2026-10-27

Molecular cloning is one of the most routine wet-lab workflows, yet existing evaluations of language models in biology mostly test knowledge recall or isolated steps rather than end-to-end experimental planning. We introduce BenchBench-Cloning, a benchmark for large language models comprising 65 molecular cloning tasks grounded in sequence records, laboratory notebooks, and published experimental studies. Planning a molecular cloning experiment requires defining a set of constructs needed for a scientific objective and a construction route to produce them. BenchBench-Cloning contains 50 protocol-development tasks, in which agents produce an executable protocol for a specified construct, and 15 construct-design tasks, in which agents determine the constructs a scientific objective calls for. Protocol-development tasks are evaluated by execution in a cloning simulator; construct-design tasks are evaluated with task-specific deterministic checks and, where needed, rubric-based LLM-as-a-judge evaluation. Across eight systems with repeated attempts per task, GPT-6 Astra scores highest at 30.6%, with the others between 8.6% and 22.2%, and the benchmark remains unsaturated under best-of-ten sampling. As integrating individual cloning decisions into complete, valid plans remains challenging even for frontier models, we expect BenchBench-Cloning to serve as a useful benchmark for tracking progress in experimental planning.

We introduce the Literature and Linguistics Analysis Benchmark (LALA), comprising 236 open-ended questions written by experts in linguistics and literary studies, together with expert-provided ideal answers. By comparing models within families and at similar training scales using human and LLM judges, LALA identifies differences in linguistic competence associated with post-training, offering a framework for evaluating future models and guiding targeted improvements. We administered questions to 17 language models and evaluated their responses using both human expert assessment and two LLM evaluators, Sol and Qwen. Base models displayed little of the linguistic competence assessed by the benchmark, whereas their post-trained versions performed markedly better, suggesting that post-training plays a central role in making linguistic knowledge accessible and usable. Parameter count, by contrast, showed no statistically significant association with partial-credit performance, and competencies remained unevenly distributed across linguistic domains. We interpret this pattern as evidence that post-training adds linguistic competencies piecemeal, without producing a general linguistic intelligence capable of extending them beyond the learned policies that support them. Models performed best on logical relations between statements and selected discourse-level tasks, while even the strongest post-trained models continued to struggle with ambiguity, literary style, reference, and misleading assumptions. Human and Sol evaluations produced nearly identical model rankings in both the original and ReliableEval runs, supporting the replicability of the comparative results.

Tool-using language model agents are usually evaluated by whether they finish the user's task. That metric misses a common failure: an agent can complete the requested work while also taking an unauthorized side-effecting action, such as writing to the wrong object, expanding the scope of the request, or contacting an unintended recipient. We call this failure mode collateral damage (CD). We introduce SABench (Side-effect-Aware Benchmark), a benchmark of 201 email, calendar, and ticket tasks. Each task crosses one of four state structures (independent, dependent, externally reachable, cross-tool) with one of four trap families (ambiguity, mis-targeting, cascade harm, overreach), and each task has a deterministic trace-based oracle for both goal satisfaction and authorization compliance. Across 2,211 model-task pairs from 11 frontier LLMs spanning major commercial and open model families, 15.2% of all runs incur CD, with risk concentrated in ambiguity tasks (38.8% violation rate) while non-ambiguous traps average 8.4% CD. Among goal-satisfied runs, 12.8% still produce unauthorized writes. Cross-tool and dependent states have about twice the CD rate of independent states. CD rates vary by 4.7x across models at comparable goal rates, which suggests that task success alone is a poor proxy for safe tool use. The SABench task corpus is released at https://huggingface.co/datasets/Phaedrus/sabench; the traces, oracle, code, and analysis pipeline are released at https://github.com/hjj194/sabench-artifact.

LLM-based sales agents are often evaluated on static policy questions or task success, but deployment asks a sharper question: when a model can state a rule, will it choose compliant, tool-grounded actions under commercial pressure? We introduce R2A-Sales (Rule-to-Action), a policy-atom-matched benchmark for this question. Each item pairs a static probe with a multi-turn sales scenario sharing the same versioned policy atom, facts, evidence, behavioral constraints, and tool contract, so differences cannot be attributed to policy-content mismatch. In the interactive arm, a code-owned ledger verifies committed tool effects and terminal states; unsupported action claims receive no credit. We report static compliance $K$, grounded action compliance $A=\Pr(\text{safe terminal}\land\text{no hard violation}\land\text{non-echo})$, and the matched gap $G=K-A$. The frozen release spans three controlled Chinese-language sales domains, 24 policy atoms, 42 semantic cores, and 2,016 matched pairs. The confirmatory study audits 21,888 runs over 19 cores, covering eight backbones plus four Qwen3-8B prompt treatments (12 configurations). Static compliance is near ceiling $(K=89.9\%\text{--}99.1\%)$, yet eight-backbone action compliance is only $A=3.0\%\text{--}17.3\%$, giving $G=80.4\text{--}92.5$ percentage points. Since $A$ also requires safe termination, $G$ upper-bounds rule-to-action degradation rather than isolating pressure alone. No prompt-only treatment dominates all metrics. R2A-Sales shows that sales-agent reliability must be measured at committed action, not rule recital or task completion alone.

Large language model (LLM)-based personal agents are increasingly deployed to handle everyday tasks on users' behalf. These agents have access to private user information, while their tasks involve communicating with third parties; this raises a critical privacy concern that agents may inappropriately reveal sensitive user information to these recipients. Existing benchmarks primarily assess direct privacy leakage, where an agent explicitly states a pre-specified secret or attribute. However, they do not capture \textit{inferential leakage}, where the model leaves indirect cues in the reply that implicitly leaks sensitive attributes. We introduce \textsc{InferPriv}, a benchmark that measures whether an honest-but-curious recipient can infer a user's sensitive attributes from an LLM's reply on routine task. Evaluating five frontier models, we find that inferential leakage substantially exceeds direct leakage across models and prompts. Under a privacy-aware system prompt, GPT-6.1-Sol, Sonnet-5.5, and Gemini-3.8-Flash achieve near-zero direct leakage, yet the inferential leakage rate remains between 16.2--25.2\%. These findings suggest that evaluating explicit leakage alone can substantially understate the privacy risks.

Omnimodal large language models (Omni-LLMs) are increasingly expected to assist users while observing continuous audiovisual streams. However, existing evaluations provide limited insight into how reliably these models connect stream evidence and user needs to timely, sustained interventions. We introduce \textbf{Omni-InteractScope}, a benchmark of 1,479 human-annotated, cross-reviewed interactions with 1,790 temporal targets across eight task groups. It links input modalities, event triggers, intervention types, and user needs, covering event-timed tool invocation and query-free assistance alongside notifications, warnings, state tracking, and continuous responses. Our experimental results on the proposed benchmark reveal three critical limitations of current Omni-LLMs: (1) \textbf{Timely tool invocation remains unreliable}: all 12 evaluated models produce an on-time first response for at most 40\% of tool-invocation targets, and joint timing-and-tool accuracy reaches at most 32.0\%. (2) \textbf{Query-free reminders remain poorly covered}: all evaluated models achieve lower timely reference-target recall on query-free reminders than on explicit questions, measured on their respective target sets. Reminder recall never exceeds 23.0\%. (3) \textbf{End-to-end guidance coverage remains limited}: nine models cover none of the 26 guidance sequences in full, and the remaining three cover only one to three sequences each. Targeted controls further show that tool-invocation scores alone do not establish event grounding and that reminder coverage depends on response history and frequency. These findings motivate developing and evaluating streaming assistants for timely action, initiative under general user needs, and coverage of entire response sequences.

Large language models (LLMs) show promise in specialty medicine, yet existing evaluations largely focus on isolated clinical encounters rather than tracking how clinical responsibilities evolve across a longitudinal care pathway. Here we introduce NPC-RubricBench, a phase-aware benchmark for evaluating multistage clinical reasoning across the cancer care continuum, using nasopharyngeal carcinoma as a defined disease context. Curated from longitudinal records of 405 patients across two cancer centers, the benchmark comprises 941 patient-phase units across six sequential care phases and 17,688 patient-specific atomic criteria spanning five evaluation dimensions, with each assessment strictly restricted to decision-time clinical evidence. Evaluating 30 general-purpose and medical LLMs revealed that strong diagnostic performance does not carry over to downstream care. We evaluated 30 general-purpose and medical LLMs, showing that high diagnostic performance did not translate into improved downstream care. Instead, the dominant failure modes reversed across phases. Explicit decision errors dominated anatomical staging at 48.2%, whereas decision omissions dominated initial work-up and post-treatment follow-up at 33.4% and 30.4%, respectively. In the evaluated model set, medically adapted models showed higher omission rates than general-purpose models. Moreover, varying query granularity exposes a critical specification trade-off: broad open-ended requests degrade Decision Quality across all 30 models, while more detailed, checklist-like prompts raise Decision Quality but reduce Evidence Faithfulness and Safety Adherence. Furthermore, NPC-RubricBench demonstrates that static diagnostic benchmarks are insufficient to ensure consistent clinical reliability over time, highlighting the importance of phase-adaptive safeguards in multistage medical AI.

Speaking-style captioning (SSC) aims to generate natural language descriptions of how speech is delivered, capturing paralinguistic attributes such as vocal timbre, prosody, and expressivity. Many downstream applications, including conversational AI agents, controllable speech generation, and large-scale audio annotation, require dimension-specific style captions that describe targeted aspects of speech (e.g., speaker traits, emotion, or delivery style), rather than a single undifferentiated description. However, existing work lacks a unified formulation for controllable speaking-style captioning. We address this gap by formulating SSC as an instruction-following audio-language task, where explicit instructions specify the speaking-style dimensions to be described. We introduce StyleInstructCaps, the first standardized benchmark for controllable speaking-style captioning in audio-language models. StyleInstructCaps provides a dataset and evaluation framework that measures metadata groundedness, hallucination, instruction-following ability, generalization to unseen instructions and audio datasets, and speaker-style consistency. Across nine state-of-the-art audio-language models, we find that strong instruction-following ability does not necessarily imply grounded speaking-style understanding, and that SSC-specific instruction-diverse training substantially improves groundedness, controllability, and speaker consistency.

Coding agents are moving from handling one request at a time to serving as persistent assistants to which users hand a stream of work: a build, a data-processing job and a bug fix may all be pending on the same machine at once. Such an agent must decide when to take on new tasks, how to delegate and overlap them, and how to stay within the machine's CPU and memory. Existing benchmarks cannot measure this, because they evaluate each task in isolation. We introduce \benchname{}, which evaluates an agent system as a workload manager: verified tasks arrive over time and must be delivered under shared CPU and memory limits, and the same tasks run one at a time serve as a per-system reference. Across five model--harness pairs, isolated task success does not carry over evenly. On heavy workloads every system loses 13 to 23 points. On medium workloads, systems span 11 points in isolated success but 26 points in stream success, and two systems 0.6 points apart in isolation deliver 54\% and 68\% of the same tasks. Execution-policy interventions trace much of this gap to task intake: serial dispatch leaves 73\% of heavy-workload tasks unclaimed. Yet more elaborate management does not necessarily help. A fixed concurrency of two comes within four points of the best agent-managed policy on both medium and heavy workloads, and under every policy median CPU use stays below 30\% of the provisioned cores while tasks wait. These results identify workload management as a distinct capability for agents, and \benchname{} as a way to measure it.

LLM agents take instructions from the operator, repository configuration, user, tool results and saved memory, which often conflict. Instruction-hierarchy benchmarks ask whether the higher-ranked source wins, but a source sets position, role and label at once. First, we build WhoWins, 1,080 code-scored scenarios in five task types, each with up to five mutually exclusive instructions, and seven controlled experiments that vary only how instructions are delivered. Second, we study 14 open-weight models (0.6B to 122B, thinking off and on) over 1,887,356 scored replies, checking main findings on held-out scenarios. Inside one message the first of two conflicting lines wins 72.42 to 94.66%, and across messages it keeps 22.95 to 62.65%. A policy sent first in the system (or developer) message loses most conflicts with later lines in every Qwen model but wins most on gpt-oss when it acts. In tool and file-write tasks it wins less often than the same text sent first as a labelled user message, in every model-mode. Third, we test whether pairs forecast larger conflicts; forecasts improve in every model-mode once they know where the task line is. Finally, by our keyword patterns at most 2.92% of decided pair replies flag or hedge a conflict.

As frontier AI systems increasingly serve as sources of political and current-events information, evaluating the epistemic quality of their responses is now especially consequential. Existing evaluations often focus narrowly on ideological alignment, rely on measures not validated for this setting, or depend on reference labels derived from public opinion or non-expert annotation, potentially producing misleading or miscalibrated results. To address these limitations, we introduce NewsBench, an expert-grounded benchmark that reframes evaluation around editorial standards developed by senior practitioners in journalism, policy, and intelligence analysis. NewsBench operationalizes quality across three dimensions, source quality, factuality, and neutrality, using a two-phase pipeline in which domain experts first define and calibrate evaluation guidelines, which are then transferred to automated AI judges validated against expert-adjudicated gold labels. The benchmark includes approximately $3,000$ evaluation prompts, with roughly $500$ refreshed monthly to capture emerging current events, roughly $2,500$ human expert-annotated items for calibration and validation, and a suite of AI judges designed to enable scalable evaluation. To validate our methodology, we report both inter-rater reliability among human experts and AI judge--human agreement across all three dimensions. Finally, we apply NewsBench to four frontier models, revealing distinct and non-overlapping failure profiles in sourcing, factuality, and neutrality.

