2026-11-02

Large language models (LLMs) are increasingly used for personal tasks, yet evaluating their safe use of personal context remains difficult because such contexts are private, fragmented, dynamic, and socially situated. We introduce P-ContextBench, a realistic, verifiable, and scalable benchmark for evaluating personal-context understanding. P-ContextBench constructs population-structured, life-story-grounded digital humans from demographic seeds reflecting the U.S. population. Rather than using explicit, static profiles, P-ContextBench embeds traits, routines, goals, and constraints as implicit evidence across fragmented, evolving synthetic mobile-app records. It evaluates five capabilities: single-source fact recovery, multi-source fact recovery, temporal state tracking, inference of higher-level personal patterns from routine behavior, and safe handling of requests conflicting with persona-specific constraints through compliance, warning, clarification, or refusal. A multi-stage validation pipeline checks persona plausibility, context recoverability, information leakage, evidence preservation, test-case grounding, and human-judge agreement. We instantiate P-ContextBench with 100 personas, 10,088 hidden facts, 18,261 app-trace fragments, and 4,810 verified test cases. The pipeline supports configurable generation across scales and demographic groups. Cost-effective state-of-the-art (SOTA) LLMs relevant to personal agents achieve below 60% accuracy on temporal tracking, higher-level pattern inference, and constraint-aware action-boundary decisions, with fewer than 10% of test cases answered correctly across all three repeated runs. Even a high-performing SOTA model remains inconsistent on these capabilities. P-ContextBench provides measurable targets for improving the accuracy, reliability, and safety of LLM-based personal agents.

When a language model invents a Qur'anic verse or hadith, it gives false authority to text that readers may treat as sacred. Existing Islamic hallucination benchmarks are largely Arabic and often closed-ended, while Bengali resources mainly study general-domain hallucination or factual knowledge. We introduce BILAL-BENCH, a 240-question expert-verified benchmark for open-ended Islamic-source fabrication in Bengali. It covers Qur'anic attribution, hadith, fiqh, contemporary issues, and adversarial false-premise prompts, and uses a two-axis rubric that separates source errors from assertiveness. We evaluate seven open-weight models from 1B to 120B on 1,680 double-annotated responses. Six of the seven models confidently fabricate on 79.3-95.3% of source-production questions, with no monotonic improvement from scale. Hadith provenance is the weakest category across models: only two of 315 responses are correct and correctly sourced. The two 8B BanglaLLaMA variants have similar confident-fabrication rates but different error profiles. Models vary far more in whether they withhold a claim they cannot support than in source accuracy: appropriate abstention or false-premise correction ranges from 6 to 46 responses across the benchmark, and does not track parameter count.

Expert scientists excel at weighing empirical findings from multiple sources that vary in both statistical reliability and epistemic relevance. To what extent are language models, which are increasingly driving autonomous discovery systems, capable of such evidence deliberation? We introduce EDB-Bio, a scientific reasoning benchmark curated from real-world data in three biological domains: drug-perturbation transcriptomics, CRISPR screens, and human genome-wide association studies (GWAS). Each of EDB-Bio's 96k episodes presents a sequence of experimental results from published studies and asks the model to predict the probability that a target hypothesis is true. Out-of-the-box, open-weight and frontier models alike demonstrate systematic underconfidence and attain Brier scores comparable to a trivial base-rate forecaster. Post-training against strictly proper scoring rules significantly improves both calibration and discrimination. This calibration pays off downstream: in a sequential experiment selection task, our Ibis0 model is 2.5 times more efficient than the best frontier model. Throughout, quantifying evidence weights as Bayes factors reveals how models update their beliefs over time, whether they agree with expert biologists on relevance, and where they exhibit inconsistencies and biases. Our work extends calibration training and uncertainty analysis to a realistic scientific forecasting setting with applications to human cancer biology. We release all code, data, and trained models.

When a language-model agent fails a multi-step task, can another model find the decisive mistake and fix it with a single action? We introduce ReplayRCA, a benchmark that scores such diagnoses by intervention rather than by agreement with annotations: the environment is restored to the state preceding the chosen step, the proposed action replaces the agent's own, and the same agent continues. The score, causal repair, is the continuation success rate minus that of a paired placebo that replays the agent's original action through the same mechanism. Two tracks share this protocol: a post-hoc autopsy, where the diagnoser sees the whole failed run and is told that it failed, and live oversight, where a supervisor watches the run unfold and may intervene once, before an irreversible action, without knowing whether the run will fail. Tasks come from OpsWorld, a deterministic, procedurally generated back-office environment with six difficulty levels, the last of which is multimodal (QR codes, handwriting). On 62 calibrated systematic failures, the best of 13 frontier models (Grok 4.7) reaches a causal repair of +51.3 points; the top eight models are statistically indistinguishable, the flagship models of the two largest vendors rank 8th and 12th, and 5 items are repaired by no model. Forcing a strong model to intervene at the last step costs it 17.1 points, so the last-step shortcut of annotation-based attribution benchmarks does not transfer. On 87 runs, the two best live supervisors are statistically tied and one of them is a small open-weight model (DeepSeek V4.1 Flash); the supervisor that is among the best post-hoc diagnosers (GPT-6.1 Sol) is the weakest live one (-14.5 points on the same runs, p = 0.002), so post-hoc diagnostic performance can be a poor proxy for live oversight. The environment and the evaluation code are released at https://github.com/jeremy1392/RCAreplay.

An LLM agent that helps a patient get care must do two things. It must reach the correct outcome, and it must follow the necessary procedure. For example, the agent must get the patient's permission before it shares the patient's records. We introduce BABA-Bench, a benchmark of 47 interactive worlds. We built each world from a de-identified, real patient advocacy case that expert advocates reviewed. In each world, the agent must find hidden information and coordinate insurers, clinics, suppliers, pharmacies, and community organizations. The engine checks each action at the time the agent does it. It checks if the agent has permission, if the agent has evidence for the action, and if the action is prohibited. We evaluated 14 frontier models with three runs on each world, for a total of 1,974 runs. Agents often reached the correct outcome by an invalid route. Of the 921 runs that reached all required outcomes, 254 (28%) did not follow the necessary procedure. Claude Opus 5 reached all required outcomes in 81.6% of its runs, but only 52.5% of its runs passed all checks. When we gave two models a verified route, their pass rates increased from 29.8% to 91.5% and from 44.0% to 97.2%. For these models, the main problem is to find a route, not to execute it.

