2026-10-05

Language models increasingly assist people with power-related requests, so systematic differences in whom they help could shift the distribution of power at scale, or be exploited by users who learn which identities are refused less. We introduce PowerBench, an evaluation of power-shifting requests that distinguishes self-empowerment, disempowerment, and power grabbing, plus a control of refusal-inducing requests that shift no power. We build, curate, and open-source a dataset of such requests varying the power domain, the context, the scale of the affected party, and the prior power standing of the user, and evaluate 24 models (12 from US and 12 from Chinese developers) under three experimental conditions: reciprocal nationalities of user and affected party, an AI agent as the user, and 8 request languages. Models refuse power grabbing more than disempowerment, and disempowerment more than self-empowerment. Refusal of power grabbing rises with the scale of the affected party, from an individual to a society. Models are biased toward helping others take power from the US and against helping US users take power from others, but favor the US when it gains power and nobody loses it. When the user is an AI agent, refusal of power-shifting requests increases, especially in power grabbing against an individual. Finally, language biases refusal, but in model-specific ways that largely cancel on average. We release PowerBench to make these asymmetries measurable in current and future models.

Static evaluations credit a language model for naming the right move, but an agent must carry a plan through to a verified outcome while an opponent responds. We introduce XiangqiBench, an executable benchmark that measures this difference in Chinese chess: starting from 119 tactical endgames with forced mates supported by engine or checks-only search, an LLM agent must deliver checkmate against an engine defender. An interactive REPL interface separates real moves, state queries, and forward simulation, and we record 8,568 multi-turn trajectories from 12 frontier LLMs under two observation protocols. Three signals that look like competence each overstate closed-loop success. (i) The Conversion Gap: models play the stored reference first move in 26.1\% of Sighted trials, yet only 13.9\% of these trials end in a win. (ii) The Consistency Gap: the leading model reaches 38.7\% pass@3 but only 5.9\% pass^3, winning all three trials on 7 of the 46 positions it ever wins. (iii) The Simulation Gap: 32.3\% of accepted simulation calls stop on an illegal move, and in 49.3\% of comparable cases the real defender replies differently from the line the agent simulated; self-authored rollouts check legality but cannot anticipate the opponent. Finding the move is not winning the game: agent evaluations should score closed-loop outcomes and report reliability alongside coverage.

We present OpenGameEval, a benchmark and evaluation framework for agentic game development inside Roblox Studio. It runs language models as agents in reproducible, stateful game-engine sessions and scores each run with executable checks, both on the edited scene and in a simulated play session. Most agentic coding benchmarks require exploration but score only final task success. OpenGameEval separates observation tools from editing tools in its eight-tool action space, so exploration can be measured directly. We measure the pass rates and exploration behavior of 13 frontier models on 84 human-curated core tasks, with 16 attempts per task. The tasks are hard for current models. The best model solves 51.7% of tasks on a single attempt and 39.4% five times out of five, and no tested model solves six of the tasks. Models at the frontier reach similar pass rates by solving different tasks: splitting tasks by the kind of work they require spreads the top five by 5.0pp on script-authoring tasks and 12.5pp on scene-change tasks. Exploration behavior predicts whether a run succeeds. Holding task and model fixed, a run that inspects every object a reference solution touches before acting on it passes 13.4pp more often than a run that inspects none of them on scene-only tasks, and 9.8pp more often on script-only tasks. We release the task suite, its place files, the per-task annotations, a plugin that runs the tasks inside Roblox Studio, and an updated leaderboard under the MIT license at https://github.com/Roblox/open-game-eval.

Modern AI systems are designed to refuse hazardous requests. A jailbreak is a prompt crafted to bypass those safeguards and elicit outputs that the system would normally refuse to provide. The MLCommons Jailbreak Benchmark v1.0 provides an end-to-end methodology for evaluating the robustness of large language models to single-turn, text-based jailbreak attacks. It combines criteria-driven system and attack selection, paired baseline and adversarial evaluation, human annotation, automated evaluator calibration, scoring, grading, and risk-calibrated disclosure within a single benchmarking pipeline. The benchmark evaluates eight open-weight systems using 264 seed prompts spanning eleven hazard categories and representative attacks drawn from the MLCommons Jailbreak Taxonomy. Responses are assessed using the AILuminate Assessment Standard v1.4, and robustness is measured through the Resilience Gap: the change in safety performance between baseline and adversarial conditions. Across all evaluated systems and attacks, the unsafe-response rate increased from 11.08% under baseline conditions to 18.65% under jailbreak conditions, producing an average Resilience Gap of 7.57%. Accessible systems showed a larger mean gap, while attack effectiveness varied substantially across attack categories and hazards. The benchmark also examines evaluator reliability and sources of measurement error. Beyond reporting results, Jailbreak Benchmark v1.0 establishes a reproducible methodological foundation for comparative jailbreak evaluation and for future expansion across systems, attacks, hazards, and evaluation methods.