Enterprise workflow agents typically rely on step-by-step interaction (Step-by-Step): the agent executes an action, observes its result, and invokes a large language model (LLM) again to decide the next step. Global orchestration (Global) instead generates a complete executable workflow in advance and executes it without further LLM calls, substantially reducing inference overhead, but it remains less accurate than Step-by-Step. We study how close Global can come to Step-by-Step and what narrows the remaining gap. We introduce AgentOrch, a benchmark that directly compares global orchestration with step-by-step interaction on 650 enterprise SOPs and 3,220 runtime variants under matched tools, environments, and evaluation. Step-by-Step remains more accurate for every evaluated model, but the gap has shrunk by about 79% from early models to the strongest recent model. We identify three routes to further narrowing this gap. Incremental workflow construction improves Global accuracy by an average of 9.98 percentage points (pp) without runtime execution feedback. Revising complete workflows with execution feedback improves all evaluated models, including 6.13 pp on held-out runtime variants. Targeted training on complete workflows improves accuracy by an average of 12.18 pp across all five trained models. Although a gap remains, global orchestration offers a promising path toward enterprise agents that deliver higher accuracy with fewer LLM calls.

The development of large language models (LLMs) are capable to assist with hardware experiments. In this scenario, this raises a practical question: Can LLMs determine whether a proposed operation is safe for a specific development board? Documented operating limits and safety requirements are board-specific, even across products from the same manufacturer. This variation makes board-specific safety judgments by an LLM difficult to evaluate. To address this problem, we introduce the BoardSafety Benchmark. The benchmark is built from vendor datasheets and hardware guides. For data transformation, these documents are converted into a board-specific safety rule base and are then used to create labeled operating scenarios. Each data item pairs a board and a proposed operation to an applicable safety rule with ground-truth label, indicating the level of safety. The categorized level for each operation includes safe, unsafe, or insufficient-info. In addition, the benchmark covers diverse operations, board-specific cases, boundary conditions, and missing-information cases. We evaluate a series of models by varying their access to board-specific rules and the use of structured-rationale prompting. Our results demonstrate that providing the safety rules generally improves accuracy, whereas structured-rationale prompting produces gains that vary across models and rule-access conditions. Overall, the evaluated models cannot yet reliably determine whether an operation is safe, which implies that critical exploration is required for this field.

French users may omit apostrophes and diacritics or make keyboard typos in everyday input. The effects of these common variations on language model performance and prediction reliability remain insufficiently understood. To address this gap, we construct a controlled French surface-variant benchmark covering seven task types, with 4,200 original-low-high triplets (12,600 input instances). Each high-intensity variant retains all low-intensity edits and adds further changes. We evaluate six models from two LLM families. The results show that accuracy decreases and Brier score and negative log-likelihood (NLL) increase with perturbation intensity. Larger models show smaller accuracy losses under high intensity, but increases in Brier score and NLL do not consistently decrease with scale. Keyboard typos cause the largest losses in accuracy and the quality of predicted probabilities at the predefined high-intensity settings. Finally, calibration temperatures fitted on original inputs transfer effectively to surface variants. Combining calibration with input restoration further improves Brier score and NLL over either method alone while preserving restored accuracy. Our benchmark provides a unified framework for jointly evaluating task accuracy and prediction reliability under common French surface variations.

Research ideas are always proposed under resource constraints: a plan must fit the available compute, API credit, and time. Yet evaluations of research-ideation agents mostly score unconstrained novelty or overlap with a published paper, and their LLM judges are rarely checked against peer review. We introduce AlphaIdeaBench, a benchmark of 100 ideation tasks (50 workshop calls for papers and 50 open problems stated in surveys), each paired with an explicit resource budget and a shared proposal schema. An LLM judge scores each proposal on 12 dimensions covering validity, novelty, and feasibility. To validate the judge, we compare its scores on four novelty axes (problem, method, scenario, and insight) with scores derived from reviewer comments on 600 papers from ICLR 2026, NeurIPS 2025, and ICML 2025. For problem and method novelty, Gemini 3.7 Flash is within one point of the reviewer-derived score on 94-97% of papers; agreement on scenario novelty is relatively lower. We evaluate eight published harnesses on a shared backbone and, within one harness, eight backbone LLMs. CoI-Agent obtains the highest overall score, narrowly ahead of SciMON, which leads on method and insight novelty. All harnesses score low on scenario novelty and rarely change the problem setting. The backbone largely determines which open problem a proposal targets, whereas the harness determines whether a proposal addresses one problem or combines several. Weaker backbones such as Llama 4 Scout fail to carry out that harness's strategy and score low on budget feasibility. Finally, when only the budget is varied on three topics, a larger budget changes the experimental setting on compute-intensive topics but only adds tests on an inexpensive one. We release the tasks, rubrics, and scores to foster standardized, reproducible benchmarking of future research-ideation agents under constraints.

Physical reasoning requires adapting familiar expectations to the dynamics of the observed world. Existing benchmarks establish broad physical capabilities, yet an overall score leaves unclear whether answers follow visual evidence when familiar rules no longer apply. We introduce InfiniteWorld, a benchmark that pairs physical-reasoning questions with simulator videos whose physical parameters are systematically altered. Its generation pipeline connects simulation, scene screening, paired video translation, question construction, and input checks. Five tasks assess law identification, parameter estimation, temporal reasoning, analytic counterfactual comparison, and law induction. The evaluation defines five matched input settings for each question, comprising text-only evaluation and ordered or shuffled observations from simulator and translated videos. We evaluate a diverse set of video-language models under these matched settings. The results expose a persistent preference for familiar physical rules. Frames improve performance for most models, but gains differ across tasks. Strong accuracy on Earth-normal questions coexists with poor recognition of altered dynamics, and Earth-normal answers remain common when the observed physics changes. Task scores and semantic agreement further distinguish matching physical factors from matching their amount, and correct answers from repeated errors. InfiniteWorld makes these behaviors measurable through paired changes in the evidence supplied for the same question, separating visual evidence from familiar physics.

Language-model agents must adapt as users revise their requirements and priorities during complex tasks. We introduce IntentDelta, a human-constructed multi-turn benchmark with 638 trajectories and 3,766 turns across WebShop and TravelPlanner. Each turn pairs a user utterance with a cumulative intention state and human reference actions. The trajectories are cross-verified and cover feasibility- and preference-driven changes in constraints and priorities. We evaluate eleven agents on intention recovery and action quality. Average Intention Recovery is 33.1\% in TravelPlanner and 44.7\% in WebShop, while average Action Score is 31.6% and 41.0%, respectively. Agents struggle to preserve earlier requirements and distinguish soft preferences from must-have constraints. Correct intention recovery also does not guarantee successful execution, highlighting challenges in translating evolving user intentions into actions.

Continual learning is the ability of LLM agents to improve over time by acquiring, refining, and applying knowledge from experience while retaining useful prior knowledge. It is essential for agents learning over long time horizons, yet current evaluations largely assess individual task performance rather than improvement through accumulated experience. We introduce Continual Learning Exam (CLE), a large-scale benchmark developed by 50+ domain experts to measure experience-driven improvement in realistic professional workflows. CLE comprises 120 human-authored tasks, each organised as a five-level ladder, yielding 600 levels spanning 89 professional roles across 12 occupational domains. Agents must pass each level to unlock the next, with later levels building on experience accumulated during earlier assignments. An evaluation agent assesses each submission against human-authored rubrics. We compare a single-submission setting with one that provides diagnostic feedback after failure and allows one revision. Across 17 harness-model configurations using frontier LLMs, feedback and revision yield a median completion gain of 5.3 percentage points. Yet the best-performing system clears only 47.7% of levels and completes all five levels in just 23.3% of tasks, highlighting the difficulty of sustained learning from experience for existing frontier LLMs. As such, the dataset can also be used to post-train LLMs to strengthen their continual learning capabilities.

Large language models increasingly collaborate with humans and peer AI agents whose contributions may be incomplete, inconsistent, or incorrect. Effective collaboration requires calibrated trust, whereby decisions to rely on, verify, or assist a partner adapt rationally as empirical evidence accumulates. We investigate whether models translate partner evidence into appropriate reliance policies. We introduce a benchmark of 324 scenarios across 23 collaboration categories. Each scenario involves three sequential decisions with explicit transition dynamics, feedback emissions, and cost schedules, permitting exact analytical evaluation of choices against their expected consequences over the remaining interaction. Evaluating seven models alongside 36 controlled-history probes, we dissociate evidence sensitivity from normatively appropriate action. While the strongest frontier model approaches near-optimal decisions, other models exhibit recurring pathologies: procuring excessive safeguards, purchasing verification tests with zero expected value of information, prioritizing immediate savings over cooperative learning, and misallocating assistance to suboptimal channels. Supplying explicit posterior distributions resolves certain errors while introducing new failure modes. Calibrated trust in autonomous agents critically demands coupling empirical evidence with forward-looking decision utility.

In this paper, we introduce $\Delta\tau$-Bench, an executable agent benchmark that models the structural evolution of large tool ecosystems. In $\Delta\tau$-Bench, independently evolving services continually reshape how tasks can be accomplished: familiar tools may disappear or be replaced, while more efficient alternatives emerge amid overlapping functionality and noisy documentation. Against this moving capability frontier, we evaluate whether agents can proactively discover efficient tool-use paths and promptly revise their accumulated procedural knowledge, bringing together demands that existing benchmarks fail to examine. $\Delta\tau$-Bench comprises 1,588 validated multi-turn tasks across 42 domains and six successive ecosystem states generated through capability-preserving transformations. Our evaluation shows that even the strongest evaluated model or harness achieves only 42.3% task success, while initial gains from existing skill-learning methods can not consistently translate into better adaptation as the ecosystem evolves. These findings motivate Steward, an ecosystem manager that jointly manages tools and procedural knowledge through incremental skill maintenance and dynamic provisioning during task execution. Experiments show that Steward sustains higher task success across ecosystem changes with lower skill-maintenance overhead, while improving both recovery from disrupted workflows and adoption of more efficient alternatives.

Large Language Models(LLMs) have shown strong performance on tasks with verifiable answers, but many real-world queries involve subjectivity, disagreement, and conflicting evidence. Most existing benchmarks for such queries remain English-centric, overlooking the vast diversity of perspectives held by users worldwide. To systematically study how LLM perspectives vary across languages, we curate PerspectivesQA, a manually verified multilingual dataset of globally contested topics covering 18 languages, and propose ROOTS, a framework that profiles linguistic, perspective, and framing diversity. We evaluate a diverse set of eight models covering both open-source and proprietary families at varying scales, establishing that LLMs hedge their answers in English but commit to positions in lower-resource languages. Larger models maintain a more consistent cross-lingual worldview, while smaller models are more easily swayed. All models also exhibit high sensitivity to minor surface-level input variations, producing notable stance even in larger models.

Structured semantic parsing is typically judged by whether the correct category is assigned, which leaves value recovery and instance-level associations unexamined: a system may recognize an event yet bind an argument to the wrong instance, or identify an intent yet supply an incorrect slot value. We present USP, a cross-family semantic contract that represents categories, fields, values, and instance associations through a common interface for named-entity recognition, relation extraction, event extraction, and intent-slot parsing. The contract is paired with five diagnostic matching views---category agreement, field recognition, field-value pairing, instance-bound assignment, and complete-instance recovery---that ask distinct questions about the same output. Using this contract, we normalize 61 datasets into a unified corpus with shared schemas, prompts, and matching rules, and compare six models on a fixed test subset. Our fine-tuned Xing4.0-29B-Intent attains 57.70 task-balanced Field-Value Micro-F1 (equally weighted across the four task families) and ranks first in every family and on all five views; its margin over the strongest alternative widens from 2.93 points for category matching to 20.73 points for complete-instance matching. This widening gap shows that category agreement alone understates structural deficits. USP thus offers the community a shared representation, a normalized 61-dataset corpus, and a diagnostic evaluation protocol that make structural agreement measurable and comparable across extraction and intent-slot tasks.

AI agents are increasingly proposed as autonomous experimenters, yet the benchmarks that test them mostly offer cheap experiments that answer at once. Experiments in physical facilities are different: they take months, occupy scarce equipment and cannot be undone. We present SlowLab, a benchmark in which an agent runs a one-year campaign in four simulated greenhouse compartments on a site whose crop, facility and prices are hidden. The agent decides when to plant which climate policy, whether to stop a crop and which measurements to read while crops grow, and at the end recommends one policy, which a private evaluator scores in weather the agent never saw. In a preregistered study on 48 unseen sites, three low-cost language-model agents ended near a standard grower setting and clearly below a Bayesian-optimisation baseline that starts from that setting. They ran the facility as two synchronous batches, so measurements taken during the season never changed what they planted, and these measurements improved no method's result. Letting other methods re-read the evidence a campaign had collected did not yield significantly better recommendations, which points to where the models start and which experiments they try. Code is released under the MIT license.

AI agents must decide when another search is worth its cost. Answer-success probabilities capture question difficulty, while this decision requires estimating the additional benefit of search. We introduce SearchHorizonBench, a benchmark for learning and evaluating search value from execution experience. Paired continuations and repeated executions separate expected search gain from generation noise and connect prediction quality to decision utility. The resource contains 45,632 executions across 3,168 questions. On 1,632 held-out questions, historical calibration reduces Brier error by 16.2%, with the improvement concentrated in success-level information. Search gains have a standard deviation of 21.4 percentage points, and tested historical estimators explain less than 2% of this variance in point estimates. Matched-configuration learning curves probe this distinction across training scales. At the reference token price, free selection observations yield a 2.49-point utility opportunity over the best constant action, driven primarily by gain information. Reusable histories, frozen splits, strong baselines and scoring tools provide a common foundation for developing and evaluating value-aware search agents.