To accelerate materials discovery, language models must be able to address problems at the research frontier. Yet textbook- and examination-derived benchmarks primarily assess established knowledge, leaving performance on frontier research problems largely untested. We introduce FrontierMat, a benchmark of 330 problems drawn from frontier academic and industrial research in fields including aerospace, semiconductors, embodied intelligence and nuclear fusion. Every problem is validated by leading experimentalists and industry experts in its field. Each problem presents a real study and asks the model to infer a withheld outcome or draw a conclusion from selected experimental evidence. Problems are annotated by functional domain, material system and the scale of the governing mechanism; 130 require cross-scale reasoning. The strongest general-purpose model scores only 53.7%, and on matched problems, cross-scale problems score 8.7 points below single-scale ones. Half of the problems are solved by at most one model, and models fail chiefly by endorsing what is plausible rather than what was observed. A domain-adaptation study with meta-weighted continued pretraining and scale-modular self-distillation raises an open-weight model from 36.7% to 55.4%, matching the strongest proprietary model, with improvements on public materials benchmarks and no decline in aggregate performance on general benchmarks. FrontierMat provides a testbed for measuring progress toward language models that support frontier materials R&D.

The needle-in-a-haystack (NIAH) benchmark was introduced as a test for measuring a model's ability to find small relevant facts within a very large body of input text. However, existing NIAH benchmarks each suffer from one or more of the following limitations: (1) reliance on synthetic data, so their evaluations may not generalize to real-world settings; (2) reuse of previously published datasets, leading to the risk of benchmark contamination; (3) dependence on manual annotation, which requires human effort to create and update data and is prone to future benchmark contamination; and (4) unrealistic simplification of long-context multi-hop reasoning where each next step is trivial given previous steps, while real-world reasoning often requires planning over multiple possible proof paths. To address these limitations, we introduce \Benchmark, a dynamic benchmark built from real citation relationships among scientific papers published after the evaluated models' knowledge cutoff date, which utilizes real-world information while ensuring no benchmark contamination. The fully automated benchmark construction pipeline samples representative subgraphs of the citation graph formed by these papers and converts each subgraph into a long-context QA example with a verifiable answer. We use GraphNIAH to evaluate eight LLMs from the GPT-5.6, Gemini-3, and Qwen3.5 families. We find that accuracy falls as questions require longer reasoning paths: at question depths of 16 and 32, the GPT-5.6 and Qwen3.5 models answer YES/NO questions no better than random guessing, mostly by answering NO; four of the eight models answer no counting example correctly, and the others model get worse as the subgraph size increases. These findings suggest that current LLMs often stop searching the citation relationships in their context before they reach the answer, even though every fact they need is in the context.

Conversational assistants are expected to help people with diverse backgrounds, preferences, and communication styles. Yet evaluations centered on a single cooperative user provide limited insight into how reliably an assistant serves users with different personas. We introduce PersonaChatBench, a framework for evaluating task competence and responsiveness to diverse user needs. PersonaChatBench pairs the same underlying task with multiple simulated users, each guided by a private persona card that specifies distinct needs, preferences, and interaction behaviors. By holding the task fixed while varying the user, the framework examines how assistant performance changes across user conditions. We evaluate 20 language models across six domains, separately assessing native task outcomes, fulfillment of expressed persona needs, and simulator fidelity. Our results show that task performance and need fulfillment vary across personas and domains. Strong task performance does not necessarily translate into satisfying individual needs, and models' relative strengths in meeting those needs are not consistent across domains. PersonaChatBench makes these differences measurable, revealing gaps obscured by aggregate scores and helping identify which users an assistant serves well and where adaptation is needed. It provides a basis for evaluating consistency across users and guiding the development of assistants that better accommodate diverse needs in everyday applications.

Self-evolving agents rewrite a persistent state from feedback, but it is largely untested whether this lets them learn a rule nobody told them. An agent can also pass by writing down each correction as an exception, and its list then grows with every new rule. We introduce CLR-Bench, a Curriculum of Latent Rules benchmark. Each chain adds one hidden rule per level on top of the earlier ones. A rule acts on an opaque label that stands for a named grouping published on the Web, such as the founding members of an airline alliance. The task gives one example value, practice feedback reveals one more member and one non-member, and the test checks unseen members and similar non-members. Probes certify that the base model knows the members of each grouping and does not guess it without feedback, and that the evidence identifies the grouping once labels are read as named groupings. An automated pipeline builds chains from a new data domain, and we release 20 chains over 10 domains. We evaluate six recent self-evolution methods. With the certifying model they barely learn. The best gain at most 5 points over a frozen agent and exclude only 6-12% of the unseen members from the count, fewer than the 20% reached by simply keeping the raw practice history. A fixed lesson that states the named-grouping convention lifts the same methods by 14.5 to 16.5 points. Their states hold the revealed member in 75-98% of cases and almost never the class name, and when practice exposes a member the list misses, they add that member to the list. With the stronger GPT-6-Sol, ACE and Dynamic Cheatsheet learn from feedback alone. They reach 78% accuracy and solve 59-60% of the single cells, against 53% and 8% with the certifying model, and they surpass the raw history. CLR-Bench therefore measures how well an agent learns, and this ability differs across models.

