetted model. Teams therefore evaluate candidates offline, on data produced by already-deployed models. This is off-policy evaluation (OPE), and it faces two distribution shifts: as a model is updated in post-training, its responses diverge from the logged ones (policy shift), and the reward definition under which it is judged changes with business requirements (reward shift). Classical OPE methods are ill-suited to this continual-deployment setting because they are defined per task and require fitting from scratch on every new logged dataset or reward definition. To address this, we propose PFN-OPE, a prior-data fitted network that amortizes OPE across a distribution of contextual-bandit tasks. We pretrain it once on tasks constructed from a pool of LLM responses scored by several reward functions, in which both shifts occur. At test-time it maps a logged dataset and one sampled target response per prompt to a value estimate in a single forward pass, with no per-task fitting. On HelpSteer2 and UltraFeedback with Qwen, Llama, and Gemma policies, PFN-OPE achieves 2.0 to 9.3 times lower error than the best baselines across all tested configurations in the reward-shifted settings.

Time Series Language Models (TSLMs) offer a promising path toward time series understanding by reasoning over temporal signals and producing natural language answers and explanations. A common approach is Chain-of-Thought (CoT), which generates step-by-step rationales linking relevant signal patterns to final answers. Although these models learn from reference CoT traces during post-training, generating faithful descriptions of input time series at inference remains challenging. Expressing high-dimensional, continuous temporal representations in discrete language tokens may cause the model to neglect task-relevant patterns or describe them inaccurately. Because later reasoning steps build on these descriptions, early errors propagate, leading to incorrect answers with plausible explanations that are inconsistent with the input signal. We propose Lapras (Latent Post-trained Reasoning Across Series), a post-training framework that equips TSLMs with latent reasoning. A model trained with Lapras reasons through a sequence of continuous thoughts in the joint time series-language space, producing text only for the final answer. It learns this through teacher-student self-distillation, where a teacher trained on CoT reference traces reasons explicitly through text. The student aligns its hidden states with the teacher's at the answer stage, transferring the teacher's reasoning ability into its latent computation. We evaluate Lapras across four TSLM backbones on five time series question answering benchmarks. Lapras improves average F1 by up to 10.79% over explicit CoT while generating 23.9x fewer tokens. Lapras's continuous thoughts can also be decoded into readable reasoning traces via standard language decoding, preserving textual explanations. Together, these results highlight Lapras as a promising post-training paradigm for efficient, effective, and interpretable TSLM reasoning.

Mixture-of-Experts (MoE) architectures have become essential for scaling large language models, with recent work demonstrating the benefits of fine-grained expert designs. Training such models from scratch is expensive, and sparse upcycling from pre-trained dense models is an attractive alternative. However, we identify a structural pathology of fine-grained upcycling: when fine-grained experts are derived from a single source model, naive routing collapses and downstream accuracy drops to near-random (e.g., on Qwen3-1.7B, Drop-Upcycling-fine-grained reaches only 23.2% average accuracy across 15 benchmarks, essentially matching from-scratch training at 22.2%, while the same method's coarse-grained variant reaches 50.2%). We propose DivMoE, the first framework achieving fine-grained MoE Upcycling with structurally-balanced routing. DivMoE introduces domain-specialized fine-grained expert initialization, deriving experts from dense models that have undergone domain-adaptive continual pre-training, and diversity-constrained routing, a hard structural constraint guaranteeing that each token activates experts from distinct domain groups. Across two base models and 15 benchmarks, DivMoE consistently outperforms six upcycling baselines (55.6% vs. 51.6% for the strongest baseline on Qwen3-1.7B) and strictly improves over the dense base model on every benchmark after Stage 2 continual pre-training -- closing the regression gap that has plagued prior fine-grained upcycling. After supervised fine-tuning on a public reasoning mixture, our 12B-parameter DivMoE model matches Moonlight-MoE (16B) at 64.5% average accuracy while outperforming a controlled NVIDIA-Upcycling baseline by 6.1 percentage points.