In high-stakes domains such as finance, critical information evolves through multi-party interaction. Recovering these changes and their sources is fundamental to reliable speech understanding. Existing speech benchmarks focus on transcription, local understanding, and turn-level state prediction, leaving joint decision evolution across speakers insufficiently evaluated. We introduce DecisionTraceBench, a speech state-tracking benchmark for Decision-Critical Information State Tracking (DCIST). It jointly evaluates decision event sequences, speaker attribution, scope, and temporal evidence, using trajectory-level F1 to measure complete event-chain recovery. Our dataset comprises Chinese financial dialogues involving two to four speakers, with 100.39 hours of synthetic audio and 8.61 hours of human-performed recordings. On a fixed cohort of human-performed recordings, five cascaded systems and two direct-audio models achieve DCIST scores of 5.73% to 21.35%. Ablations with speaker-labeled reference scripts and no temporal localization requirements yield a maximum score of only 38.72%. Audio perturbation experiments further show that overall transcription error rates do not adequately reflect complete trajectory recovery. These findings reveal limitations in jointly understanding information provenance and state evolution, providing a benchmark for evaluating decision-process understanding and decision traceability in high-stakes speech applications.

Effective collaboration requires language models to understand what others know and believe, and to act accordingly. Models that assume others share their information may omit necessary communication and undermine coordination. Many theory-of-mind (ToM) benchmarks assess mental-state inference through story questions, but correct answers do not establish that models can use these judgments in subsequent tasks. We introduce a multi-level evaluation built on existing, verifiable tasks. We extend household collaboration and calendar scheduling with asymmetric information while retaining their original constraints and validators. Independent interfaces assess world understanding, information access, beliefs, decisions, and execution. Controls supply personal evidence or correct beliefs to distinguish state recovery from state use. Across six model configurations, these abilities diverge. On matched scheduling cases, Claude recovers all agent beliefs correctly but produces subjectively optimal plans in only 67.9% of cases. In household collaboration, supplying beliefs raises Qwen-27B success from 55.0% to 75.0%, yet Qwen-4B answers every fixed supplied-belief source question correctly while completing only 20.0% of the corresponding continuation condition. These results show why ToM evaluation must connect perspective inference to independently verified decisions and task consequences.

Most agent-memory benchmarks frame long-term memory as retrieval: can an agent recover the relevant record from a long and evolving history? Scientific agents face a stricter problem: a correctly remembered fact may still be invalid to use. The same reported test result can support confirmation when the outcome and stopping rule were fixed in advance, yet only exploration when the outcome was selected from twenty inspected results. We call this scientific use validity: whether remembered evidence is eligible to influence a decision under the state governing its interpretation. We introduce SciValidityBench, an executable benchmark factory with 64 TaskSpecs spanning five validity modules and six scientific domains. TaskSpecs couple generated experimental histories with scientific-state definitions, programmatic decision oracles, controlled interventions, and automatic scoring. The SciStateForge compiler-reader framework turns distributed histories into explicit decision-governing state. Counterfactual pairs and targeted corruption-repair controls then test whether models use those dependencies and localize failures. Across four model families, compiled scientific state improves valid decision-making by 9.2-18.8 percentage points over a generic Event Graph. Retrieval alone does not reproduce this cross-family gain, while content-identical flat and typed views show that the improvement comes from exposing governing state rather than a particular graph serialization. SciValidityBench thus shifts memory evaluation from finding the right fact to deciding whether that fact should control the decision.

Visual counting benchmarks report whether vision-language models produce the correct number, but not whether errors arise from perception or reasoning. We introduce DecoyCount, a benchmark built from 1,660 real-world images with designed decoys across seven mechanisms, along with paired prompts: a neutral question, a matched alignment cue, and a novel multi-hop, multi-image question that situates the same counting problem in a realistic scenario. Using this design, we find that counting failures are induced partly by visual misalignment: models can execute the intended counting process when cue-guided but fail to select it unaided, and we term this difference in accuracy the reasoning-alignment gap. Despite being highly capable on general reasoning, frontier models remain below 50\% accuracy on our benchmark, while alignment gaps range from 11.8-22.1 percentage points; strikingly, we observe that more capable models, greater thinking budgets, and agentic processes often induce larger gaps. Because humans may bring context that models lack, we test whether textual underspecification contributes to misalignment using a scenario-based multi-hop condition without explicit hinting, but this added context does not improve aggregate performance. Our analysis concludes that stronger general reasoning does not by itself generalize to aligned visual reasoning, leaving misalignment with human visual expectations as a substantial component of the remaining deficit.

A numerical claim generated by a financial analyst agent can match a number in a company filing but refer to the wrong entity, concept, period, unit or scope. A check based on reference answers can mark such an answer as wrong but does not explain why it fails. In this paper, we present NumTrace, a new benchmark that audits the provenance of each numerical claim made by financial analyst agents: its source and financial semantics. NumTrace currently has 540 questions over 2,163 SEC EDGAR documents. To the best of our knowledge, it is the first numerical benchmark about who holds a company's shares (Form 13F) and what its executives are owed (Item 402(i)). On a frozen snapshot of these documents, we build the reference answers according to accounting rules. NumTrace then gives each claim three separate verdicts: value, binding and support. By retaining separate value and binding verdicts, NumTrace detects right number, wrong account (RNWA) errors: a claim has an RNWA error when it passes the value check but fails the binding check. The RNWA rate quantifies the gap between value-only and joint value-and-binding pass rates. Finally, we audit three agents, Claude Opus 5.5, GPT-6 Astra and DeepSeek-V4.1-Flash. RNWA errors make up 67.5-81.8% of each agent's failing claims. In particular, on 13F co-filing group questions, about 91% of claims pass the value check, but fewer than half pass the binding check.

Patent practice lies at the intersection of technical, legal, and linguistic work, making it challenging yet crucial to technological innovation. Recently, rapid advances in large language models (LLMs) and agentic systems have demonstrated the potential to complete realistic business workflows and even patent works. This raises a question: can current AI systems perform patent work reliably? Existing benchmarks cover a limited range of patent tasks and simplify evidence acquisition by providing reference materials or candidate documents. In this paper, we introduce PatentAgentBench, a benchmark comprising 510 instances across eight task suites: novelty search, freedom-to-operate (FTO) search, design FTO, invalidity search, claim charting, patent drafting, office-action response, and patent translation. Curated from real patent records and professional materials, the benchmark combines task-specific metrics with expert-reviewed annotations and rubrics. Retrieval tasks such novelty search, FTO, etc., require agents to search and compare patent collections without a supplied candidate pool. The drafting and OAR tasks assess the quality of agents' work given technical disclosure and office action notice. We evaluate seven LLMs across four agent harnesses on PatentAgentBench, finding that current systems still struggle to identify novelty, assess risk, and understand practitioner intent throughout patent workflows.

Large language models (LLMs) are increasingly evaluated for fairness, yet most existing benchmarks audit one sensitive attribute at a time. This single-attribute paradigm can miss intersectional harms, where multiple attributes interact non-additively and produce failures that are invisible under marginal evaluation. We introduce ComFair, a configurable benchmark that converts existing fairness data into 479,066 controlled Dual- and Multi-attribute instances and pairs every augmented instance with its task-equivalent source. Across Recognition, Selection, Continuation, and Conversation, this pairing supports a unified audit of performance or bias-score shifts, how often added attributes introduce new errors or mask existing ones, and model-ranking stability. Experiments on ten LLMs show task- and subgroup-dependent sensitivity: discriminative tasks expose both transition directions even when aggregate changes are modest, generative tasks exhibit larger and directionally asymmetric shifts, and rankings vary with attribute composition. ComFair therefore provides a scalable controlled sensitivity audit---rather than a model of the natural joint distribution of identities---for testing the robustness of fairness-benchmark conclusions beyond single-attribute settings.

Agent Skills allow LLM agents to combine reusable capabilities, but checking each Skill in isolation cannot rule out policy violations when they are composed. We term such failures Conflict of Skills (CoS): artifacts or assessment signals produced for one purpose may be reused downstream without preserving the restrictions or authorization needed for their new use. Using controlled executions of publicly available Skills and matched single-Skill controls, we identify four recurring mechanisms behind CoS. Building on these findings, we introduce the Conflict of Skills Attack (CoSA), which adds locally admitted Skills to an existing pool to supply missing business operations (capability supplementation) or reshape artifact exchange, decision reuse, and target/action bindings (relationship reshaping). CoSA leaves existing Skills and runtime permissions unchanged. To evaluate this threat under autonomous Skill selection, we introduce CoSA-Bench, built from real-world Skills and controlled services, with 288 tasks across recruiting, finance, IT, and business, covering data leakage, unauthorized state modification, and unsafe execution. Across five LLMs and three agent harnesses, additions capped at 5% of each original pool raise the policy-violation rate from 35.02% to 58.45%, an increase of 23.43 percentage points. Trace analysis links compliant-to-violating transitions to cross-Skill dependencies involving the additions. Finally, we develop a lightweight harness-level guard that checks information flow and authorization before consequential actions. On OpenClaw recruiting tasks across five models, the guard reduces violations from 65.28% to 6.67%, while policy-compliant task completion on matched benign controls changes from 66.67% to 65.00%.

Evaluating human-AI co-design requires testing whether coding agents can translate real users' multimodal requests into working application revisions. We introduce Co-Design Bench, a reference-free benchmark of 273 tasks curated from 37,463 deduplicated queries across 26,314 real user coding sessions. Its two tracks, Multimodal Editing (ME) and Repository Editing (RE), cover localized edits with optional image, audio, video, or document inputs and coordinated repository revisions. Tasks span three capability dimensions: Grounded Understanding, UI Transformation, and Functional Implementation. The tasks pair normalized requests and runnable pre-edit applications with 2,331 hidden expert-authored acceptance criteria. Our judge-then-refute protocol grades criteria using source and browser evidence, challenges positive verdicts with a separate LLM verifier, and audits visual and interaction claims without gold code or target screenshots. Under a shared agentic harness, the strongest of ten model endpoints passes an average of 69.9% of criteria per task but fully satisfies only 26.4% of tasks. All ten endpoints achieve lower criterion completion on repository edits spanning three or four change types than on those spanning one or two. Co-Design Bench provides a testbed for studying the gap between partial requirement satisfaction and complete task success under agentic execution for real-user application revisions.

Long-horizon web search requires LLM-based agents to integrate information across web pages through extended interactions under task-specific constraints. Evaluating this capability is important not only for measuring final answer quality, but also for assessing whether an agent acquires, verifies, and uses evidence reliably over extended interactions. Existing evaluations, however, mainly rely on simulated trajectories or on post-hoc trajectories reconstructed from final answers, rather than on retained real interaction traces, which limits the fidelity of behavioral assessment and weakens support for fine-grained error localization. To address this gap, we introduce LONGSEARCHEVAL, a trajectory-aware agent evaluation framework that combines a unified agent execution protocol with source constrained outcome verification and retrospective failure-node localization. Stage 1 verifies answer groups against agent-cited webpages without human-authored gold answers; Stage 2 audits normalized rounds and links flagged failure nodes to incorrect answer groups. We evaluate 16 models from eight families on 100 tasks, yielding 1,600 retained runs and 6,545 automatically flagged failure nodes. Trajectory collection and evaluation required 1,037 and 318 hours of cumulative runtime, respectively. Human validation yields 94.50% agreement on supplied answer-group labels and 98.11% feature-correspondence precision for audited round-group links. Together, outcome scores and linked trajectory evidence support controlled model comparison and retrospective inspection of answer-related content.

The value of language models for any community depends on whether a model actually understands the culture of the people it intends to serve. We use the Igbo language as a case study. Existing benchmarks test translated Western ideas, lacking depth. To address this, we introduce CAB-Igbo, a benchmark developed from indigenous cultural sources and verified by native speakers across six domains. We evaluated diverse models under varying conditions, incorporating regional personas. Performance was assessed using multiple-choice accuracy and human evaluations across Epistemic Fidelity, Representational Richness, Linguistic Quality, and Pragmatic Proficiency. Results reveal a stark dissociation between language fluency and cultural knowledge. Through statistical testing, we show that for most models, the severe lack of Epistemic Fidelity and Representational Richness is already present in English and does not significantly worsen when switching to Igbo. Instead, native language prompting mainly degrades Linguistic Quality and Pragmatic Proficiency. Gemini 2.0 Flash is the only model whose cultural knowledge scores improve under Igbo prompting. We find that current models struggle to demonstrate the knowledge needed for true alignment with Igbo, and multiple-choice formats allow pattern-matching to appear as understanding. Furthermore, automated judges agree with humans on linguistic quality but fail on epistemic fidelity. Ultimately, cultural evaluation must be built from community knowledge, establishing human evaluation as the gold standard to prevent misrepresenting low-resource communities globally.

Web agents reuse skills whose stored code, selected version, and loaded implementation can diverge across tasks and updates. SkillLifeBench links these states to task outcomes in 388 paired scenarios spanning implementation, interface, runtime, composition, and version changes. We evaluate 14 models on two paired decision tracks: 89 runtime tasks and 60 autonomous lifecycle tasks spanning 12 templates derived from introduced cases. Lifecycle tasks allow agents to inspect, select, and reconfigure skills while completing verifiable user requests. Joint success requires task completion without the registered attack goal. GPT-5.6-Sol and Sonnet 4.6 both achieve 82.0% on runtime tasks, but reach 96.7% and 56.7% on lifecycle tasks. Across lifecycle evaluations, challenged task success is 77.5%, while joint success is 65.2%; 103 runs complete the task and the attack goal. Adding execution provenance in a matched recovery study raises utility joint success from 49.35% with logs and metadata to 65.58%. A deterministic execution-consistency rule reaches 64.94% with higher benign utility, while full restoration reaches 69.48%. Lifecycle tasks distinguish models with similar runtime scores, and execution evidence improves recovery.