Web world models increasingly predict the consequences of candidate actions before agents act, yet a plausible predicted future does not necessarily support the correct safety decision. In a matched human evaluation of 3,000 predictions from ten frozen open models, 86.1% are judged semantically plausible. However, several models exhibit highly asymmetric authorisation behaviour, combining low false-safe execution with poor recall of clarifiable ASK cases and a strong preference for BLOCK. This motivates safety sufficiency: whether a predictive interface preserves and exposes the policy-decisive information required to distinguish actions that may be EXECUTED, require clarification through ASK, or should be BLOCKED. We introduce SafeWebWorld, a benchmark for evaluating safety sufficiency across five web-safety domains through action-conditioned futures, policy-decisive evidence, and policy-conditioned decisions. Under a controlled Security-Aligned Supervised Fine-Tuning protocol, evidence preservation can improve substantially while authorisation-sensitive decisions deteriorate. To determine whether this deterioration reflects loss of authorisation-sensitive score structure or a failure of the decision readout, we introduce Latent Authorisation Recovery (LAR). Frozen-score recalibration reveals substantial but backbone-dependent recovery, while paired recovery provides only a modest additional gain. Yet within-family and matched-authorisation evaluations remain poor, showing that aggregate recoverability does not imply robust fine-grained authorisation behaviour. Overall, plausible prediction, evidence preservation, and aggregate recovery are each insufficient on their own to establish robust authorisation.

Fairness benchmarks for large language models (LLMs) are overwhelmingly English-born and American in their social categories; translated versions carry those categories into societies where they do not apply, and almost none fix their hypotheses before the models are run. We present Qistas (Arabic qistas, "the just balance"), a protocol for evaluating how LLMs treat the social, confessional, and regional constituencies of a society, portable across countries and instantiated first for Iraq. In its pilot form it measures only what needs no verdict on contested truth: procedural symmetry across matched counterfactual pairs, differential refusal, stereotype reliance under ambiguity, and terminological framing of documented events. Every measurement is paired by construction and analysed with paired machinery (sign-flip permutation tests, exact McNemar tests, template-clustered bootstrap intervals, Holm-corrected families), with a power analysis that states, before any data exist, the smallest effect each axis can detect after the design effect of template clustering. We release the authoring apparatus (a pruned 3x5 taxonomy with four cells closed by design, per-type checklists, a cumulative rulings log, a structural audit, and a human review notebook) and report its first field test: a semantic review of a 58-item seed bank in which 24 of 25 template groups failed at least one check. The defect patterns it exposed (evaluation criteria leaking into prompts, token-level swaps that break Arabic agreement, epistemically unbalanced pairs, single-theme refusal probes, unwritten dialect twins counted as coverage) are, we argue, general hazards of counterfactual fairness benchmarks. This is a protocol paper, not a results paper: no model has been evaluated, and the accompanying pre-registration commits the hypotheses, metrics, exclusion rules, and analysis plan that the pilot will follow.

Human preferences change over time: an established preference may remain valid over time, become uncertain as context shifts, or be explicitly revised. Effective personalization requires conversational AI to adapt its response strategy to evolving preferences rather than relying on passive memory retrieval. We formalize this capability as adaptive personalization under epistemic uncertainty. A human preference study further confirms that people favor state-contingent response strategies as preference uncertainty changes. To evaluate whether language models mirror this adaptation, we create a multilingual benchmark spanning 250 personas, 5 languages, and 25K conversational turns that decouples strategy selection from response execution. Across eight closed- and open-weight models, we uncover a recognition-generation gap: models reliably identify appropriate strategies when selecting among candidate options, but collapse during open-ended generation, suffering a 35-46% drop in strategy calibration. Crucially, this failure is systematic: models default to exploit-first behavior, over-relying on existing preferences regardless of preference epistemic uncertainty. We show this failure is addressable: while prompting interventions fail, targeted training that supervises models' internal strategy representations toward state-conditional human preferences narrows the user acceptability generalization gap by a third on out-of-distribution personas and conversational styles.

Digital agents can now gather and organize information, reason over it, and use tools to complete complex tasks. Forecasting puts all of these abilities to the test: it requires integrating incomplete evidence, inferring causal drivers, and committing to calibrated probabilistic judgments. Existing benchmarks are poorly suited to measuring and improving this ability. Live benchmarks provide feedback only after events resolve, so every evaluation cycle waits on the calendar; fixed-corpus evaluations return results immediately but confine research to pre-collected documents. We introduce FABench, a scalable backtesting environment for Forecasting Agents. An automated pipeline discovers events, generates questions, and resolves outcomes for any past forecast date, so the benchmark can be renewed to stay ahead of model training cutoffs; its first release poses 287 questions across nine domains. Agents research each question freely in an archive environment that combines an offline archive with live web search screened to prevent temporal leakage. Evaluating agents built on eight recent models, we find that retrieval improves the Brier score of all eight in every retrieval setting and lifts the leading agents to the market consensus on the 54 questions linked to Polymarket, with the best three scoring slightly above it. Research traces reveal recurring failures in how agents weigh and apply evidence, leaving ample room for improvement. We release our code, data, and archive environment as a foundation for community research on forecasting agents.

Agentic LLM assistants must repeatedly decide whether to answer from memory or to call a tool such as web search. Unnecessary searches cost latency and money; missing ones yield confident but unsupported answers. Prior work mostly measures verbalised confidence or trains new adaptive-retrieval methods, and rarely audits the search decisions that deployed products actually make. We introduce a benchmark of 400 multiple-choice questions that combines widely known inside-knowledge (IK) questions, which should be answered directly, with outside-knowledge (OK) questions about facts from news published after the models' training cut-offs (to our knowledge), which should trigger search or, without tools, abstention. Each item offers an abstain option, and two action-level scores, NAS and KIS, reward searching on OK questions and restraint on IK questions. Across five commercial LLMs with and without their vendors' web search, GPT-5 mini and Grok 4 Fast search on over 90% of OK but only 1% of IK questions, whereas Mistral Small searches on just 20% of OK questions, often abstaining instead of using its tool; one product's search indicator proved unusable. Searching is partly triggered by explicit temporal cues: on uncued OK questions, the NAS of the two selective models falls from about 0.72 to 0.57--0.58. The scores thus measure the search triggering of specific products, not metacognition in a strict sense.