Retrieval-augmented generation (RAG) grounds language models in external corpora. Agentic RAG enables iterative search, yet exposes the model to isolated chunks without document structure, making it difficult to distinguish relevant evidence from chunks that merely resemble the query. Structure-aware methods such as PageIndex navigate document structure but cannot scale to the structures of large corpora, which do not fit in the LLM context. Hence, they first commit to a single document using a document retriever and cannot recover from a wrong choice. We propose RIT-RAG (Retrieval-Induced Tree RAG), which combines content retrieval with structural navigation. Offline, RIT-RAG builds a tree for each document from its table of contents or sitemap. At query time, it retrieves a broad set of chunks and uses their positions to induce manageable sub-trees, potentially across multiple documents. An LLM agent navigates these sub-trees, selectively reads promising nodes, and reformulates queries when needed. Thus, retrieval proposes where to look, while the agent decides what to read. Across financial, scientific, and customer-support benchmarks, RIT-RAG achieves the highest answer accuracy among vanilla, graph-based, and agentic baselines. On EntQABench, our new benchmark of 2.84 million technical-documentation webpages, it improves accuracy by 6.8 to 11.4 points over the strongest baseline across three LLMs.

We introduce SAIL, an open model with 35B total and 3B active parameters for literature research, scientific coding, and multi-step research workflows. SAIL is developed through a science-aware improvement loop: agents built on frontier AI models analyze its task failures and construct training tasks that address the underlying capability gaps. The diagnosis examines search and evidence selection in literature tasks, scientific assumptions and reasoning in coding, and planning and revision in longer investigations. The agents draw on paper collections and scientific code repositories to build problems, interaction trajectories, and executable tasks with the required environments and tools. We repeat this loop over multiple development cycles and train SAIL through supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning. SAIL achieves competitive performance across scientific research tasks with substantially fewer parameters than leading open-weight models.

Chain-of-thought (CoT) reasoning often improves language-model performance by giving models additional computation before answering. However, explicit CoT expresses this computation as a sequence of autoregressively generated tokens. Latent reasoning replaces these tokens with compact continuous states, but most autoregressive latent-reasoning methods retain a left-to-right dependency among latent vectors. We introduce LLoCoT: a looped latent-reasoning framework that replaces left-to-right latent generation with iterative refinement of a compact latent workspace. A shared transformer is reapplied for a small number of refinement iterations, jointly updating the latent slots based on the prompt and the evolving workspace state. Using the refined state, a probabilistic head predicts a distribution from which latent tokens are sampled in parallel and used to condition an autoregressive decoder for answer generation. Training uses continuous representations derived from explicit CoT together with a final-answer prediction loss and likelihood-based supervision of the latent states. Across HumanEval and MBPP, LLoCoT achieves the highest mean among the evaluated methods, performing on par in accuracy with Reasoning SFT, our explicit-CoT baseline, while outperforming the base model, answer-only SFT and NF-CoT. Relative to Reasoning SFT, LLoCoT reduces time to the first answer token by approximately $36\times$ and reasoning-phase latency by approximately $42\times$, while increasing end-to-end throughput by $9.2\%$. This design replaces serial thought generation with parallel latent-slot refinement while retaining probabilistic latent modeling and autoregressive answer decoding.

Verifier scores now serve as both benchmark metrics and training rewards for large language model (LLM) agents, and a change in score is routinely read as a change in capability. It may instead reflect a change in the evaluation. We introduce TRACE, a protocol that turns a score change from a verdict into a testable diagnosis: it applies a targeted change to one part of an evaluation, compares paired runs, checks whether the agent's behavior changed, and rescores unchanged trajectories to test whether the scoring rule is responsible. In a controlled suite of 25 synthetic tasks, renaming tools lowers a scripted agent's score by 0.250 even though it performs exactly the same operations; restoring the original names at scoring time closes the entire gap, while the same mutation exposes a genuine behavioral failure in a second agent. On public $\tau^2$-bench tasks with four LLM agents, an initial 30-task study finds mixed reward changes whose one clear effect does not replicate. In a larger follow-up on 88 new tasks with repeated runs per condition, renaming tools or reformatting tool outputs leaves reward unchanged to within $\pm$0.10 for seven of eight agent-change pairs, whereas tool names that deliberately mislead lower every agent's reward by 0.20-0.44, showing that the setup can detect real effects. Identical reruns flip 15-36% of task outcomes, so single-run comparisons cannot separate presentation effects from run-to-run variation. Two frontier LLM judges give consistent verdicts when a fixed trajectory is presented differently, yet disagree with each other on 57% of the same records, largely because one grades procedure rather than outcome. TRACE thus separates what a score change says about the agent from what it says about the measurement.