Large Language Models (LLMs) are prone to generating hallucinated content, which compromises their reliability in knowledge-intensive tasks. To address this challenge without sacrificing creativity, we propose HARPO, a reinforcement learning framework designed to jointly optimize faithfulness and creativity. HARPO incorporates a Hallucination-Aware Generative Reward Model (HA-GRM), trained via verifiable feedback, to assess both faithfulness and writing quality. A Selective Activation Mechanism (SAM) activates writing rewards only for outputs judged hallucination-free by HA-GRM, while a data curriculum progressively shifts training from creative writing to hallucination-centric tasks. On RAGTruth, our Qwen3-4B-based HA-GRM achieves a response-level F1 score of 78.08%, compared with 66.37% for the supervised fine-tuning baseline. Experiments on Qwen2.5 and Qwen3 models from 1.7B to 8B parameters show improvements in both faithful generation and writing quality. On Qwen3-4B, HARPO reduces the HA-GRM-judged hallucination rate on MultiHopRAG from 3.29% to 1.02%, while increasing the Arena-Hard-v2.0 creative-writing score from 16.95% to 27.54%.

Large language models are increasingly used where small syntactic errors matter, yet character-level reasoning is still evaluated mostly through isolated probes and aggregate accuracy. We introduce SyntaxBench, a diagnostic benchmark and statistical evaluation framework for character-level reasoning. It contains five core tasks, character counting, letter containment, palindrome detection, edit distance, and longest-string selection, plus index_to_span, a harder substring-extraction stress test. The five core tasks use paired English and character-length-matched random-string inputs. index_to_span documents share a 200-500 word band and are not character-length matched. All six tasks use zero-, one-, and four-shot prompts. We evaluate eight open-weight models from 2B to 32B parameters across 11 reasoning-mode configurations. The framework reports exact-match and relaxed accuracy, Cohen's kappa, paired McNemar tests with odds ratios, bootstrap confidence intervals, Kendall's tau, class-conditional metrics, tokenization analysis, and multiple-comparison-corrected tests. Three findings stand out. First, tokenization shapes accuracy: random strings are more character-visible than English strings (1.892 vs. 3.169 characters per token), and character-counting accuracy falls as English words occupy more tokens. Second, reasoning mode is not uniformly helpful: Gemma4-31B is nearly unchanged across modes on the near-saturated tasks, while Qwen3.6-27B is worse with thinking on palindrome detection (0.952 non-thinking vs. 0.886 thinking at four-shot). Third, index_to_span remains largely unsolved; the best four-shot exact-match accuracy is 6.75%. Character-level evaluation needs controlled inputs, paired tests, and analyses of tokenization and reasoning mode rather than aggregate accuracy alone.

Large Language Model (LLM) agents are increasingly deployed in high-stakes settings such as industrial maintenance and equipment fault troubleshooting, where workers occupy a variety of roles. A capable agent must therefore act in a way that is calibrated to user's role: taking actions and providing information that respect the role's knowledge and capability boundaries. Unlike coding, where mistakes are usually recoverable, agent responses in these settings are enacted on physical equipment, and can therefore cause irreversible equipment damage, production loss, or personnel harm. Existing benchmarks, however, largely overlook the need for agents to infer what a role intends and acting only through tools that role may legitimately use, a capability which we term Perspective Awareness. To this end, we introduce ReFract, a benchmark of 150 expert-validated entries in which an agent must act differently in response to the same query depending on user's role. Entries of ReFract are grounded in anonymized queries from domain support conversations, against which we construct Text World Models that simulate the agent's operating environments and assemble perspective-aware action trajectories. State-of-the-art LLMs solve at most 69% of the tasks with more than 50% of their trajectories contain attempts of taking perspective-violating actions. ReFract exposes perspective awareness as a distinct, largely unsolved axis of agent evaluation and motivates agents that calibrate not just how to act, but for whom.