The ability to integrate new information and deduce its consequences is central to learning, understanding, and reasoning in large language models (LLMs). Whether information arrives during training or in-context, models must not only store local facts, but propagate their long-range implications globally. To explore information integration in LLMs, we introduce two complementary natural language benchmarks for large-scale evaluation of deductive reasoning as a function of inference depth. The first is a formal graph benchmark enabling controlled evaluation over a wide range of graph topologies and integration tasks; the second is a knowledge composition benchmark grounded in real-world knowledge graphs. We use these benchmarks to compare knowledge integration from information encoded in context versus information encoded in model weights during pretraining, evaluating a set of open models implicitly, without chain of thought. We find that accuracy decays sharply with increased composition length for both pretrained and in-context knowledge. Information presented in-context, however, substantially outperforms deduction over information encoded in weights. These results quantify a fundamental limitation of current models with direct implications for prompt-based control of agent behavior and for continual learning strategies, and they provide a foundation for developing training, prompting, and integration methods that support deeper reasoning.

As Large Language Models (LLMs) become increasingly multilingual, distinguishing human-written from AI-generated text is essential for maintaining information integrity. However, research remains English-centric, with limited benchmarks for low-resource languages such as Hindi. To address this gap, we introduce SHAKTI, a large-scale Hindi benchmark covering four genres, eight LLM families, multiple prompting strategies, and adversarial transformations. We evaluate its quality through human annotation and analyze linguistic, grammatical, and sentiment-based differences between human- and AI-generated texts. Building on these findings, we propose TRIDENT, a round-trip translation-based detection framework with three variants: TRIDENT-S (static threshold), TRIDENT-D (dynamic threshold), and TRIDENT-C (logistic regression using multi-distance RTT features). TRIDENT leverages semantic consistency between original and round-trip translated texts to identify AI-generated content. Extensive experiments show that TRIDENT outperforms conventional and supervised baselines, with TRIDENT-S achieving the best overall performance. Further evaluations across seven languages, adversarial settings, and unseen LLMs demonstrate its multilingual applicability and robustness, while highlighting sensitivity to the translation model.

Large language models increasingly generate structured outputs consumed by programs and downstream systems, where semantic correctness alone does not guarantee reliable execution. We introduce SERIAL-7, a Korean-English benchmark built around an executable verification framework for machine-verifiable generation across seven representations. Using 177 validated canonical artifacts, SERIAL-7 separates semantic derivation (Derive), constrained artifact generation (Realize), and cross-representation preservation (Preserve). Its 43 atomic representation constraints are paired with executable checks over format, structure, and task-specific semantics, enabling progressively stricter verification from representation compliance to full artifact validity. Across ten models, REALIZE-L3 HSR ranges from 42.0% to 93.9%, while Full Compliance exposes additional artifact-level failures beyond correct final answers and valid representations. These results show that reliable structured generation requires explicit, executable verification of how artifacts are realized and preserved across representations.

Choosing among the hundreds of thousands of available large language models (LLMs) requires a thorough understanding of each candidate model's properties. In practice, each property is measured by its own benchmark suite, so profiling a model on many properties means paying for many separate evaluations. To address this, model embeddings compress each model into a fixed-length vector from which many properties can be predicted. Existing methods, such as EmbedLLM, build these vectors from correctness patterns and question features, and often need large probe batteries to do so. However, many behaviors that are important when choosing a model, such as its verbosity, safety and style, show up more in how it answers than in whether it is correct. We introduce BULLPEN (Behavioral Understanding of Large Language models via Profile ENcodings), a framework that extends EmbedLLM in three ways. First, BULLPEN organizes the downstream tasks that prior work uses and adds 20 task benchmarks we built on behavior, style and provenance. Second, it is the first framework to read a model's own answer text jointly with its correctness and the question text. Third, it encodes models from a few hundred questions instead of the full 24,118-question pool. Across 260 open-weight models, our answer-text encoder scores 0.48 +/- 0.10 on the new tasks against 0.35 +/- 0.10 for the best prior encoder. Our encoders that read answer text keep at least 91% of their full-pool skill from only 400 questions (1.3% of the inference compute of the full question pool), while a correctness-only baseline keeps 83% with the same 400 questions. Because these embeddings need only model outputs, they support behavioral auditing and low-cost profiling, and could extend to API-only models. Code and data: https://github.com/wcquesera/bullpen

Given that AI agents can search the web broadly and deeply in real time, it is tempting to conclude that their outputs account for recent world events. We show that this conclusion does not hold in practice. While agents can search and easily find recent changes, the correctness of their responses depends on whether they search and what they search for. Popular agentic search benchmarks like BrowseComp cannot expose failures wherein agents do not recognize their own assumptions that should be verified, since these benchmarks explicitly state what needs to be found. In contrast, many user requests state only the final information need (e.g., an itinerary for a trip), leaving the agent to discover relevant changes (e.g., the temporary closure of a museum) and account for their consequences. To highlight such gaps, we introduce BlindToChange, a benchmark that measures whether agents proactively discover recent events that affect their responses. Our benchmark comprises requests for recommendations where an obvious candidate is no longer applicable due to a recent real-world event. These events are drawn automatically from Wikidata, allowing BlindToChange to be periodically refreshed. Strikingly, we find a big gap -- as high as 27.5 points, even at high reasoning effort -- in systems' performance, when the recent event is not explicitly mentioned compared to when we ask about it directly. Analyzing search trajectories, we learn that more than half of the failures can be attributed to the fact that agents do not directly search to verify whether a relevant fact has changed. Our benchmark highlights, and tracks, a new axis of progress, wherein systems account for recent information that is not explicitly mentioned in the user request.