Large Language Models (LLMs) increasingly ship with explicit "thinking modes", yet their counterpart, "no-thinking", has received far less attention. We study LLMs' no-thinking behavior along two axes. a. How to measure no-thinking? Prior work typically defines no-thinking through proxies such as a disabled thinking mode or the absence of long traces. These proxies are unreliable: disabled thinking modes may still emit reasoning, while long traces may contain filler rather than genuine inference. We instead normalize each response into a pre-answer trace and final answer, and evaluate it at three levels: (i) Empty-Thinking Rate for strict answer-only compliance; (ii) instruction-aware Question-Pre-answer Relevance for similarity between the question and pre-answer trace; and (iii) LLM-as-judge Explicit Inference Rate for visible explicit inference. Together, these metrics distinguish answer-only output, relevant but non-inferential text, and explicit inference. b. How does no-thinking vary across tasks and models? We evaluate six prompting interventions on six LLMs across Boolean, multiple-choice, and open-ended questions. We find that explicit no-think controls cannot reliably eliminate visible inference. Models instead exhibit "Thinking Inertia": explicit inference persists even under strict controls and becomes more prevalent as the answer space opens. Accuracy remains stable on Boolean and multiple-choice tasks, whereas open-ended tasks reveal a trade-off between answer-only compliance and task accuracy. Rewriting the same questions across answer spaces shows that supplying candidate answers makes answer-only responses easier to produce. These findings establish no-thinking as a non-trivial capability: stopping explicit reasoning cannot be assumed from model settings or instructions alone and deserves systematic evaluation alongside reasoning ability.

Safety evaluations of large language models commonly summarize harmful-output behavior with attack success rate (ASR). Yet the same non-harmful outcome can arise for very different reasons. A model may recover a harmful task and refuse it, fail to recover the task, or respond to something else entirely. Distinguishing these cases becomes especially important under intent-obscuring prompts, where a low ASR does not reveal whether the evaluated task was actually engaged. To make this distinction explicit, we pair ASR with operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to be answered. Across interfaces, this paired view reveals substantial variation hidden by ASR: similar ASR values can correspond to sharply different recovery rates. Controlled English reconstructions show that recovery consistently improves as compressed prompts become more explicit, whereas ASR does not follow the same pattern. A complementary contrast comes from FormalLogic, where high recovery can still coincide with frequent harmful assistance. Together, these results show that non-harmful outcomes are not equally informative about model safety, motivating the joint reporting of intent recovery and ASR in LLM safety evaluation. Code and experiment inputs are available at https://github.com/kevinjiang0121-cyber/IRIS.

On-policy distillation (OPD) trains a student on its own generated responses using token-level teacher supervision. However, uniform weighting overlooks differences in token learning value, while existing weighting methods rely on predefined mappings from prediction signals to token weights. These mappings are not learned from the effectiveness of the resulting student updates, limiting their ability to adapt to evolving learning needs. In this paper, we propose MetaOPD, a bilevel optimization framework that jointly learns the student model and a lightweight token-weighting network. The inner objective updates the student through weighted OPD, while the outer objective optimizes the weighting network using validation loss on reference solutions after a virtual student update. Differentiating through this update connects weighting decisions to their effects on post-update performance, allowing the mapping from prediction signals to token weights to evolve alongside the student. Experiments on six mathematical reasoning and three out-of-domain datasets, covering two student scales and seven baselines, demonstrate the effectiveness of MetaOPD, with Avg@8/Pass@8 gains over OPD of 1.99/5.97 percentage points for the 0.6B student and 2.25/6.41 points for the 1.7B student.