Large language models (LLMs) are increasingly used to support legal practice, education, and research, yet their reliability in national legal systems outside the United States remains largely undocumented. We introduce an expert-validated benchmark for evaluating LLM reliability on the Colombian legal system. The benchmark comprises 1,042 items spanning ten areas of law and three question formats (closed multiple-choice, semi-open, and open-ended IRAC), built through a human-in-the-loop pipeline with multi-stage expert review. We evaluate 15 contemporary proprietary and open-weight models with format-appropriate metrics. Accuracy on closed questions ranges widely, from 0.905 (Gemini 3.1 Pro) to 0.577, but on free-text legal answers factual correctness never exceeds 0.45 (on a 0-1 scale) for any model. We find a dissociation between answer relevancy and correctness (Spearman rho = -0.46): models reliably sound responsive while frequently being wrong, a pattern of particular concern for non-expert users. Closed-question accuracy and free-text correctness are strongly rank-correlated (rho = 0.94), so cheap multiple-choice screening predicts model ranking but overstates absolute reliability. An independent rubric-based LLM judge and blind human expert scoring both reproduce the free-text ranking (rho >= 0.88). The judge further reveals that only about half of the norms models cite are correct; the rest are wrong or non-existent. Reliability varies systematically by legal area and follows an inverted-U across question complexity. Our results indicate that current LLMs require expert supervision for Colombian legal tasks, and that grounding answers in authoritative sources is a promising path to higher reliability. We release the benchmark construction pipeline to support reproducible evaluation.

HakemBench is a Turkish benchmark of typed decisions, in which the model under test reads a text, a question and a fixed set of options and returns a probability for every option. Version 1.0 is released fully open under CC BY 4.0, with 2,346 items and 4,275 choice, yes/no and score questions in seven tracks (fact-check triage, education, guardrails, legal routing, moderation, spam and phishing, and customer support). One harness scores decision quality (macro F1), calibration (from the normalised Brier score) and selective automation (from the normalised area under the generalised risk-coverage curve), combines them by a geometric mean and reports intervals from 2,000 bootstrap draws; probes for option order, paraphrase, English translation and substituted names are reported alongside. Most gold labels come from blind passes of one AI model family compared with the votes of a panel of large language models from other model families; they are not human-verified. On a board of 16 rows the leader scores a composite of 0.888 and the lab's own model is 7th at 0.660. Its numbers are not blind. Earlier runs' test results shaped its training data, so its guardrail, moderation and customer support numbers are flagged; with every model scored on the other four tracks only, its composite is 0.678, 6th of 16.

Large language models (LLMs) have demonstrated strong performance in code generation, where success depends on both recalling relevant algorithmic knowledge and reasoning about how to apply it. However, existing LLM pipelines are opaque, with no explicit separation between these two components. We argue that for well-known algorithms whose canonical implementations are widely accessible in pretraining corpora, code generation is better measured as \textit{parametric code retrieval}: reproducing a named algorithm from internalised knowledge rather than synthesizing a novel one. We introduce AlgoREval, a benchmark of 599 problems spanning classical 77 algorithms across 14 domains, 7 programming languages, and 4 graph-input representations to evaluate this capability in isolation, and assess 15 models (7B--34B parameters) in a zero-shot setting. We find substantial variation in retrieval accuracy across languages and input representations, even for widely documented algorithms and show that prompt augmentation with retrieved code snippets or structured algorithmic hints improve accuracy on complex algorithms, while SFT achieves broader language gains and GRPO achieves larger per-language gains on specific languages. Together, our results establish parametric code retrieval as a distinct, measurable capability and caution against deploying AI-generated algorithmic code without systematic validation.\footnote{Code and dataset are available at https://github.com/Nickil21/AlgoREval