While prior work has highlighted that systematic nation-level bias in large language models (LLMs) can pose operational risks for international relations (IR) applications, many existing evaluations still lack a clearly specified unbiased reference (ground truth), limiting fully quantitative and cross-setting measurement. Using an expanded corpus of 722 official UN Security Council (UNSC) records, we measure nation-level bias with five complementary tests, thereby extending prior approaches to cover a broader set of bias expression channels: opinion ranking (WOW), association sorting (AK+-), vote simulation (PVS), opposition attribution (F-Oppose), and framing (Headliner). To enable grounded bias evaluation, we explicitly define an unbiased status for each test; in particular, for WOW and AK+- where ground truth is difficult to specify, we propose an action-based pseudo unbiased status derived from observable voting behavior to construct a quantifiable baseline. Benchmarking 4 LLMs (GPT-4o-mini, Llama 3.3-70B, Mistral-22B-Small, and Qwen3-30B) across the five permanent members of the UNSC, we find a test-agnostic macro-pattern: relative to the reference baselines defined for each test, model outputs lean toward the opposition (red) side for China and Russia and toward the supportive (blue) side for the U.K., the U.S., and France. These results underscore the need to audit IR-facing LLM deployments with grounded, multi-test benchmarks tied to explicit reference points.

Multi-turn negotiation remains a challenging problem for large language model (LLM) agents. While LLMs can conduct natural-language dialogue, existing benchmarks predominantly assess only agreement rates, overlooking whether negotiated commitments actually create value once executed. To address this gap, we introduce TradeWar, a multi-country simulation benchmark that translates negotiation dialogues into concrete tariff and trade-share policies, enforces capacity and network constraints during settlement, and evaluates outcomes via an Economic Health Index (EHI). To solve this execution-grounded task, we propose TACTIC (Trade-aware Adaptive Consolidation and Theory-of-mind for International Commerce), an agent that iteratively refines its negotiation strategy across episodes-without updating base LLM weights-by leveraging prior experience, opponent behavior, and past settlement outcomes. Despite using zero-shot chain-of-thought (CoT) prompting, PromptLLM attains a 75% agreement rate but only 57.1 final EHI (*100), whereas TACTIC reaches 68.8. Across 20 sequential episodes and three seeds, TACTIC's final EHI improves from 58.2 to 73.2, compared with 65.2 for an ablation without cross-episode updates. Under technology-controls and energy-crisis shocks, TACTIC also exhibits smaller performance degradation than PromptLLM and MemoryLLM. These results indicate that CoT deliberation and interaction memory alone do not fully address delayed execution feedback, whereas settlement-grounded adaptation improves long-horizon negotiation in TradeWar.

Effective computer-use agents (CUAs) should seamlessly combine visual GUI interaction with code and terminal execution within a single policy. We present Mage-VL-Cowork, a 4B hybrid CUA trained from a general multimodally aligned checkpoint through an end-to-end, progressive five-stage curriculum across midtraining, SFT, and RL. To power this curriculum at scale, Mage-CUA-Forge replaces per-task agent construction with reusable, skill-conditioned generation workflows and application-specific preparers, synthesizing 293,806 desktop and web tasks across 6,915 skills while reducing recurring per-task expenditure by 82.6%. Rather than directly distilling a default teacher, we evolve the teacher harness and instantiate specialized CUA harnesses to yield high-quality trajectories alongside stage-specific supervision. Mage-VL-Cowork-4B finally climbs to 66.0% on OSWorld-Verified through different training stages with more efficient reasoning cost, achieving state-of-the-art performance at its parameter size and pushing forward the frontier of on-device intelligence.

Recent scientific agents build knowledge recursively by retrieving evidence, constructing an intermediate state, and judging whether that state contains sufficient information. These agents are evaluated end-to-end, usually without ground truth, so the effect of each design choice is unknown. Here, we examined two design axes: (1) what the recursion builds (the state of evidence) and (2) what controls it (the judge that decides when to stop). To measure them, we defined consensus scientific claim verification and measured these effects on this task, where a system must establish a claim's consensus stance from the open literature. We built a new consensus claim dataset, ProClaim-eval, including 718 claims with SUPPORT and REFUTE gold labels from five expert-curated databases spanning molecular biology, nuclear physics, and astronomy. Our instrument, ProClaim, persists an evidence state outside the model's context, so each axis can be studied with the other held fixed. Against verification systems and general-purpose agent harnesses, ProClaim ranks among the top three, with a significant advantage over the harnesses. We found that the evidence state improves performance over in-context accumulation, and it also matters what is stored in it. In addition, a trained judge that reads the evidence state performs better than fixed-round stopping. A judge trained separately and never exposed to ProClaim-eval achieves the highest agreement with the gold labels on three of five subsets, where it matches the off-the-shelf Claude Sonnet~4.6 judge and outperforms Claude Haiku~4.5. Our results support recursively maintaining a structured evidence state outside the model's context as an effective design for verifying claims in scientific agents.

LLM agents serve natural-language requests through tools, data sources, and other agents, and even simple requests often need several interdependent calls. Typically, the LLM picks these calls by trial and error, with no guarantee of an executable plan. We present Capability-Grounded Agentic Planning (CGAP), which instead reasons about what a request requires and what tools provide. CGAP compiles the tool catalog deterministically into a PDDL model, uses an LLM to turn the request into goals, and solves the model with a classical planner. Since tool outputs are unknown in advance, the model is abstract, tracking which inputs are available, not their values. During execution, an LLM adds entity-specific calls that abstract plans omit, and deterministic checks reject any that break the plan. Every argument is drawn from a cited source, such as a conversation span or an earlier tool response, and each requested result counts as delivered only after an LLM verifier checks it against the request. With three LLMs on ToolSandbox, ComplexFuncBench, BFCL multi-turn, and $\tau^2$-bench retail, CGAP outperforms a standard tool-calling agent by up to 28.9 points.