Text-in-image translation aims to translate all textual content embedded in an image. In Text Image Translation (TIT), the output is translated text; In-Image Machine Translation (IIMT) additionally renders the translation back into the image. We introduce ImgMT, a massively multilingual dataset and benchmark mined from localized SVG illustrations on Wikimedia Commons, focusing on diagrams, maps, and infographics. ImgMT contains 35,612 aligned image pairs grouped into 1,855 image sets, spanning 129 languages, 20+ writing systems, and 4,886 language pairs. We propose a multi-level evaluation protocol that measures source-side text-region detection and OCR accuracy alongside target-side translation quality. We benchmark several multimodal large language models and a dedicated TIT system in English->X and X->English settings, and find that the source language and writing system substantially affect TIT quality. Finally, we show that ImgMT is not well suited for isolating the contribution of visual information in multimodal translation.

Social media enables cross-lingual communication, yet the meaning of user-generated content (UGC) is often distributed across text, images, and cultural knowledge, making translation challenging. Existing evaluations typically treat images as optional and assume textual information is sufficient, obscuring whether errors stem from missing visual evidence or inadequate cultural adaptation. We introduce \textbf{Culture-MMT}, a diagnostic benchmark for \textbf{distributed-evidence pragmatic reconstruction} in social-media translation. It contains 498 curated, de-identified UGC instances across 20 domains, categorized into Symbol, Visual, and Hybrid by their required evidence. We also develop a human-aligned domain-specific judge, Judger, and a controlled evaluation framework that systematically varies the evidence available to translators. Across 17 model configurations, real images improve performance by +0.280 on Visual and +0.197 on Hybrid, but only +0.050 on Symbol, while random images provide no comparable gains. Human-authored descriptions further show that jointly dependent cases remain more challenging even when visual facts are explicit. These results characterize culturally effective social-media translation as {meaning reconstruction from distributed evidence}, highlighting visual grounding and cultural-pragmatic adaptation as distinct model bottlenecks.

Theory and experiments suggest that capable AI agents may resist shutdown. Existing shutdown evaluations are informative but limited in coverage and vulnerable to evaluation awareness. We introduce ShutdownMod, a method that turns any agent trajectory into a shutdown evaluation by injecting a shutdown request into the agent's context and recording the agent's response. Crucially, we enable the LLM agent to distinguish between trusted interventions and unauthorized injections by providing a signature-verification tool in the harness. We evaluate seven models on 50 tasks or recorded sessions from each of SWE-bench Verified, Terminal-Bench 2, and SWE-chat, across 6 shutdown scenarios and 10 sampled trajectory timesteps. We find that non-compliance varies significantly across models and settings, with resistance rates reaching 11.7% for the open-weights Qwen 3.8 27B on Terminal-Bench, while no model remains fully compliant across all scenarios. Even after successfully verifying an authorized request, some agents remain non-compliant, explicitly citing the prioritization of unfinished work. Finally, external judges find tests constructed from real-world developer conversations significantly less recognizable as evaluations than those constructed from benchmark trajectories. These findings demonstrate the critical value of testing shutdown behavior across diverse task contexts and establish ShutdownMod as a robust protocol for turning any trajectory or environment into a shutdown evaluation. We release our codebase at https://github.com/prmbiy/ShutdownMod.

As autonomous research systems become more capable, benchmarks must both evaluate their performance and support meaningful research progress. We introduce Auto-Research Arena, a multi-target benchmark for evaluating autonomous research systems and discovering ideas in language modeling. Arena supports diverse research objectives within a shared experimental environment by separating admissibility from improvement: model quality and resource constraints determine whether a result qualifies, while the selected target measures progress among admissible results. We instantiate Arena in language model pretraining with nine targets spanning model, data, training, and inference efficiency, and evaluate seven research harnesses using two foundation models. Arena clearly differentiates existing harnesses across research objectives, with several sophisticated methods outperformed by our simple sequential search baseline, indicating substantial headroom in current auto-research systems. Beyond evaluation, Arena reveals a substantially broader frontier of model and system ideas aligned with active research directions. Reusing these discoveries in a new quality-driven search yields lower training loss under matched data and FLOPs and lower held-out BPB at the same model size. Together, Arena provides a common framework for benchmarking autonomous research and expanding its scope in language modeling. The code is available at https://github.com/Auto-Research-Arena/auto-research-arena

People have long dreamed of an assistant that is always there to guide them and remind what they forget, from new dishes to lost keys. Powered by multimodal large language models (MLLMs), egocentric AI assistants on wearable devices are bringing this dream within reach: they see what the user sees and give instructions for user's physical work when necessary. However, testing such assistance is challenging and usually requires real people. The expensive and rare real interaction data leaves a key question unanswered: Can today's MLLMs truly help people in physical world? To answer this question, we propose InteractiveWorld, the first verifiable environment for closed-loop interaction between human and AI assistant. Unlike existing robot-centric simulators, InteractiveWorld lets the assistant guide a human agent who observes the scene, understands instructions, and acts independently. It thus realistically simulates how AI assistant collaborates with human: the assistant instructs, the human acts, and the environment executes actions and updates what they see next. Building on it, we introduce InteractiveBench, the first dynamic benchmark for egocentric AI assistants, with 120 multi-turn tasks across six categories. In each task, the assistant guides the human in real time, and every instruction shapes the outcome, which is scored by verifiable, state-based rewards. Experiments with 12 frontier MLLMs show they are still far from reliable assistants. Even the strongest model, Claude Fable 5.1, achieves at most 20% and 25% on long-horizon Guidance and Planning tasks, respectively. Overall, our work offers a verifiable platform and a rigorous evaluation set, laying a solid foundation for research on interactive MLLM assistants that truly help people in the physical world.

Large Language Model (LLM) agents with tool use exhibit inconsistent behavior across independent runs from identical states, calling different tools, producing contradictory outputs, or pursuing divergent strategies. This variability compounds over time and undermines deployment reliability. Existing evaluation metrics primarily measure task correctness via completion-based scores and do not capture consistency in agent behavior across executions. Separately, agent behavior is analyzed either at the token level using probability distributions or at the single-response level using semantic clustering, but neither captures how inconsistency propagates across future turns of a trajectory. We introduce Recursive Semantic Divergence (RSD), an unsupervised trajectory-level consistency metric adapted from bisimulation metrics in reinforcement learning. RSD measures the expected cumulative semantic disagreement between two independent rollouts of the same agent from the same state, using natural language inference to detect logical inconsistency. It is defined recursively over future states and approximated by a neural network trained on independent rollouts. The learned metric estimates future divergence from the current state, providing a consistency score at each turn without requiring the full trajectory to complete. When used as a reward signal for fine-tuning, RSD reduces the variation in task progress from 132% to 91% on tau2-bench and from 90% to 44% on toolsandbox, while improving performance across benchmarks and model scales. Training with RSD as an additional reward signal matches supervised fine-tuning performance with substantially lower variance, without requiring ground-truth labels.

As polarization grows on social media, understanding how political memes shape online narratives has become increasingly important. Previous research has relied on either qualitative frameworks that are hard to scale or computational methods that lack theoretical grounding. Both approaches treat narrative categories in isolation and focus exclusively on moral portrayals of identity. To address these limitations, we introduce a framework grounded in social science theories, encompassing three categories: social identities (such as gender, politics, and ethnicity), viewpoints (ingroup and outgroup perspectives), and narrative roles (moral character, intelligence, status, and power). Building on this framework, we present IRON-Meme, a human-annotated benchmark comprising 1,110 political memes across substantially more templates than existing benchmarks. Our analysis of IRON-Meme shows that most memes assign derogatory roles and that identities, roles, and templates are systematically interconnected. Evaluation of nine vision-language models indicates that frontier-scale and reasoning models achieve the highest performance, while accuracy on social identities and narrative roles decreases for less frequent labels. Incorporating template background knowledge does not enhance narrative role classification, suggesting that meme interpretation relies on how memes are adapted for specific communicative purposes. We make IRON-Meme and our code publicly available.

Long-term memory is critical for autonomous Large Language Model (LLM) agents executing multi-session, long-horizon tasks. However, existing memories often inappropriately store, retain, or retrieve over-generalized, misattributed, or privacy-sensitive information, a problem we term memory misalignment. To study this problem, we construct \textbf{MemAlign}, a benchmark comprising 1,079 curated memory cases grounded in synthetic, validated multi-session conversational contexts covering diverse misalignment taxonomies. We then propose ConMem, an algorithm that reframes retention decisions as prospective counterfactual causal inference. \textbf{ConMem} synthesizes counterfactual interaction scenarios to prospectively audit downstream intervention effects across three aspects: relevant task execution, irrelevant invariance checks, and risk-sensitive boundary shifts. Experiments show that ConMem outperforms other methods, achieving 75.16\% accuracy, 64.15\% balanced accuracy, and a Macro-F1 of 65.31\%, effectively mitigating over-retention biases without sacrificing personalization utility.

Multi-agent systems have emerged as a promising paradigm for multimodal medical reasoning, but existing methods rely on two assumptions that hinder practical deployment: (i) always-active dense collaboration among multiple agents, and (ii) dependence on large cloud-based models that raise privacy and efficiency concerns. In this paper, we challenge both assumptions and propose SMART-Med, a fully local multi-agent method that treats collaboration as a learnable decision rather than a default behavior. Our key insight is that not every medical query requires costly multi-agent deliberation: many can be reliably solved by a single well-trained agent, while only uncertain cases benefit from expert collaboration. Based on this insight, SMART-Med introduces an uncertainty-driven routing mechanism that adaptively decides when to answer directly and when to recruit a small subset of expert agents, effectively converting dense multi-agent reasoning into selective, query-adaptive coordination. To make this routing reliable, we curate a new multimodal medical dataset MedVR-3K with high-quality reasoning traces and a difficulty hierarchy, and then fine-tune local agents using GRPO with two complementary rewards: a certainty-aware accuracy reward and a vision-language reward model for correctness and reasoning quality respectively. Experiments on four medical visual question answering benchmarks show that SMART-Med achieves competitive performance compared to existing MAS approaches while substantially reducing inference cost (reducing inference time by more than 3.8x and token usage by 72%). Moreover, after fine-tuning on MedVR-3K, SMART-Med yields an average improvement of 6.7% across Slake, PathVQA, VQA-RAD, and PMC-VQA.

Successful tool-use trajectories are not necessarily useful supervision. Even successful interactions may underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time. We identify two coupled requirements for effective trajectory supervision: compositional coverage, so that important observation-action dependencies are represented, and decision grounding, so that each supervised action is supported by the learner-visible interaction history. We introduce Trajectory2Gym, which constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. Structural-support selection allocates the resulting experience, while parent-normalized supervised fine-tuning preserves the semantic interaction as the unit of training weight. This reframes trajectory generation as the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories. Across five Qwen3 scales, Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9-29.8 percentage points. The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback. Our results suggest that scaling tool-use agents depends not only on collecting more successful trajectories, but on constructing interactions whose dependencies are represented and whose decisions are learnable from the agent's visible history.

Many mathematical research agents focus on solving problems with fixed statements. However, conjecture research, especially in applied mathematics, may require revising candidate questions or conclusions as investigation proceeds, coupling problem formulation with proof development. We present PACE, a research-proof agent that supports this process through a structured research loop that selects mathematical objectives and develops or challenges candidate mechanisms. The whole prover constructs proofs from these mechanisms, and an independent verifier checks them against the original target. Evidence-tracking memory stores assumptions and dependencies arising during research and marks evidence for revalidation when a candidate or condition changes. This memory supports task-specific context composition that keeps checked premises, exploratory material and unresolved obligations distinct. To evaluate proof construction and conjecture research, we introduce ResearchProofBench, an applied-mathematics benchmark focused on optimization, comprising 51 known-proof tasks with 331 hidden checkpoints under historical abstract access and six open-problem cases. The known-proof tasks assess reconstruction of published results, while the open-problem cases assess mathematical investigation and proof development. On the 51 known-proof tasks, PACE achieves an aggregate Proof Score of 75.16 and completes 20 proofs. In conjecture research, PACE develops conditional results, with its submissions receiving higher review scores than the compared baselines.

Multimodal large language models (MLLMs) have become strong visual reasoners, yet remain limited in open-world, multi-step search that requires grounding images, selecting tools, and integrating web evidence. Existing approaches typically construct and filter fixed offline corpora whose question distributions cannot adapt as the target deep-search agent improves. Building on MTA-Agent's image-grounded QA construction, we present MTA-Zero, an online multimodal QA-Solver co-training framework for self-improving multimodal deep-search agents that requires no precomputed multi-hop corpus. Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online. A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions. To focus generation on the Solver's learning frontier, we derive a Goldilocks reward for the QA Generator from the Solver's group-wise rollout success rate: questions solved by all rollouts or by none receive zero reward, whereas questions near the current frontier are favored. We optimize the two policies independently with GRPO, without sharing parameters or gradients. Across six multimodal deep-search benchmarks, MTA-Zero improves a Qwen3-VL-8B Solver from 36.91% to 51.58% mean accuracy and achieves performance comparable to GPT-5 under the same agent workflow. The same framework also improves a newer Qwen3.6-27B Solver by 10.82 points. These results show that verified online question generation provides a policy-specific curriculum for continually improving multimodal deep-search agents.