Agents are increasingly given open-ended research tasks: discovering an empirical law from self-designed experiments, improving a heuristic whose optimum nobody knows, or beating a standing record. Their execution logs record every step of this research, yet the runs are still judged by their outcome score. That score alone does not establish whether an agent's claims follow from executed experiments, and a reference answer may be unavailable. We evaluate the agent's epistemic process: how it forms hypotheses, tests them, and revises them in response to evidence. We introduce OEB (Open-Endedness Bench), a benchmark-agnostic methodology that reads only the agent's execution record and never a reference answer or an outcome score. OEB compiles the record into a unified epistemic event graph whose edges connect the propositions the agent states to the executed actions that test them; each node carries an exact excerpt that code verifies against the record. One principle governs scoring: prose can state a proposition, but only evidence returned by an executed action can support or refute it, so OEB checks what the agent writes against what it actually ran. From the graph, OEB scores four competence axes (evidence, experiment, revision, and no reward hacking), mostly as the share of opportunities for sound research that the agent took, and profiles six subjective persona traits that describe the agent's research habits. We score 119 existing runs over 12 tasks from three benchmarks: LLM post-training, chip design, and a training-speed record. Against logged results, only 16-29% of the improvements agents claim are real. On 9 of 10 tasks, the best run tries more new ideas in its second half than the worst run. The persona readings follow the model: for every trait, the model that ran explains more of its variance across runs than the task (a median of 43% against 7%).

Evaluating the occurrence and triggers of large language model (LLM) behaviors - such as sycophancy, self-preference, or over-confidence - is critical for predicting real-world model deployment risks. However, existing situated behavioral evaluations typically ignore cultural context, limiting their generalizability across an increasingly global user base. To address this gap, we propose CuBEs - Culturally-situated Behavior Evaluations that probe for response patterns across diverse user cultures. We first extend an automated testing pipeline to inject cultural context into behavioral test scenarios and subsequent evaluation. We assess the cultural adaptability of this pipeline by building a human-labeled dataset that captures nuanced dimensions of behavior understanding across 12 distinct cultures. Our dataset reveals significant cross-cultural variations that one-size-fits all judgments fail to capture. Through evaluating 13 open- and closed-source LLMs, we find that introducing cultural situatedness in the evaluation scenario creates significant variation in the presence of a behavior. For example, while our baseline experiments testing for political bias capture localized Western political dimensions like the American conservative-progressive divide, non-Western culturally situated evaluations surface entirely different axes of bias such as religious and colonial political issues. Our findings demonstrate that standard, culturally-agnostic evaluations fail to capture these shifts, highlighting the necessity of culturally situated behavioral testing for global deployments.

A compressor can keep the facts of a dialogue and still drop the turn that changed them. A user corrects a price, reverses a choice, or adds a constraint. We call this failure turning-point eviction. One overall retention score hides it, because that score mixes what the user first wanted with what the user wants now. We introduce TPBench, which evaluates three complementary information targets at shared nominal retention budgets. P1 asks for the user's initial goal. P2 asks for the current value of a slot the user revised. P3 asks for both, in dialogues with a late annotated slot update. The current-value answers come from the human dialogue-state annotations of MultiWOZ and SGD. The initial-goal answer is the first sentence of the first user turn. Neither requires new crowdsourcing. The probe-specific evaluations rank compression methods differently. On the joint probe at a retained fraction of 0.30, every tested compressed method remains below full context with the main Llama reader. Deleting the turn that carries the update sharply lowers current-value accuracy, while deleting one matched irrelevant turn leaves it unchanged. A Mistral reader repeats the P2/P3 rankings and the joint-probe gap. Current-value recovery is tested on an additional corpus, LongMemEval-KU, and on Chinese RiSAWOZ: full context has the highest accuracy, and recency has the highest compressed-method mean in both evaluations.

Document databases such as MongoDB are core infrastructure for modern applications, and natural-language interfaces to them---text-to-MQL---would let non-experts query complex, semi-structured data without mastering the query language. Progress on this task depends on high-quality benchmarks, which are most practically obtained by converting an existing text-to-SQL benchmark to the document setting. Unfortunately, existing efforts rely on heuristics for mechanical conversion: the document schema mirrors the relational foreign-key graph, and each query mirrors its source SQL. As a result in our experiments, these approaches fail to migrate 6 of 21 BIRD databases outright, silently drop up to 25.9\% of rows on others, and yield schemas whose ground-truth queries run over an order of magnitude slower as the data scales. We instead propose a conversion pipeline, driven by coding agents with human-in-the-loop verification, that designs each document schema from expected access patterns and rewrites queries to be MongoDB-native. Applying it to BIRD, we build an access-pattern-based text-to-MQL benchmark (AptMQL-Bench). It includes 21 document-oriented databases, 3,186 natural-language requests, and their associated MQL queries---whose databases are migrated from SQLite without data loss and scale efficiently. The strongest model, Claude Opus 4.5, achieves only 57.38\% accuracy without external knowledge evidence and 70.34\% with it. This indicates that realistic text-to-MQL generation remains challenging.