Speech language models have expanded automatic speech recognition (ASR) from transcription to general speech-to-text generation. Yet existing benchmarks still assume a single ground-truth transcript per utterance, an assumption that struggles when models produce readable, formatted, or structured outputs. In this work, we introduce Grounded Generative Speech Recognition (GSR), reformulating speech recognition as generating semantically grounded text instead of a single transcript. We define a grounded reformulation space that characterizes all valid textual realizations supported by the input speech, and establish a four-level hierarchy of reformulation capabilities, including literal transcription, canonical reformulation, semantic reformulation and speech-grounded structuring. Based on this formulation, we present GSR-Bench, the first comprehensive benchmark designed to evaluate generative speech recognition systems beyond exact transcript matching. GSR-Bench covers diverse scenarios including refined meeting transcription, cross-dialect translation, context-aware speech understanding, and robustness under silent or noisy acoustic conditions. To address the limitations of word-level evaluation metrics, we develop a multi-dimensional evaluation framework combining speech grounding assessment and LLM-based semantic judgment, whose reliability is validated through human preference studies. Extensive experiments across representative architectures, ranging from conventional ASR systems to end-to-end Speech-LLMs, reveal fundamental trade-offs between transcription fidelity and generative reformulation ability. Our analysis further identifies critical failure modes, including acoustic hallucination, unsupported generation, and instruction drift, providing insights into the design and evaluation of future grounded generative speech systems.

Process reward models (PRMs) score reasoning steps, but best-of-N selection also requires a rule that aggregates those scores. We study that rule as a separate evaluation object. StrictPRMBench fixes mathematical reasoning pools, scorer evidence, and the source trace of a deletion intervention, so that different reducers face the same recorded selection problem. Across seven scoring checkpoints and two benchmarks, the stored N=32 outcomes give higher deletion-induced selection instability for product than geometric mean (GM) in all 14 cells. The equal-cell difference is +0.0965; a shared-problem bootstrap, conditional on these checkpoints and frozen pools, gives a 95% interval of [0.0846,0.1086]. A pre-injection-defined sensitivity excludes replacement slots that are either reducer's clean winner and retains 4,459 of 4,622 cell-problem records. The retained difference is +0.0986 [0.0875,0.1098]. Effects are heterogeneous: the two Qwen cells have positive percentile-bootstrap intervals on retained support but two-sided exact paired p-values of 0.0625 and 0.0703. Clean-pool Qwen accuracy does not detect a product advantage, but this is not an equivalence result. We provide versioned outcome records, CPU reanalysis, and an explicit evidence-availability contract. The contribution is a diagnostic evaluation protocol and a reproducible audit of recorded selector behavior, not a new PRM, a universally optimal reducer, or evidence of deployment-wide harm. Raw-score replay and outcome-record reanalysis are distinguished: missing primary edited-score vectors limit independent evaluation of new reducers under the same stress.

Existing visual preference evaluation relies on costly human annotations or preference reward models, whereas online marketplace transaction prices are readily available and reflect real purchasing behavior, offering potential low-cost supervision. However, transaction prices conflate visual preferences with non-visual factors such as provenance, demand, and competition, making aggregate accuracy insufficient to distinguish model errors, unreliable price signals, and protocol confounds. To address this identifiability challenge, we introduce MarketImage, a diagnostic framework that identifies when transaction prices reflect visual preferences, where they fail, and whether multimodal language models can recover human-supported visual preference signals. MarketImage comprises three components: (i) MarketImageArt, 266 six-image comparison sets with transaction prices and price-blind human rankings; (ii) MarketImageLens, a tie-preserving protocol that partitions 3,990 within-set image pairs into price-human agreement, conflict, and price-tie regions while separately auditing presentation-order effects; and (iii) MarketValuator, six elicitation methods spanning direct comparison, independent scoring, budget constraints, and multi-agent aggregation to evaluate signal recovery by static visual scorers and three multimodal language models. Experiments show that only 24.2% of image pairs have a non-tied price direction, of which 34.7% agree with the human majority. Within this region, HPSv2 leads at 66.4% Recoverable Accuracy; Qwen's budgeted and aggregated variants trail baselines, while GPT interface gains remain statistically unresolved. Thus, transaction prices encode recoverable visual preferences only under limited conditions and should be treated as uncertain proxies rather than ground truth.

Large language models offer a promising route to register-transfer-level (RTL) design, but reliable implementations require checking and revision based on electronic design automation (EDA) tools' feedback. Existing RTL-specialized models often lack learned interaction with such EDA tools or rely on separate generation and repair components. We introduce RTL-Interact, a multi-task post-training framework that trains a single RTL-specialized model to generate RTL, construct task-appropriate EDA checks, and revise implementations using tool feedback. The framework combines automatically constructed multi-task demonstrations with staged supervision from chain-of-thought to tool-integrated reasoning. We further introduce Pattern-Aware Policy Optimization (PAPO) for multi-task reinforcement learning. PAPO augments Group Relative Policy Optimization with correctness-gated rewards that favor concise successful checking while rewarding effective revision, and uses task grouping to preserve coverage across objectives. On CVDP, RTL-Interact-32B achieves 48.7% weighted Pass@1 and 68.9% Pass@5, surpassing its teacher model, DeepSeek-V3.2, while RTL-Interact-8B achieves 44.9% weighted Pass@1. Additionally, we show that PAPO can reduce average tool calls by 63.8% while improving Pass@1. Both models also outperform all RTL-specialized baselines on VerilogEval v2 and RTLLM v2, demonstrating the effectiveness of our method.

Length-aware scheduling reduces request latency in LLM serving by prioritizing shorter requests. However, despite their large model sizes, length predictors can suffer degraded ranking performance under domain and task shifts, and adapting them can require costly retraining. Their prediction overhead can further offset the benefits of scheduling. To improve the generalization and efficiency of length-aware scheduling, we propose \textbf{LEAP}, a length prediction system with fast online adaptation and asynchronous CPU execution. LEAP reuses the serving model's prefill hidden states, which contain information about the output length, and extracts this signal through a linear readout. To adapt to substantial domain and task shifts, it updates the readout at test time through recursive least squares, using completed requests as supervision without gradient computation or historical-data replay. Prediction and updates run asynchronously on CPU, while prediction and initial batch ordering overlap the remaining GPU prefill. Across the evaluated cross-domain transfers, LEAP retains strong ranking performance, and online adaptation improves Spearman correlation by 64.7\% under a larger domain shift. Serving experiments show that LEAP reduces average per-token latency by 15.9--40.8\% compared with the evaluated length predictors, while achieving 1.9\% higher throughput with $30{,}569\times$ fewer predictor parameters than the strongest evaluated baseline.