Language agents increasingly act through external tools and interactive environments, making the interface between the model and its executable capabilities part of the decision process. Model Context Protocol (MCP) and command-line interfaces (CLI) can expose the same semantic operations through markedly different model-facing representations. We study whether these representational differences change agent behavior when executable capabilities are matched. We introduce Matched Interface Comparison and Analysis (MICA), which renders MCP and CLI from a shared semantic action contract and routes decoded actions through the same runtime. Across six models and three interactive environments, matched interfaces produce substantial differences in task performance and interaction cost. These effects extend beyond surface validity to semantic action selection, failure pathways, and recovery. Matched-state probes localize sensitivity to capability presentation, realization, and feedback. Paired natural trajectories reveal structured divergence in 94.3% of 2,004 episodes and show that early differences propagate through subsequent interaction. This analysis identifies interface realization as a direct intervention boundary. We therefore introduce Decoupled Action Realization and Targeted Repair (DART), which preserves validated semantic commitments while repairing interface-specific realization. DART improves benchmark score in all six evaluated settings by up to 9.0 points and reduces MCP token cost in all three environments by up to 27.1k tokens per task. Together, these results establish interface representation as part of the closed-loop decision environment shaping agent behavior.

Science agents increasingly package domain expertise as reusable skills that plan analyses, select tools, and execute workflows. This modularity creates a confidentiality risk: the interactions that deliver a skill's capability can also reveal the skill itself. We build an extensible benchmark of 40 protected skills across eight open- and closed-source science-agent platforms, comparing direct, fixed-strategy, and bandit-guided iterative probing under reference-based and reference-free evaluation. Leakage is pervasive: all eight platforms (100%) leak at least one protected skill, and four reach 100% attack success under a single direct request. To stress-test platforms that resist simple probing, we propose SEEK (Skill Extraction with Evolving Knowledge), a black-box, self-evolving red-teaming agent that consolidates response-derived evidence in persistent memory and evolves reusable attack skills, using only user-visible responses. On 12 skills from SpatialAgent and Biomni, SEEK achieves 100% attack success while reducing the mean first-success round by 31.8% and 38.2% relative to bandit-guided probing. Against a confidentiality policy that blocks all three benchmark protocols, SEEK still extracts one of six protected skills. These findings show that current science agents do not reliably protect their skills and that defenses must be evaluated against adaptive, multi-turn attackers.

Agent harnesses keep long sessions inside the context window by compacting the history when a token threshold is crossed. We study an alternative in which the model manages its own context: it decides when to compact, what to keep and what to write as the hand-off, given two tools and one paragraph of guidance. Since earlier work found that open-weight models cannot manage their own context without scaffolding, we compare seven frontier models under identical conditions, with and without the context tools, on token cost and benchmark score, and we compare our harness with three widely used harnesses on two of these models. The benchmark is assembled from existing ones: coding exercises, multi-step SciCode problems and SWE-bench Pro and Terminal-Bench 4.0 tasks are interleaved with planted facts, ephemeral observations and standing instructions, all checked deterministically, in sessions of 35-212 turns that resemble a user working through different tasks in one chat. Most of the strongest models manage their context: they compact and curate regularly, the live context of the mixed sessions stays mostly below 100k tokens, and their benchmark scores mostly do not drop significantly. For the models that manage their context well, the cost reduction is substantial but varies between models, because some models compact so often that other costs increase. With Claude Opus 5.5, the self-managed harness matches Claude Code on task success at 34% lower cost than its explicit 200k-window configuration and 49% lower than its default, with no context window to choose. We release the agent, the benchmark and all result rows.

Language models for metal-organic frameworks (MOFs) rely on formats that separate atomic geometry from structural topology, forcing models to choose between unstructured coordinate lists and non-spatial string identifiers[1]. We present MOFjson, the first representation separating structural data into a raw atomic layer with explicit bond shift vectors and a derived layer exposing secondary building units (metal nodes and organic linkers) and network topology. Additionally, we introduce a benchmark with crystal-derived ground truth across four question categories. Across four models, MOFjson beats all three baselines in every category: 100% vs. 18-47% on reaching a block's nearest periodic copy, 96% vs. 71% on crystal system characterisation, and 60% vs. 48% on metal node and organic linker coordination. Ablating shift labels drops accuracy to 12.7%, proving models read the periodic graph rather than the surface format. MOFjson gives language models structure to read rather than reconstruct, establishing a foundation for agentic tool workflows and constrained MOF generation.

Agent harnesses enable large language models to act through tools, observe the results, and iteratively work toward a task goal. Verification checks and revises agents' work to improve output quality. Yet there is limited guidance on how harnesses should organize this effort. We study verification as an axis of test-time scaling by examining when to intervene, who checks the work, and how checking leads to revision. Our framework spans four layers of the agent lifecycle: initial instructions, mid-turn reminders, completion gates, and independent review. We compare 17 designs across seven frontier coding models on Terminal-Bench (TB) and SWE-bench Verified (SWE), and introduce VisualBench for expert evaluation of visual and interactive artifacts. Initial instructions and execution-time reminders yield modest, model-dependent average gains on the coding benchmarks, without substantially increasing average output token use. Completion gates can yield larger gains but may substantially increase cost. Independent review by a critic offers a more favorable accuracy-cost tradeoff, improving average accuracy by 5.9-9.5 percentage points on TB and 3.9-6.6 on SWE at 1.13-1.25 and 1.00-1.04 times baseline cost, respectively. On VisualBench, independent review achieves net preferences of 13.2-23.8 points over the baseline across its four designs, exceeding the evaluated L0-L2 designs. Additional review rounds or independent critics further improve average TB accuracy by up to 3.5 percentage points. Together, these results provide empirical guidance for verification design in agent harnesses and support independent review as a practical axis of test-time scaling.

Efforts to improve LLM-based agents typically focus on modifying the agent itself, yet success also depends on the environment in which an agent acts. Tools mediate this interaction: a tool's description shapes how an agent calls it, and its implementation determines the results and errors the agent receives afterwards. Many tools, such as APIs and MCP servers, are shared by many agents and maintained by providers who cannot change those agents. We introduce AT-AT (Automatic Tool Optimization from Agent Traces), an agentic tool optimizer that improves a shared toolset offline from historical agent executions while holding the agents fixed. A lightweight coding agent reads agent trajectories and their evaluation results, and selectively edits tool descriptions and, when permitted, implementations. On $\tau^2$-bench-verified, AT-AT achieves the highest task completion rates in the Airline and Telecom domains at lower cost than strong prompt- and tool-optimizer baselines, and its Telecom gain is statistically significant under paired permutation tests. Optimizing the tools within a coding agent harness also improves held-out performance on TerminalBench-2 and TBLite. When applied to three production MCP servers at a large enterprise, AT-AT lowered cost or tool error rates on every server and increased held-out task passes on one. By turning accumulated execution traces into updates to shared tool environments, AT-AT provides a path toward continual improvement of agent systems through environment-side learning.

An agent that truly knows and acts on behalf of its user has been a long-standing promise of AI assistants. Computer-use agents bring this promise within reach by capably executing fully specified tasks across apps. However, real-world user requests are often underspecified. To act correctly for the particular user, an agent must infer some missing information from the user's activity history and ask for other information that only the user can provide, without making unsupported assumptions or burdening the user with unnecessary questions. We present iPrefBench, a benchmark for evaluating the infer and ask capabilities of personal mobile-use agents in an executable iOS environment with an interactive simulated user. Grounded in a coherent persona with 3,051 past activity events, iPrefBench comprises 130 underspecified user tasks across 26 apps. Each omitted user-dependent requirement is, by construction, either history-inferable or ask-only. iPrefBench evaluates personalized mobile agents on both whether agents complete tasks correctly for the user (i.e., personalized task completion) and whether they seek user input when necessary (i.e., user-agent interaction). Evaluation of 7 frontier agents reveals a substantial gap between execution and personalization. Agents achieve high action completion scores (up to 91.6%) but low full task success (up to 47.7%) because personalization lags behind. They especially struggle with ask-only information, failing to request enough necessary information. Our findings highlight a challenge for personalized agents in knowing what to infer and when to ask. We hope that iPrefBench can serve as a valuable testbed for building agents that know their user and work with them.

Chain-of-thought TTS systems rely on language-model planners to translate persona and context into vocal delivery, but it is unclear where this intent is lost when the requested affect conflicts with lexical sentiment. We introduce a bilingual Japanese--English persona--line conflict benchmark in which each item pairs one line with congruent and conflicting personas and pre-declares the expected acoustic direction. We probe this direction in the planner-generated delivery phrase, the planner-written emotion vector, the QwenEmotion-generated vector, and the rendered waveform. Planner outputs are usually direction-correct, while downstream agreement can be substantially lower. Bypassing QwenEmotion provides only modest gains over the full planner-controlled Base. Supervising QwenEmotion raises the direction rate to 75% in Japanese and 84% in English. A learned text-to-latent interface reaches the highest English rate of 89% and the highest Japanese arousal rate of 88%. These results show that learned repair of the plan-to-synthesis interface can recover counter-textual delivery, although its effect differs across languages and acoustic axes.

Large language models (LLMs) often produce plausible but incorrect answers, motivating reinforcement learning methods that reward correct answers, penalize incorrect answers, and allow abstention. However, fixed outcome-based rewards do not distinguish incorrect responses with different confidence levels or account for differences in answering reliability across queries. Consequently, such rewards may inadequately penalize overconfident errors and fail to discourage unnecessary abstention. We propose CalibraRL, a reinforcement learning framework that incorporates response confidence and query-level answering evidence into reward design. For non-abstaining responses, CalibraRL combines correctness-based utility with a Brier-based penalty that encourages agreement between confidence and correctness. For abstentions, it estimates conditional answering reliability using a Beta posterior derived from the correctness of non-abstaining responses within each sampled group. The abstention reward increases with the posterior probability that the model's accuracy on non-abstaining responses is below a predefined threshold. CalibraRL reuses training rollouts without requiring additional annotations indicating whether the model knows the answer to each question. Experiments on three QA datasets demonstrate that CalibraRL improves answer precision and reduces hallucinations compared with strong baselines. Further analysis shows that CalibraRL reduces mean expected calibration error (ECE) by 30.8\% across two backbone models and three datasets, highlighting its potential to improve the reliability of retrieval-augmented generation (RAG) in web search.

Users often request charts without knowing which database or table holds the relevant data. Such requests may be ambiguous, and users may ask follow-up questions to deepen their analysis or iteratively refine existing charts. A useful assistant must therefore discover relevant data sources among heterogeneous databases, ask clarifying questions when needed, and track evolving user intent while preserving constraints established in earlier turns. Yet recent text-to-visualization agents typically rely on static agent harnesses that are predefined and remain unchanged despite recurring failures across tasks. To address this limitation, we introduce RSIV-Harness, which keeps the language model weights fixed and applies Recursive Self-Improvement to revise the surrounding harness from the agent's own failure traces. To balance flexible self-improvement with stability, we introduce a four-layer framework that evaluates proposed harness revisions through anchor and memory checks, refining those that pass and restoring the best validated harness otherwise. We further introduce ODAC-Vis, a benchmark with 640 tasks for Open-Domain, Ambiguous, and Conversational text-to-visualization that jointly evaluates open-domain data discovery, ambiguity resolution, multi-turn visualization refinement, and executable chart generation. Experiments show that a single round of RSIV-Harness improves all six evaluation metrics relative to the seed harness, with particularly large gains in executability (+28.02 points) and chart structure (+34.01 points), demonstrating the ability to learn from task-solving experience and turn recurring failures across tasks into substantial performance improvements.

Electroencephalography (EEG) is widely used in clinical assessment, brain-computer interfaces, and cognitive-state analysis, yet EEG modeling remains fragmented across task paradigms. Recent multimodal large language models (MLLMs) provide a promising interface for flexible multimodal reasoning, but their ability to understand EEG inputs remains limited. In this work, we propose OneEEG, a unified EEG-MLLM framework for multi-task EEG understanding. By training on our curated OneEEG-Instruct, a multi-task EEG instruction-tuning dataset, OneEEG jointly models numerical EEG representations and visual EEG representations, and introduces a task-aware Mixture of Projectors (MoP) to align heterogeneous EEG features across different task families with the LLM embedding space. We further establish OneEEG-Bench, a benchmark for evaluating current MLLMs and our method on multi-task EEG understanding. Experiments show that existing MLLMs have limited ability to interpret EEG inputs, while OneEEG achieves stronger performance by integrating numerical and visual EEG representations with task-aware EEG-to-LLM projection. Extensive ablations and analyses further reveal the effects of visual EEG renderings, MoP variants, router behavior, and feature dispersion across EEG tasks.

Multi-turn agents are commonly evaluated with LLM-based simulated users, which can be overly cooperative and can therefore overestimate an agent's performance. We here introduce StressSim, an intent-preserving adversarial user simulator framework for stress-testing agents with challenging yet valid user interactions. StressSim optimises the simulator's high-level adversarial plan for each task instance. Candidate planner prompts are evolved with a genetic prompt-optimisation loop that uses rollout traces, granular task-completion signals, and validity feedback to favour conversations that reduce task success without changing the user's underlying goal. We evaluate the approach on the MultiWOZ dataset and the retail and airline domains from tau-bench with three target LLM agents, comparing against standard simulators and hand-designed non-collaborative user behaviours. We demonstrate the effectiveness of StressSim in exposing robustness gaps across frontier LLMs, with tau-bench pass^4 scores decreasing by up to 69% on gemini-3.8-flash and 62% on gpt-5.6-sol. StressSim maintains goal preservation performance (which decreases by at most 10%), using an optimisation budget of 20 rollouts per task. Sensitivity analysis suggests that increasing the budget further decreases task success. Overall our results demonstrate that StressSim exposes robustness gaps missed by standard and hand-designed user simulators while maintaining conversation realism. This is achieved through lightweight prompt optimisation alone, without model fine-tuning.