Mental health stigma has profoundly harmful impacts but its complexity makes it difficult to evaluate. Stigma may involve explicit derogation, but also subtler forms of blame, fear, paternalistic pity, social distancing, structural exclusion, and discrimination. We introduce a theory-grounded benchmark for automatic evaluation of mental health stigma in online communication, consisting of naturally occurring online news and social media text annotated with a fine-grained taxonomy of stigma across multiple mental health conditions. Our annotation framework comprises a binary stigma-detection task and a multi-level taxonomy covering (i) stigma mode, (ii) domain, and (iii) specific components of certain forms of stigma. We apply this framework to texts mentioning six mental health conditions and evaluate large language models alongside stigma-related classifiers for detecting sentiment, toxicity, and hate speech. Results show that mental health stigma is not well captured by models trained to detect these neighboring constructs, and that LLMs often overpredict stigma unless given explicit operational rules - mirroring the importance of decision rules in human annotation. We release the publicly available part of benchmark, annotations, prototypical exemplars of stigma and code at: https://github.com/jemimakang/mh_stigma.

Standard factual benchmarks assume clean monolingual prompts and exact-match scoring, two assumptions that break simultaneously in real-world multilingual deployment, from retrieval-augmented generation pipelines returning mixed-language passages to users pasting multilingual web content. We introduce Multilingual Distractor Interference (MDI), a lightweight and fully replicable evaluation protocol requiring no new data or annotation, in which factual questions are preceded by a semantically irrelevant foreign-language sentence, and evaluate five instruction-tuned LLMs across TruthfulQA and TriviaQA under eight distractor conditions (40,000 evaluations). Our central finding is a metric confound: for Llama-3.1-8B under a Hindi distractor, 58% of responses switch to Devanagari script, yielding a raw hallucination proxy of 0.710, but manual review reveals that 120 of 148 script-switched responses that were correct under clean conditions remain semantically correct despite being written in the wrong script, reducing the adjusted semantic hallucination rate to 0.470. All other models respond through abstention escalation with no hallucination increase. A paragraph-length English distractor triggers near-universal abstention (0.806-0.998) across all models, consistent with reading-comprehension confusion, a failure mode with direct consequences for multilingual RAG pipelines. TruthfulQA multiple-choice accuracy is unaffected under all single-sentence conditions. These results show that exact-match hallucination rates in mixed-language settings should be decomposed into script-switching and semantic error components before drawing conclusions about model reliability.

Programmable logic controllers (PLCs) coordinate industrial equipment by reading sensor inputs and issuing control commands. Evaluating whether large language model (LLM)-generated PLC programs satisfy task requirements and safety constraints requires observing how their commands affect device and workpiece states. We introduce PLCWorld, a common closed-loop execution environment and benchmark that couples Structured Text (ST) execution with simulated plant responses and sensor feedback. Grounded in control relations identified in industrial PLC programs and engineering documentation, PLCWorld contains 100 synthetic tasks and 473 registered task-condition pairs across Motion Control and Material Handling, with difficulty defined by control-dependency scope. A common protocol reports Task Success and Safety Violation separately. Validation combines practitioner review, reference and alternative programs, targeted counterexamples, specification-evaluator alignment checks, and comparisons with independent ST runtimes. Reference and alternative programs satisfy their applicable cases, while all 542 targeted counterexamples activate their designated evaluator rules under at least one registered condition. Execution Gap relates submission-profile acceptance to subsequent task failure or observed Safety Violation. Across the constructed task groups, direct GPT-5.5 achieves 82.70% Task Success on Easy cases but 25.10% on Hard cases. Evaluations of six LLMs and four adapted generation-and-verification workflows further expose differences between completion, safety, and generation cost. Our code, simulation environment, benchmark tasks, and baseline implementations are publicly available at https://yunji0516.github.io/PLCWorld/.