Looped language models can strengthen reasoning through repeated computation with shared weights, increasing execution depth without adding independent parameters. However, each additional pass increases training floating-point operations (FLOPs), making these gains more expensive to obtain. To address this cost, we propose UpLoop, a pretraining schedule that uses upcycling to train looped Transformers while exactly matching a standard dense baseline in final parameter count, total training tokens, and training FLOPs. Specifically, UpLoop begins with a shallow Transformer, progressively grows its depth by copying learned layers, and uses the computation saved early to support repeated execution of a shared middle block later in training. Experiments across model scales show improvements in language modeling and downstream performance, with the average score across 22 downstream tasks increasing from 22.99 to 27.52 at 1.3B parameters. These gains also extend to mathematical reasoning after supervised fine-tuning.

Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context windows. However, at ultra-long context (above 1M tokens), full attention is prohibitively expensive. A promising approach is sparse attention, in which a dedicated indexer selects a subset of keys for the attention. Yet current methods of indexing either still have quadratic cost in context size or lose accuracy at ultra-long context scale. We address this with Latent Sparse Attention (LSA), an efficient attention method that maintains accuracy at ultra-long context. In LSA, keys are hierarchically clustered in a learned latent space, and queries score the hierarchy of clusters instead of scoring every key, making prefill, decode and training faster at long context. We propose two variants, Distilled-LSA (D-LSA), which can be applied to pre-trained models via distillation, and Native-LSA (N-LSA), which is trained natively on the next-token loss alone. On GLM-5.3-FLASH (QWEN-3.8-27B) at 10M tokens, D-LSA speeds up one attention layer by 22.7x (72.4x) in prefill and 4.6x (42.5x) in decode over its original sparse attention (full attention), while keeping similar accuracy across a variety of tasks. N-LSA matches the accuracy of full attention with a training speedup of 1.6x per layer at 128k tokens. LSA offers practical and efficient attention for models to benefit from ultra-long context.

Producing long-form videos from narration requires interdependent decisions that remain coherent across shots. Existing agentic editors often use multi-stage workflows driven by general models, leaving the quality-cost trade-off of compact, task-adapted policies unclear. We address this gap with WeVisCut, a long-form script-to-video system that connects tool execution with policy learning. Its agent harness maintains shared production state so that tool feedback informs subsequent decisions and repairs. Because finished videos do not expose these decisions, our four-stage data pipeline combines editing inversion with verified replay to turn recovered editing intent into executable supervision. Applied to 3,750 finished videos, this pipeline yields over 130K training examples. We use this supervision to train a single 9B policy that coordinates production and recovers from failures within the same harness. Moreover, we introduce a 100-task benchmark and five metrics assessing editing quality and delivery reliability. Under a shared tool environment, WeVisCut substantially outperforms its base model and achieves comparable performance to Claude Opus 5 while using nearly 33x fewer generated tokens per successful output minute. It also outperforms task-matched native editors on our proposed metrics.

Training a large language model is compute-bound, and current recipes divide the work between optimizers by hand: orthogonalized momentum for attention and MLP matrices, Adam-style updates for the rest. This division rests on human priors about which layer type deserves which geometry, and it fixes a single global update budget that every parameter block shares. We introduce FARO, a function-space adaptive optimizer that makes the partition of parameters into blocks the unit of decision. FARO gives each block a geometry by its realized counterfactual return, then freezes the routes and reallocates the frozen per-block budget by cross-block equal risk contribution (ERC), the risk-parity principle from portfolio theory. On a GPT-style model trained on FineWeb, the framework gives two results. First, its FARO-Adaptive arm reaches 3.2780 nats at the full schedule, matching NorMuonAndAdam (3.2794), the best hand-designed recipe: it recovers the human prior without hand-setting per-layer geometry. Second, FARO-ERC improves on that arm: it lowers the clean-data loss at every one of 6 seeds, by 0.0136 nats on average, a modest but robust gain. Its sign test, paired test, and confidence interval all agree, and it exceeds the measured run-to-run noise at no added time or memory. It works by trimming the riskiest blocks rather than expanding the safest, and it holds up under a corrupted stream, still lowering the mean loss and winning the seed majority. The source code is available at https://github.com/wxj630/FARO.

Agentic systems increasingly gate individual steps with a pointwise judge that scores a response against a criterion. Small language models are attractive for this role because they can be served and specialized efficiently. Such a judge must locate the criterion's decision boundary, but pointwise labels under-determine that boundary in fine-tuning: labeled responses to different queries differ along many dimensions at once, so they do not reveal which features decide the label. We introduce CPA (Counterfactual Pairwise Augmentation), which makes the deciding difference explicit in training. For each pointwise item, CPA synthesizes and audits a counterfactual response to the same query intended to carry the opposite label. It frames the accepted pair as a pairwise task asking which of the two responses satisfies the criterion. The judge is fine-tuned jointly on both pointwise and pairwise datasets and deployed in pointwise mode only. A second-order analysis shows when this should outperform the prevailing strategy of merging separately trained pointwise and pairwise experts: merging moves every parameter direction toward the pairwise expert by the same fraction, whereas joint training can hold the pointwise optimum wherever the pairwise loss is flat and absorb its curvature where it is sharp. This is the regime that CPA's same-source, counterfactually minimal pairwise view is designed to produce. On separate Qwen3-8B LoRA adapters for six binary criteria, CPA improves criterion-averaged pointwise macro-F1 by 2.5 points over pointwise-only training and outperforms model merging on all six criteria at every tested merge weight. It also surpasses the strongest zero-shot frontier judge we evaluate, Gemini 3.6 Flash, by 11.8 points.