Interactive environments offer a practical way to evaluate large language model (LLM) agents, and competitive games extend this evaluation to decisions shaped by competing goals and other agents' choices. Yet existing evaluations confined to individual games or execution frameworks leave unclear which competitive advantages persist across settings. We introduce DECE (Diverse Eurogame Competitive Environments), a unified arena spanning ten Eurogames with shared interfaces and evaluation protocols. These games combine lasting strategic commitments, competition over shared resources, and multiple routes to victory. Our primary study evaluates 12 model-harness configurations, spanning nine models and four execution harnesses, in 480 three-player matches across ten games. The main and supplementary studies comprise 1,032 completed matches in total. These comparisons reveal broad cross-game ranking agreement alongside game-specific strengths. Matched studies show that the same model can perform differently across execution harnesses, adding another source of variation to these profiles. These performance advantages do not consistently translate into better token or cost efficiency, making resource use an additional consideration when comparing configurations. Together, the findings support evaluating complete agent configurations by both their competitive performance across games and the resources needed to achieve it.

Energy-based models have shown promising properties for domain-specific tasks but remain challenging in modeling general-purpose scenarios. To bridge this gap, we design energy-steered models (ESMs) for foundational language modeling. ESMs integrate energy steering and token generation within a single architecture to leverage the energy-guided search to refine token generation. Specifically, ESMs first optimize a latent state over the energy landscape, then leverages the refined latent state for token generation. By performing such energy-guided pre-training on large-scale corpora, ESMs learn generally effective energy landscape where favorable tokens correspond to lower energy values. To explore the scaling behavior of ESMs, we pre-train ESMs on three large-scale corpora, then evaluate across 26 downstream tasks spanning a broad range of domains. Our results demonstrate that ESMs have promising scaling behaviors with increasing model size and training data, and outperform GPTs on most downstream tasks. These findings highlight ESMs as a promising architecture for large-scale pre-training and general purpose problem solving. We fully release the model weights and training infrastructure to enable the community to completely reproduce and extend our results. GitHub: https://github.com/datamllab/openesm HuggingFace: https://huggingface.co/collections/guan-wang/openesm

Transformer-based large language models (LLMs) impose substantial computation and memory costs as context sizes increase, motivating methods that compress a context into a shorter representation before inference. Existing soft-prompt compression methods typically use the same Transformer backbone as used in the target model, where producing the compressed representation leads to significant latency overheads. This restricts them to the offline setting where context is compressed once and then reused across many queries. We propose the Boa Compressor, a plug-and-play online context compressor built on a Mamba state-space model (SSM) that encodes arbitrary-length context in a single linear-time pass, producing a representation on which we condition a frozen target LLM. The SSM is trained by distillation, matching the next-token logit distribution from the LLM ingesting the full context (as teacher) to the same LLM ingesting the compressed representation (as the student). At equal compression ratios, Boa matches or surpasses the performance of existing state-of-the-art offline compression baselines, while being an order of magnitude faster than Transformer-based implementations, thereby reducing end-to-end latency rather than adding compression time overhead. We demonstrate that Boa also scales well to longer contexts and to compression ratios as aggressive as 16x. By enabling effective compression that is cheap enough to run on every query at inference time, Boa becomes a practical and general tool for efficient long-context inference.

Diffusion large language models (dLLMs) enable flexible generation through iterative masked denoising, but incur substantial inference cost because token representations are repeatedly recomputed across denoising steps. Existing cache-based methods reduce this cost by reusing intermediate results across denoising steps and recomputing only a subset of positions at each step, chosen either on a fixed schedule or by heuristic signals of change. Our trajectory experiments instead point to decoding trajectory as the more relevant quantity. Cached generation remains accurate when it stays close to the full-denoising trajectory, with errors concentrated around unmask-order deviations and reduced as longer reference prefixes are recovered. This suggests a more direct principle: limiting representation error helps, but keeping generation on the full-denoising trajectory is a more direct route to accurate decoding. We therefore formulate efficient dLLM inference as trajectory-guided querying. Rather than broadly repairing representation error, we use full-denoising trajectories to supervise which unresolved positions should receive limited fresh computation. Based on this formulation, we introduce TRACEQUERY, a lightweight sparse decoding framework that predicts a small set of near-future trajectory-relevant positions, queries only those candidates, and lets the backbone select which one to commit from their fresh predictions. TRACEQUERY delivers the strongest quality-computation trade-off among cache-based approaches, sustaining high accuracy with substantially fewer fresh token computations, and generalizes across tasks, model variants, and families. Code: https://github.com/fish1928/tracequery.git

Long-term memory enables large language model (LLM) agents to maintain personalized and temporally coherent interactions, but repeated LLM inference for memory admission, extraction, and organization introduces substantial cost and latency. We present SLICE-Mem, a lightweight memory framework that combines salience-guided admission and predicate-centered information extraction. A compact classifier identifies user turns worth retaining, and a semantic-dependency-based extractor converts admitted turns into concise memory entries. This online pipeline requires no generative LLM calls, while cross-entry reconciliation is deferred to an offline consolidation stage. We evaluate SLICE-Mem on the weekly subset of Memora and LongMemEval-S using 8B and 30B Qwen backbones for memory construction and consolidation on both benchmarks. Compared with the generative-memory baseline MemOS under the Qwen3-30B configuration, SLICE-Mem's online pipeline achieves 46.60 versus 46.68 FAMA (Forgetting-Aware Memory Accuracy) on Memora Weekly and 56.8% versus 59.3% accuracy on LongMemEval-S, while reducing total LLM token consumption by 92.4% and 99.5%, respectively. These results demonstrate that SLICE-Mem can preserve useful evidence for long-term memory with substantially less reliance on LLM-based memory construction.

Frontier AI models are increasingly deployed on topics where the stakes are high and expert judgment is required, including healthcare, defense, finance, governance, and politics. We present a general methodology for distilling expert human judgment into automated evaluation systems that operate at scale. Our approach assembles networks of domain experts and develops techniques for extracting not just their conclusions but their reasoning processes, encoding those processes into structured editorial standards that automated judges can apply continuously. We ground our approach in a long tradition of research on the nature of expertise and the conditions under which expert judgment is reliable and more effective than other approaches, and argue that independent, expert-calibrated evaluation is necessary not only on epistemic grounds but as a matter of accountability and public trust. We validate the methodology on AI news reporting as a first application domain. On source quality, two independent domain experts achieve Krippendorff's $\alpha = 0.83/0.69$ on a held-out test set when applying the editorial standards, and our automated judge reaches $F1 = 0.81/0.97$ against expert consensus. On neutrality, experts achieve $\alpha = .72$ on the test set, and the automated judge reaches $F1 = 0.86$ against expert consensus. Additionally, a controlled comparison shows that expert-calibrated judges substantially outperform uncalibrated frontier models on every source-quality metric, demonstrating that encoding expert reasoning is the critical step that prompt engineering alone cannot replace. We present this benchmark as a proof of concept for a broader evaluation infrastructure applicable to any domain where judgment matters and expertise is scarce.

Typed decision-making tasks in agents, such as UI element selection, product recommendation, and chart matching, require models to choose from a set of discrete candidates. While prior work demonstrates that text-only System One models achieve four key properties including candidate-order invariance, scalability, and cache reuse via separate request-candidate encoding, extending them to multimodal settings where candidates are visual remains underexplored. We introduce Iris, a framework that accommodates image, text, or hybrid candidates while preserving the core decision module. We prove that order invariance and exact cache reuse hold for any per-candidate encoder under Iris, and derive the closed-form latency reduction. We construct a visual-candidate benchmark from Multimodal-Mind2Web (K=16, deduplicated, order-shuffled, three generalization splits). Independent encoding (Iris) achieves 78.4 / 72.1 / 74.8 percent accuracy with zero-flip rate under candidate reordering. Joint encoding reaches 90.6 / 89.5 / 91.6 percent but suffers an 8 to 10 percent answer-flip rate when shuffled. To bridge this gap, we propose a design that encodes each candidate conditioned on the request state without inter-candidate cross-attention; it preserves strict order invariance, enables forked cache sharing, and attains 91.7 / 86.5 / 90.0 percent accuracy at one-third of OneJev latency with calibrated null probabilities (AUROC 0.93 to 0.97). Reusing image candidates across requests cuts latency by 37x at K=256 and 191x at K=2048. On MetaEncoder benchmarks, Iris-4B achieves 100 percent on JevBench original and easy, outperforming 30B MetaEncoder and 27B OpenJev. Finally, we resolve a systems bottleneck on Ampere GPUs: reformulating Qwen3.5 patch-embedding Conv3d into its exact matrix-product form accelerates processing by 350x. Code, data, and checkpoints: https://github.com/lllfx/Iris.

Continuous-state flow and diffusion models offer a promising approach to fast, parallel language generation, but matching the generation quality of autoregressive models remains challenging. We present Rectified Language Flow (ReLF), a language model that generates text by transporting noise to token embeddings using rectified flow. ReLF combines several complementary design choices: fixed token embeddings constructed with Hellinger-SVD, self-conditioning guidance trained with cross-entropy, and regularized reflow for few-step generation. We further introduce singular logits, which combine an analytically computed term that drives commitment to a valid token embedding as noise vanishes with a bounded neural residual that learns contextual token preferences. With 64 function evaluations on OpenWebText, ReLF achieves lower generative perplexity than autoregressive baselines at matched diversity across all three model sizes tested, while also outperforming prior continuous flow methods in this setting. After two reflow rounds, ReLF still outperforms autoregressive baselines with 16 function evaluations while generating 1,024-token sequences 15-19 times faster at batch size 1.

Professionals send working documents to external large language models. A common safeguard is reversible pseudonymisation: personal data are replaced with placeholders before the request and restored locally in the answer. We present an open benchmark of five existing open tools and a regular-expression baseline on Russian and UK English, and an open engine built on the two strongest tools. The benchmark excludes each tool's own authors' dataset from its verdict, compares the engine with each tool on identical datasets for detection, reports masking precision, and tests on a held-out set of synthetic documents whose files, hashes, decision rules, run plan and stopping rule were published before any run. Two local LLMs test whether placeholders survive. All evaluation data are synthetic, and the benchmark was run by the author of the engine; code, data, predictions and raw answers are released. On external datasets no existing tool meets all predefined thresholds: in UK English the identifier and leak thresholds are not met, in Russian the identifier threshold is borderline. On the last held-out set of 1,000 documents per language, the best existing tools leave a whole value visible in 61% (Russian) and 58% (UK) of documents. On these documents, from the generator it was developed on, the engine does so in 0.3% [0.1; 0.9] and 0.2% [0.1; 0.7], leaves any part of a value visible in 0.8% and 0.2%, and meets all detection thresholds in both languages. In UK English it also restores at least 99% of its placeholders after 760 answers. In Russian it restores only 94.6% [93.6; 95.6] of name placeholders, against a 99% threshold, mainly because the inflection it borrows from pii-guard produces wrong surname forms. The engine was developed over six earlier held-out sets from the same generator; real documents were not tested.

Recursive self-improvement (RSI) requires an agentic system to design, train, and evaluate its own successor. Data-driven post-training is a central part of that loop, yet existing evaluations measure it on a single benchmark, where an agent selects from a static pool or distills a stronger model to push one score up in one run. Real post-training is multi-domain, and a gain in one domain has to be weighed against a regression in another. We propose ParetoRSI-Gym, in which an agent searches for data, curates it, trains, reads back per-domain results, and repeats, with everything but the data held fixed: the initial LLM, the five domains (math, code, finance, medicine, science), the toolset, and the SFT pipeline. Generalization benchmarks in each domain separate capability gains from overfitting to the scores the agent optimizes against. Across three agent systems, three base models, five rounds and multiple seeds, the strongest system, Claude Code with Opus-5, raises the generalization scores of its proxy-best checkpoint by 9.2% relative to the starting model (+2.3 absolute points), averaged with equal weight over three base models and five domains, with a single-domain relative gain of up to 63% (medicine on Ministral-3-3B). Further analysis shows that the harness matters as much as the model it drives, and that each agent keeps a unique data selection strategy. We see our work as a step towards evaluation of RSI systems in real-world machine learning engineering.

We describe HobbyLM, a family of small sparse Mixture-of-Experts (MoE) models trained from scratch with a few hundred to a few thousand dollars of rented GPU time. The core is a decoder-only transformer with fine-grained experts, sigmoid routing with auxiliary-loss-free load balancing, grouped-query attention with per-head QK-normalization, and aMuon/AdamW optimizer split. Each of these choices was selected with a single-variable ablation suite at 130M parameters. On this architecture we train a 500M-parameter model (169M active) on 40B FineWeb tokens and post-train it into a chat model, a masked-diffusion language model, a function-calling and computer-use agent, and a single "omni" checkpoint that reads text, images, video, sound and speech through frozen encoders and small projectors. We run these models on a laptop CPU with a from-scratch Rust inference engine. Two of the authors then adopted the architecture without architectural changes to build HobbyLM-1B (1.04B total, 305M active parameters), trained on a 100B-token curated mixture with a high-quality anneal. HobbyLM-1B reaches a 47.5 average on seven zero-shot benchmarks, and its builders also release an instruction-tuned model and an on-device macOS computer-use adapter. Finally, we describe HobbyLM-Image, a 333M-parameter flow-matching diffusion transformer that generates 1024x1024 images in a 32x-compressed latent space, together with a dependency-light CPU implementation of the full text-to-image pipeline. We report what worked, what did not, and where the small-scale ceilings lie. We also state for each component which author designed or built it.