Large language model (LLM)-based agents have demonstrated strong capabilities on complex tasks. They typically perform reasoning before each action throughout an interaction trajectory. However, reasoning may not be necessary at every turn, as reasoning produced earlier can continue to support subsequent actions. A key challenge is therefore to determine when existing reasoning remains sufficient and when a new reasoning step is needed, without relying on costly generation-based verification. We find that decreases in the likelihood of subsequent reference actions after removing additional reasoning closely track whether those actions remain recoverable given earlier reasoning, providing an effective and lightweight signal for estimating cross-turn action support. Based on this observation, we propose Reasoning Adaptation through Cross-Turn Estimation (RACE), a training approach for adaptive agent reasoning. RACE introduces a Likelihood-Guided Progressive Reasoning Cover Detection (LoGiC) procedure that progressively identifies reasoning turns whose removal has limited impact on the current and subsequent reference actions. The resulting removal signals are incorporated into both supervised fine-tuning and agentic reinforcement learning, enabling the policy to learn when to reason and when to act directly. Extensive experiments on four representative agent benchmarks show that RACE substantially reduces reasoning cost while maintaining or improving task performance.

Large language model (LLM) routing distributes inference work across different models, advancing the cost-quality Pareto frontier of LLM serving. While coarse-grained routing at the session or query level has been widely adopted in production systems, recent algorithmic work shows that fine-grained token-level routing can yield substantial efficiency and quality gains. However, efficiently serving token-level routed inference poses significant challenges to existing systems. Built on single-LLM assumptions, current systems suffer from severe step desynchronization and frequent batch admission delays under token-level routing, and they also impose high implementation complexity on developers. To address these challenges, we design TokenRouter, an efficient and developer-friendly serving system for token-level routed LLM inference. TokenRouter follows the principle of request-centric programming, model-centric execution: developers describe routing logic from the perspective of a single request, while the runtime launches a subserver for each LLM and dispatches requests asynchronously. Each subserver employs a delayed-batching scheduler, whose optimal hyperparameters are derived from a mathematical throughput model of the system. Across diverse routing algorithms, workloads, and model pairs, TokenRouter achieves 2.01-64.15x higher decoding throughput than existing systems, substantially advancing the serving efficiency of token-level LLM routing. Our code is available at https://github.com/thu-nics/TokenRouter.

The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding. Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance. Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM decoding. Specifically, at the algorithmic axis, SparseDecoding constructs calibration matrices from layer-wise activations collected during the dense-model autoregressive generation, excluding prefill, thereby aligning the pruning objective with the decoding activations. At the system axis, we develop an optimized N:M sparse matrix-vector kernel with bitmask indexing and fixed-step traversal. Substantial empirical results on representative LLMs (Llama-3.1-8B, Llama-3.3-70B, Qwen3-14B / 32B) demonstrate that our method consistently outperforms standard fixed-text calibration on the long-form generation benchmarks while achieving up to 1.48x end-to-end wall-clock decoding speedup on A100 GPUs.

When retrieved evidence contradicts an agent's prior beliefs, does it revise its answer, acknowledge uncertainty, or persist with an incorrect conclusion? Existing evaluations of agentic systems focus primarily on task success, offering limited insight into how agents handle such conflicts. We propose to evaluate agents on epistemic humility (EH): the agent's willingness to recognize, act on, and communicate uncertainty during task execution. We operationalize EH through three trajectory-level behavioral dimensions: Identify, Solve, and Escalate (ISE). Through knowledge conflict, situations where the backbone language model's parametric knowledge contradicts the evidence it encounters, or where two contextual sources disagree, we evaluate two conflict settings: (1) controlled conflict and (2) naturally occurring conflict during multi-step agentic execution, each paired with matched no-conflict controls. Evaluating four agents, we find that higher task accuracy does not necessarily correspond to greater epistemic humility: some high-accuracy configurations recognize conflicts during execution but do not communicate unresolved uncertainty in their incorrect final answers. Trajectory-level analysis further reveals that agents frequently detect conflicts in early steps of execution but fail to maintain or resolve them in later steps. Finally, we show that model-level interventions can improve EH, but often at the cost of task accuracy, suggesting that epistemic humility emerges from the interaction among the backbone model, agent harness, and evaluation environment.