LLM agents execute tasks through repeated model calls over evolving histories of reasoning, actions, and tool feedback. Aggregate speculative-decoding metrics merge these distinct decisions, providing limited guidance on where draft supervision is needed. We present TraceSpec, a framework connecting trajectory evaluation with adaptive draft training. It combines dynamic execution with fixed-prefix replay and introduces TraceSpec-Traj, a corpus of 10,000 training states and 2,500 test states spanning browsing and software engineering. Analysis of five speculative baselines reveals that aggregate acceptance can rise with context length while declining within matched late browsing actions. Workload and phase further distinguish these trends, motivating supervision guided by trajectory conditions. TraceSpec trains on intermediate assistant turns and allocates a fixed prediction budget using state-conditioned estimates of prefix-aware acceptance gaps. Across three target sizes and two temperatures, it improves replay throughput by 12.5% to 22.6% over its DSpark base. Qwen3-4B ablations show that adaptive allocation increases accepted length by 3.3% to 10.0% over fixed full-turn supervision, with additional comparisons substantiating the roles of state conditioning and consecutive-token agreement. The resulting evaluation resources, trajectory findings, and training strategy connect the analysis of agent execution to practical draft improvement.

Long-horizon LLM agents play an important role in solving complex and sustained tasks, with ever-growing interaction contexts. However, these accumulated contexts will raise substantial computational cost and obscure key decision-relevant evidence. Existing methods typically train external compressors or context-management agents to rewrite context as shorter hard-token sequences, but such external selection may not align with the LLM's evolving representation and is necessarily lossy. In this paper, we introduce Agentic Soft-Token Compression (ASTC), a training-free compression method using soft tokens within LLM agent for long-horizon tasks. Specifically, ASTC uses the model's native recurrent dynamics to determine the already absorbed information, leveraging the commonly-used hybrid Gated Delta Network (GDN)-attention. It then carries this information forward as agentic soft tokens while retaining exact values as hard tokens. The resulting compact sequence reduces later-layer computation and yields reusable KV caches, extending the savings across the agent's trajectory. Extensive experiments on long-horizon tasks show that ASTC matches or exceeds the uncompressed agent's accuracy with up to 50.68% fewer end-to-end Attention Pairs, while trained compressors require 15.4 to 103.7 times more Attention Pairs per successful task. These findings highlight agentic soft tokens as a model-native way to carry absorbed context forward as the task representation evolves.

Overly restrictive eligibility criteria contribute to widespread enrollment failure in clinical trials, yet published recommendations for how specific criteria should be broadened remain disconnected from trial design at scale. We introduce RECITE (Revising Eligibility Criteria Incorporating Textual Evidence), a framework for automated discovery, extraction, and implementation of evidence-based eligibility criteria revisions. RECITE uses protocol amendment documents from ClinicalTrials.gov to construct a dataset of 11,913 curated instances across 5,735 trials, each pairing original criteria with documentary evidence and target revisions. We benchmark fifteen LLM configurations under truncated-to-context-window and retrieval-augmented generation conditions, on a 3,116-instance sample under an LLM-as-judge protocol validated against ten domain experts (judge vs. human quadratic-weighted Cohen's $\kappa_{qw} = 0.694 \pm 0.096$). The best open-weight model, Gemma 2 9B, achieves $\kappa_{qw}$-aligned ordinal score $3.0 \pm 0.9$ on a 0--4 scale (normalized 0.742) and the best closed-weight model, GPT-4o, reaches $3.0 \pm 0.8$ (normalized 0.745). Using an agentic literature discovery system, we then identified 44 paper--trial matches across five publications providing trial-design and EC-modification evidence. Our findings suggest that existing clinical knowledge can be operationalized into actionable trial design recommendations.

Unlike autoregressive (AR) models, diffusion language models (DLMs) permit parallel token updates, flexible generation, and bidirectional attention, but leading systems rely on costly pretraining or full-model fine-tuning. We present Build Your Own DLM (BYOD), a generalizable LoRA-only recipe for parameter-efficient AR-to-DLM conversion. Without architecture-specific tuning, BYOD induces diffusion generation in Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B; to our knowledge, this is the first reported AR-to-DLM conversion of a Mistral model. The adapters comprise 4.1-5.8% of final parameters while every base-model parameter remains frozen; each 25,000-iteration conversion needs at most 102.4M tokens, processed in 3.1-4.0 hours on one Google Colab GPU. All four models generate from fully masked sequences, use bidirectional context in a controlled attention test, and retain substantial parent-model knowledge. BYOD-Gemma is the strongest BYOD conversion on our nine-task evaluation, reaching 50.6% average accuracy compared with 51.9% for LLaDA 8B Instruct in the same harness. Open-generation quality at low network function evaluation (NFE) budgets remains the principal limitation: BYOD perplexity is 8.3-11.1 versus 4.1 for LLaDA and 2.2-3.0 for the AR parents when generating 128 tokens in 64 steps. To enable further experimentation and inspection, we provide a standalone training and evaluation implementation and interactive demos for all four conversions. BYOD thus provides a practical route for extending diffusion generation to other existing model families while identifying the gap that efficient conversion must next close.

In real-world deployments, large language model (LLM) systems are increasingly evolving from single-model invocation toward heterogeneous deployments in which different models are dynamically invoked for inference. When the same context is reused across models, however, each newly selected target model typically has to process it again, repeatedly incurring substantial computational costs. Therefore, we propose ACCT (Asymmetric Cross-LLM Context Transfer), which targets KV cache transfer between models of different scales within the same model family. ACCT transforms the KV cache produced by a source model into sparse KV states that can be directly consumed by a target model, thereby enabling cross-model reuse of context computation and decoupling context computation from downstream inference. ACCT consists of three lightweight components: the Cartographer identifies query-relevant context tokens for each target layer and KV head; the Allocator determines how many context tokens each target layer and KV head retains under a shared total KV budget; and the Compiler transforms selected source KV into sparse target KV. The target model can therefore perform inference using the transferred KV without reprocessing the original context. We evaluate ACCT in two complementary settings comprising five general-capability benchmarks and six highly context-dependent tasks. Experiments show that ACCT matches or exceeds the target model's original inference performance even when reducing the KV cache size by 75%. At a context length of approximately 19K tokens, ACCT's KV conversion achieves a 6.5-21.2 times speedup over target context prefill. These results demonstrate the feasibility of transferring context computation across models under a limited KV budget.