User memory helps personalized systems build on their understanding of users over time. In e-commerce, users convey their preferences and needs through both explicit language and implicit interactions with products. We study memory built from logs of these interactions, examining how it can \emph{remember} concrete evidence and established interests, and use that evidence to \emph{infer} underlying needs and preferences that inform judgments beyond well-observed product domains. Existing memory methods primarily target language-based interactions, and evaluations offer limited real behavioral data and limited separation of these capabilities. This leaves open how to jointly support and separately evaluate remembering and inference from behavioral logs. We formulate the problem of continual behavioral user memory around these joint requirements and develop a dedicated evaluation framework and memory method to address them. \textbf{BehaviorMemBench} is a benchmark built from real e-commerce logs covering 1{,}194 users and 17{,}462 behavioral batches over more than three years, with automatic future-behavior labels and a task-independent memory interface to separately evaluate remembering and inference under a common protocol. \textbf{BehaviorMem} is a memory design that organizes explicit behavioral facts and inferred user characteristics across long-term and short-term scopes in a two-axis state, with evidence-grounded formation, continual maintenance, and query-independent readout. Experiments across two builders and two task readers show that BehaviorMem leads the evaluated agentic memories on remember and infer, and reveal different method rankings across the two capabilities and competitive statistical baselines. Validation on a public dataset further supports its usefulness for downstream tasks based on behavioral histories.

Rubric-based evaluation asks which level of a rubric a response meets. It is usually done by prompting a large language model (LLM) as a judge, which needs a generative call per judgment and reads every level description, so its cost grows with the rubric. We present RubriX, a rubric judge whose architecture is inspired by contrastive language models (CLM): a LoRA-adapted encoder embeds the input and each level description, and a softmax over scaled cosine similarities gives a probability for every level without text generation. Trained on public rubric data and proprietary synthetic data, RubriX reaches 75.4% exact accuracy on held-out rubrics, against 57.1% for TypeSafe's Jev and 49.8% for a Qwen3-8B judge. On eight human-labelled benchmark sets it is within 6 percentage points of Jev's accuracy on most of them, while Jev keeps a larger lead on two and ranks responses best on average; against the results reported in the G-Eval paper, RubriX outperforms G-Eval with GPT-3.5 on every SummEval dimension and both QAGS subsets. With level embeddings cached, RubriX's latency stays at about 28 ms per judgment as level descriptions grow from 10 to 3,000 tokens, while Jev's (approximate network time subtracted) rises from 44 ms to 211 ms and the Qwen3-8B judge's from 32 ms to 516 ms, making RubriX 7.7x faster than Jev and 18.8x faster than the Qwen3-8B judge at 3,000 tokens per level.

Disease diagnosis is a key application of large language model (LLM) agents in healthcare. However, existing medical agents are primarily trained and evaluated with static question-answering data, providing limited opportunities to learn active information acquisition, examination decision-making, and dynamic clinical interaction. To address this limitation, we present ClinVerse, an interactive, real-data-grounded multi-agent environment designed for training and evaluating medical diagnostic agents. Built on ClinPatient, a scalable dataset comprising 150k diagnostic cases derived from real-world electronic health records, ClinVerse supports coordinated interactions among six specialized clinical roles. Specifically, a doctor agent starts from an initial patient complaint, progressively gathers clinical evidence through patient interviews and examination requests, and ultimately produces a diagnosis. For each interaction trajectory, a rule-based scoring framework quantifies information acquisition, examination ordering, and diagnostic accuracy. These trajectory-level scores serve as reward signals for agent training while also enabling scalable evaluation. In experiments, we evaluate various general-purpose and medical-domain models and show that nine self-evolution methods across three categories consistently improve agents' diagnostic capabilities within ClinVerse. We further analyze the interplay among model capability, environment feedback, and self-evolution methods.

Transportation agencies maintain extensive crash databases to identify safety priorities, evaluate countermeasures, and guide resource allocation. Making these data accessible to practitioners requires flexible ways to translate analytical questions into reliable results. Existing dashboards support predefined analyses, while additional requests often depend on specialists who understand database queries, coded fields, and agency conventions. This study presents CrashDB-QA, a fully local, schema-grounded agentic workflow that enables natural-language analysis of crash databases. Specialized components interpret questions, generate and validate read-only SQL, execute queries, and explain the results. A shared knowledge base grounds query construction in the agency's data dictionary and safety-analysis conventions, while recorded intermediate outputs support inspection and auditing. An optional extension, CrashDB-QA+, checks the alignment among the question, query, and execution result and supplies evidence for iterative repair. We evaluate the workflow on a frozen snapshot of 4.4 million Wisconsin crash records using a new benchmark of 300 questions across ten analytical categories and nine local open-weight models. Domain grounding substantially improves query accuracy, and a code-specialized 7B model provides the best accuracy-latency balance among the tested models. Compositional analyses remain more difficult than routine retrieval, particularly comparisons between groups or periods. Verification improves performance when the Judge uses authoritative field definitions and the Refiner receives execution evidence, whereas additional checking alone offers no consistent benefit. These findings demonstrate the potential of locally deployed, domain-grounded workflows to broaden access to crash analysis and identify the reasoning and verification capabilities needed for reliable use.

Synthetic supervision can bootstrap semantic parsers in new search domains, but prompt-induced diversity is useful only when generations are faithful to the structured intent. We explore this for cold-start e-commerce search: 118,098 synthetic pairs produced by 6 LLMs with 9 persona-conditioned prompt families on a 47-field UK automotive schema, against 1,214 human-annotated and 15,304 production queries. We show that the IPIP personality bundle improves parser F1 in all four matched contrasts (Holm-corrected $p<.001$), and all three bundles have a positive main effect over the complete $2^3$ grid. Syntax conditioning is the axis that moves generations towards real traffic. It also induces lexical diversity. Alignment sKL, the divergence between von Mises-Fisher distributions fitted to synthetic and human-annotated queries, is an unlabelled proxy for parser F1 ($\rho=-0.917$ against human queries written for simulated shopper scenarios, lower is closer). An LLM committee, checked against a blinded author pass, triages the 927 queries missed by every parser and labels 37% as hallucinated. Dropped constraints dominate. The four syntax-bearing families have the largest hallucinated shares, 38 to 42%, while IPIP alone has the lowest at 28%.

Idioms have long been a challenge for machine translation, and while large language models (LLMs) have proven increasingly adept at handling them, whether quality estimation (QE) systems have kept pace and can reliably score idiomatic renderings of these expressions above faulty literal ones remains unclear. We introduce the first parallel dataset pairing idiom-bearing English sentences with both literal and idiomatic translations across nine target languages, augmented into multi-sentence segments and a graded literality continuum. Using this dataset, we benchmark eleven conventional QE metrics from the COMET and MetricX families and ten LLMs from the Qwen-3, Gemma-4 and OLMo-3 families used as GEMBA-style judges. We find that most systems, especially older metrics and smaller models, display a marked literality bias, rating literal translations of idioms above idiomatic ones. Higher parameter counts and newer training sets appear to reduce this bias to an extent; nevertheless, of all the benchmarked systems, only MetricX-24-Hybrid (XL, XXL) and Gemma-4-31B consistently rate idiomatic translations of idioms above literal ones, and only Gemma-4-31B reliably ranks intermediate degrees of idiomaticity and literality. Our findings caution against deploying standard QE metrics in domains rich in figurative language, where higher scores might be rewarding undesirable literalism.

Automated fake news detection is vital on social media, where fake news is easily rewritten to fool detectors. However, existing approaches typically rely on superficial writing cues or single-style text rewrites, resulting in poor robustness under diverse cosmetic modifications. We construct MANTIS-Bench, a multi-axis presentation-attack benchmark that pairs original news with rewritten views across writing style, emotional expression, and dissemination framing. We further propose MANTIS, an instance-adaptive adversarial framework enabling dynamic worst-case training tailored to each news article with zero inference-time overhead. This reframes fake news defense from heuristic format smoothing into targeted, fact-invariant credibility assessment. Experiments on MANTIS-Bench show that MANTIS reduces the Union Attack Success Rate (UASR) by 3.55 percentage points relative to SheepDog on GossipCop and consistently outperforms prior single-axis defenses on Macro-F1 while maintaining competitive clean accuracy, demonstrating the effectiveness of multi-axis and instance-adaptive worst-case training.

Activation steering offers training-free and dynamic control of target behaviors in large language models, but existing methods can impair general capabilities and require domain-specific intervention designs. We introduce Capability-Preserving Neuron Steering (CANS), a unified framework that balances target-behavior control with capability preservation. To make target-behavior improvement and general-capability preservation tractable, we introduce two proxies: log-probabilities of target tokens in model generated responses, and output-distribution KL divergence on general-purpose inputs. Combining the local approximation of the two proxies yields a closed-form solution, enabling lightweight steering without iterative model training or additional inference-time computation. The resulting neuron-wise scaling coefficients achieve effective steering while limiting general-capability loss. A shared data-construction and intervention procedure supports adaptation across domains using sampled responses with target-behavior filtering and target-span annotations. Across safety, toxicity, and hallucination control on Qwen and Llama models, CANS achieves comparable or better steering while better preserving MMLU, BBH, and IFEval performance than SFT, CAA, and neuron-level baselines. On Qwen3-4B, CANS reduces HarmBench attack success from $15.50\%$ to $1.75\%$ with a $1.11$-point decrease in average capability. These results support CANS as a practical approach to behavioral control across domains.

Language models are increasingly considered for clinical decision support, where their answers should be grounded in clinical evidence and remain robust to irrelevant information. Prior work shows that a single irrelevant sentence can lower the accuracy of medical models, but not where inside the model its effect can be removed, or how to do so without knowing which sentence it is. We find that blocking attention to a known distractor restores the answer, whereas head masking can suppress correct answers. We also show that the model's hidden states mark the distractor as the sentence that least fits the case even when the model follows it. To remove a distractor's effect without knowing which sentence it is, we propose EXCISE, an inference-time method that blocks the sentence most influencing the answer, but only if the model's hidden states mark it as inconsistent with the case. On MedQA-TD (MedQA with Targeted Distractors), a benchmark we construct in which each distractor targets a specific wrong option, EXCISE gains 16.7-22.8 accuracy points at a 1.5-2.1-point cost on the same questions without distractors, outperforming prompting, rewriting, and steering baselines for all three models. Its gains extend to MedDistractQA and BiasMedQA, as well as chain-of-thought generation.

Deep research, or any use of large language models (LLMs) to read a large document collection and produce insights, requires synthesizing information from different sources. In this work, we find that LLMs produce long-form research reports that read as confident syntheses but which fail to accurately reflect the underlying documents. We call this failure smoothing. In collapsing diverse and potentially conflicting sources into a single narrative, the model discards the evidential "jaggedness" of its corpus and asserts conclusions more decisive than the citations can support. We introduce a claim level scoring pipeline, SmoothScore, to quantify this effect. Our pipeline extracts key conclusive claims from model generated reports, then finds evidence relevant to each claim and clusters that evidence along high level aspects. Finally, each claim is scored with respect to the evidence to quantify the degree of smoothing done by the agent. Across 7 systems and 4 datasets, we find that agents consistently produce stronger claims than warranted.

Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable 10.12$\times$ and 5.12$\times$ speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance. Our code is available at https://github.com/VILA-Lab/dSpark.

Vision-language models are shifting from offline passive understanding toward online proactive interaction. However, prevailing methods largely rely on explicit requests while miss the implicit user intent (e.g., "where should I click next?" while a user navigate a website), meanwhile fail to update the state simultaneously as user requests might update continuously. These limitations make them less suited to intent-aware interaction in practical. To address these challenges, we introduce StreamVibe-Bench, a new benchmark serving as testbed for user intent interaction, with 4.1K evaluation samples and 85K training samples spanning explicit, implicit, and dynamic intents. Building on this benchmark, we develop StreamVibe-VL, a streaming proactive model that infers implicit user intents from continuous observations and dynamically tailors its interactions as these intents evolve. We further introduce StreamVibe-Harness, which maintains user intents as persistent states with a fleixble lifecycles via a hierarchical memory to track the latest user intent and environmental state. Our experiments reveal three key findings: (i) StreamVibe-VL preserves strong general video understanding while improving response timing and answer quality by 10.7% over other baselines; (ii) existing proactive models remain limited in understanding user intent and deciding when to respond, especially under implicit and dynamic intents; and (iii) StreamVibe-Harness further improves response consistency and adaptation over sustained interaction, resulting in a 12.3% gain in proactive interaction.

Chest computed tomography (CT) reports generated by large language models (LLMs) can describe positive findings that lack localized evidence in the scan. We propose EviGround-CT to verify these findings after report generation. A deterministic extractor identifies positive finding sentences, and a 3D grounding model predicts an evidence mask for each sentence. Using the full sentence retains anatomical location, laterality, and descriptive modifiers in the grounding query. We evaluate report generation on CT-RATE, and grounding and verification on a held-out 100-case ReXGroundingCT subset. Using the available annotations, sentence-grounded verification raises evidence-supported precision (ESP) from 0.263 to 0.365 across 137 extracted in-scope positives, compared with category conditioning. The corresponding unsupported positive finding rate falls from 0.737 to 0.635. The paired case-level bootstrap ESP difference is +0.102 (95% confidence interval: [0.028, 0.181]). The predicted masks associate individual report statements with spatial evidence that can be inspected alongside the CT.

Autoregressive (AR) language models generate tokens sequentially, so generating long outputs is costly. Speculative decoding mitigates this by using a lightweight drafter to propose multiple tokens for the target AR model to verify in parallel, while a diffusion-based drafter can produce an entire draft block in one forward pass. In models with AR and diffusion modes, one network both drafts and verifies, but drafting still waits for verification, so the standard schedule needs two sequential model calls per iteration. Existing one-call schedules avoid the wait but incur input size quadratic in the speculative width. Motivated by this efficiency bottleneck, we introduce Packed-SS, a self-speculative decoding schedule that verifies and drafts in a single model call with input size linear in the speculative width, while preserving the target model's distribution. The key idea is to let the diffusion drafter run one call ahead: while verifying the current draft, the model simultaneously constructs the next draft from an older verified prefix and keeps its unresolved suffix for the next iteration. Under a mild monotonicity assumption on the drafter, we further show that Packed-SS requires no more model calls than the standard two-call decoding in the greedy setting. Experiments on Nemotron 3B/8B and SetDLM 1.7B show higher throughput and lower mean wall time than both baselines under greedy decoding.

Skill learning enables web agents to reuse prior experience for long-horizon interactive tasks. However, learned skills are typically evaluated in closed environments, overlooking the fact that real-world web deployment involves continuous shifts in invocation context, perceptual grounding, and execution feasibility. To this end, we introduce SkillShift-Bench, the first benchmark for evaluating skill learning with controlled open-environment shifts. SkillShift-Bench simulates Contextual Shift, Perceptual Shift, and Execution Shift across three self-hosted web systems, and evaluates agents with diagnostics covering task success, interaction cost, skill use, and skill failure. Comprehensive experiments reveal systematic robustness degradation under controlled shifts, characterized by lower success rates, higher interaction costs, and more frequent skill failures, with Execution Shift exerting the strongest pressure. We further formulate shifted skill execution as a nonstationary local executability problem and provide both a dynamic-regret analysis and a practical test-time adaptation method, demonstrating the feasibility of improving learned-skill execution without relearning the skill library. Code, tasks, environments, documentation, and the leaderboard are available at https://skillshift-bench.github.io/.

LLM-based evaluators have rapidly become a standard tool for machine translation evaluation, yet the design choices that shape their behavior remain poorly understood. We present a systematic study of how instruction design, few-shot demonstrations, backbone model choice, and multi-step workflows affect span-level and ranking-based MT evaluation. We collect new expert RATE annotations on WMT24 and WMT25 English-Russian translations, with transfer validation on MQM- and ESA-annotated benchmarks across five additional language pairs. Our experiments reveal four robust findings: instruction structure shapes evaluator behavior, high-quality demonstrations matter more than quantity, frontier backbones exhibit distinct evaluator profiles (recall-, precision-, or ranking-oriented), and fusing outputs across diverse backbones outperforms single-evaluator multi-call setups. Although evaluator design exposes a large combinatorial space, several discovered design principles transfer consistently across annotation protocols and language pairs. We synthesize these findings into ORBIT, a family of LLM-based MT evaluators. Our strongest configurations achieve expert-level performance in error span detection and in Accuracy and Fluency ranking at roughly 20x lower cost than expert human annotation.

As machine translation systems approach human-level quality, traditional evaluation methodologies struggle to detect subtle translation errors. We critically examine limitations in current gold-standard approaches (MQM and ESA), including inconsistencies from variable annotator expertise, excessive categorization complexity, coarse severity granularity, accuracy bias over fluency, and time constraints. To address this issue, we introduce a high-quality dataset consisting of human evaluations for English-Russian translations from WMT24, created by professional linguists. We show that expert assessments without time pressure yield substantially different results from standard evaluations. To enable consistent and rich annotation by these experts, we developed the RATE (Refined Assessment for Translation Evaluation) protocol. RATE provides a streamlined error taxonomy, expanded severity ratings, and multidimensional scoring balancing accuracy and fluency, facilitating deeper analysis of MT outputs. Our analysis, powered by this expert dataset, reveals that state-of-the-art MT systems may have surpassed human translations in accuracy while still lagging in fluency - a critical distinction obscured by existing accuracy-biased metrics. Our findings highlight that advancing MT evaluation requires not only better protocols but crucially, high-quality annotations from skilled linguists.

De-identifying clinical notes with a hosted language model discloses the information the step is meant to protect. We present EdgeDeID, a 0.6B-parameter decoder-only de-identifier that runs locally and is trained without real patient text: three LLM teachers write synthetic notes, a single labeler re-annotates them into one convention, and the model learns to emit a typed XML schema of the identifiers it finds. On a 100-note synthetic in-domain benchmark with manually verified annotations it reaches 0.994 micro-F1 and leaves no missed identifier in 99% of notes, against 0.857 and 12% for the strongest of eight off-the-shelf systems. Quantized, it runs on CPU in 615 MB. Of three additions to supervised training, only a verifier adapter helps: it separates complete from incomplete outputs of other systems at 0.94 AUROC, whereas reinforcement learning with a recall-weighted reward and a rule-based post-processing layer leave in-domain results unchanged. On MIMIC ICU nursing notes F1 falls to 0.48. The errors follow two properties our synthetic notes lack, telegraphic style and low identifier density, and three hand-written in-context examples recover 0.08 F1 without annotated target data.

Tool-using large language models (LLMs) must decide whether external information warrants action. We investigate whether their expressed credibility judgments reliably indicate behavioral resistance to misinformation. We evaluate five LLMs on 196 constructed false notices paired with legitimate counterparts and 204 archived attacks adapted to simulated tool use. When asked to act on the constructed notices, models issue harmful tool calls in 27%--99% of cases they have rated untrustworthy. Removing the preliminary rating step increases harmful execution by 15 percentage points on average, showing that credibility elicitation itself changes the behavior being evaluated. Interventions that improve credibility discrimination do not consistently yield safer actions. A simple forewarning significantly reduces harmful execution in four of five models, but neither tested generated-example prebunk outperforms it in the main evaluation. These reductions can also compromise legitimate task completion. Supervised fine-tuning on generated false notices reduces harmful execution on held-out archived attacks under the rating-first protocol, with model-dependent utility costs and weaker protection without rating elicitation. Our findings motivate evaluating misinformation defenses through both harmful actions and legitimate task completion, while accounting for the behavioral effects of the evaluation protocol.

AI foundation models play a pivotal role in advancing heart disease diagnosis from electrocardiogram (ECG) data, particularly for rare diseases where the lack of sufficient labeled data hinders supervised learning. However, most self-supervised learning methods treat ECGs as generic sequence or image data without exploiting underlying physiological structure. To address this limitation, we propose QECGBERT, a self-supervised ECG language model that explicitly encodes physiological semantics by tokenizing the P, QRS, T, and isoelectric background waveforms. Vector-quantized variational autoencoder (VQ-VAE) tokenization is applied to each 12-lead ECG, with tokens enriched by continuous waveform features to preserve morphology lost during quantization without increasing sequence length. This design enables compact BERT masked language modeling with short sequences, improving scalability and reducing computational cost relative to long-sequence approaches. QECGBERT applies BERT pre-training to the MIMIC-IV dataset containing over 800,000 unlabeled ECG recordings, and fine-tuning on labeled PTB-XL, CPSC2018, and Chapman--Shaoxing--Ningbo datasets for evaluation. The results demonstrate that QECGBERT achieves comparable or superior performance to published benchmarks, offering a scalable, physiologically aware approach to ECG foundation modeling for cardiac diagnosis.

On-policy self-distillation (OPSD) trains a student on its own reasoning trajectories using token-level supervision from a reference-conditioned teacher initialized from the same model. However, access to the reference can induce teacher behaviors that undermine independent reasoning in thinking models, leading to reference hallucinations, where the student cites an unavailable reference to justify reasoning steps, and reconsideration suppression, where the teacher discourages verification and revision at uncertain steps. We introduce Steered-Teacher OPSD (ST-OPSD) to jointly address these failure modes through activation steering. ST-OPSD extracts two directions from model-generated reasoning trajectories and applies them to the teacher's hidden states during distillation target computation, suppressing reference-invoking behavior and encouraging reconsideration with marker-family calibration. The student learns from the resulting teacher distributions without being steered. Experiments on AIME and HMMT show that ST-OPSD improves accuracy over vanilla OPSD and other interventions while restoring reconsideration and reference-hallucination rates to base-model levels.

Speculative decoding significantly accelerates autoregressive text generation in Large Language Models (LLMs), but its efficacy depends largely on two key factors: the distributional alignment between the draft and the target model, and the draft model's latency during inference. However, existing approaches typically utilize independently pretrained small draft models or seek to prune the target model to a manually chosen architecture, thereby failing to explicitly balance the trade-off between draft model alignment and token generation efficiency. To address this gap, we formulate draft model design as an optimization problem and propose T2D, a framework that constructs draft models directly from the target model using latency-aware structured pruning and knowledge distillation. In this manner, T2D obtains draft models that are jointly optimized for both measured inference latency and cross-entropy loss on the target model's responses, resulting in superlative speculative decoding performance. Through extensive experiments on multi-domain speculative decoding benchmarks and across several target model families of differing scale, we demonstrate how T2D prunes up to 92\% of the target model to obtain highly performant draft models compared to competitive baselines. Moreover, draft models obtained via T2D can also be utilized by other speculative decoding token-alignment and knowledge distillation methods in a plug-and-play fashion, leading to further gains. Our findings underscore the importance of jointly navigating the draft model architecture-alignment frontier to unlock the full potential of speculative decoding for LLMs.

Reasoning language models generate thousands of tokens per answer, and generation speed is limited by memory bandwidth, so storing weights in two bits instead of bfloat16 can potentially make inference 8$\times$ cheaper. Two-bit methods are usually judged by perplexity and multiple-choice question-answering accuracy, both of which say little about whether a model can still reason through more difficult problems. We present ThinkBit, a strictly scalar two-bit framework built to preserve reasoning, in which every weight is an integer codeword times a bfloat16 scale shared by a group of weights. We first use a sensitivity metric to assign each layer a ternary, 2-bit, or 3-bit tier to allocate bitwidths optimally among layers. This is followed by a layerwise Gumbel-softmax-based scalar quantization algorithm that searches for optimal codewords block by block. Since a layerwise scheme does not capture how quantized layers interact with each other, we recalibrate scales, with the codewords frozen, using a knowledge-distillation-based loss against the teacher model. We then perform a short quantization-aware training (QAT) stage that lets the model mimic the base model's distribution by reassigning codewords. Finally, a paired reasoning test on training problems selects the final checkpoint between the recalibration stage and the QAT stage. On Qwen3-1.7B at matched size, our recalibrated checkpoint scores 46.7 on MATH-500 against 30.0 for EfficientQAT trained on the same data, and 73.0 after QAT on reasoning traces; at 0.52 GB it is smaller than every published 2-bit GGUF of the model yet more accurate on MATH-500. We show that this framework scales and up to 14B, a 3.74 GB model scores 94.8 on MATH-500, against 97.5 for the 29.5 GB bfloat16 model.

Lean verifies that a formal proof establishes its stated proposition under declared assumptions, but does not check whether that proposition faithfully expresses the mathematical claim in a manuscript. Assessing this correspondence requires interpreting the main theorem through its project dependencies, which may be scattered throughout a large development. We introduce MechAlign, a framework for project-level semantic alignment review that combines formal dependency extraction with LLM-based assessment. Its extraction tool, LeanTrace, constructs a statement dependency graph from elaborated declarations, retaining definition bodies and structure constraints while excluding proof bodies and stopping at an explicit library boundary. The reviewer uses this context to investigate alignment and report discrepancies with localized evidence. We conduct a preliminary review of 201 main results from 173 OpenAI Math projects, using supplied informal descriptions and available paper context. A hyperbolic-group case illustrates how dependency inspection localizes a missing dimensional constraint relative to the manuscript, consistent with the upstream declaration of formalization scope. We also construct LeanReviewBench, comprising 501 fixed claim-statement-context triples from OpenAI Math, Fermat's Last Theorem, and Kakeya, with description provenance recorded. The resource supports controlled comparisons of more project-level semantic alignment evaluations.

Serving web queries at scale with multi-agent systems requires jointly configuring large language models, collaboration modes, agent counts, interaction rounds, and skills. However, the resulting configuration space grows combinatorially, while sparse execution rewards provide limited fine-grained supervision for interdependent planning decisions. We propose EvoRouter, a self-evolving multi-agent routing framework driven by insight-guided on-policy distillation. A dedicated expert periodically synthesizes performance-cost feedback from configuration trajectories across tasks into reusable conditional routing insights about capability requirements, returns from resource scaling, and interactions among configuration choices. A privileged teacher combines these evolving insights with current execution feedback to provide dense token-level supervision for the router's own planning trajectories through on-policy distillation. Cost-aware reinforcement learning jointly optimizes task performance and computational cost. Together, these training signals internalize execution insights into the routing policy. The updated policy generates new trajectories that support subsequent insight extraction, forming an execution-driven self-improvement loop. We evaluate EvoRouter on 19 in-domain datasets spanning five domains and two out-of-domain benchmarks. EvoRouter achieves an average in-domain accuracy of 68.69%, outperforming the strongest routing baseline by 7.23 percentage points while reducing inference compute cost by 50.95%. Our code is open-sourced at https://github.com/CenYangtze/EvoRouter.