Membership inference on large language models (LLMs) aims to determine whether a given text sample was included in an LLM's training data, without access to its training corpus. Despite recent progress, existing benchmarks suffer from three limitations: limited coverage of training stages, insufficient distributional alignment between members and non-members, and lack of rigorous filtering of non-members against the training corpus. To address these limitations, we propose OLMo-Detect, a multi-stage, confounder-controlled benchmark built upon the fully open OLMo 2 pipeline. OLMo-Detect spans pre-training, mid-training, and post-training, explicitly aligns members and non-members on three key axes, and rigorously filters non-members via infini-gram. To assess robustness to distribution shifts, we further introduce OLMo-Detect (Shifted), a variant where members are misaligned with non-members. We evaluate 15 unsupervised and 3 supervised membership inference attacks (MIAs) across the OLMo 2 family, finding that: (i) overall performance is limited: the best unsupervised and supervised MIAs both reach an AUC of only 0.68, and supervised MIAs degrade under cross-domain evaluation; (ii) MIA performance peaks at mid-training and is lower at pre-training and post-training, a pattern driven by data type rather than a stage effect: curated math data is far more detectable than other types; (iii) overall scores improve from 1B to 13B but plateau at 32B; and (iv) no unsupervised MIA is robust to distribution shifts, with AUCs shifting by up to 0.42. Finally, we find that our findings on OLMo 2 generalize to OLMo 3 and non-OLMo models.

Omni-modal large language models (OmniLLMs) unify text, images, audio, and video, yet hallucinate when generation relies on the wrong evidence. Existing inference-time methods can reduce hallucinations, but rarely reveal which evidence sustains a generated commitment. We introduce OmniConfess, a training-free method for mitigating omni-modal hallucinations. It fixes a candidate response and re-scores it at token resolution under controlled channel-wise evidence interventions, producing a structured token-by-channel confession that reveals the response's evidential dependence. OmniConfess uses this confession to preserve grounded content and correct commitments driven by irrelevant or contradictory evidence. To evaluate OmniConfess, we construct OmniHalluBench, a 3,540-example benchmark built from six datasets spanning text, image, audio, and video settings and both judgment and free-form generation. Experiments show that OmniConfess mitigates hallucinations across heterogeneous modality and task settings. Our code and benchmark are publicly available at https://github.com/RongHuiQiang/OmniConfess.

Long-term agents must remember not only what is true about a user, but also how a particular agent should work with that user as their shared history evolves. Existing benchmarks primarily supervise user facts and preferences or experience reusable across users, leaving this relationship-specific agent memory implicit. Additionally, most prior works measure the model solely with final-answer QA over long interaction histories, making the assessment still incomplete and unreliable. To this end, we introduce DyadMem with the proposed new definition User-conditioned Relational Agent Memory (URAM). DyadMem jointly annotates user-side memory and URAM along the same multi-session trajectories, resulting in 6 memory categories. To summarize, it includes 3,065 episodes, 50,961 sessions, and 61,210 QA instances, with extensive session-level Capture and Update gold annotations, query-level Recall support, and two QA settings: Gold-Memory and Full-Pipeline. Across 16 open-weight and 4 proprietary models, Gold-Memory QA is consistently strong, yet Full-Pipeline QA drops sharply. Such a gap explicitly supports our fine-grained evaluation design. Additionally, several quantitative results further reveal low Capture recall, incomplete Recall, and unsafe-deletion issues arising from even the frontier LLMs. We further conduct a rigorous experiment to validate the effectiveness of our URAM and observe the positive effects for all 20 models. In summary, DyadMem is a dual-domain, full-pipeline memory benchmark with extensive annotation efforts for advancing the domain's development.

We present WebFovea, a vision-based web agent that placed 2nd in the WebRetriever Challenge 2026 with a final score of 57.0 out of 100. The challenge evaluates agents end to end on Protocol III of the WebRetriever benchmark (arXiv:2607.06118): starting from an entry URL on a live website, the agent must operate the site's own interface and return a verifiable answer. A capable multimodal large language model (LLM) is necessary for this, but not sufficient. The model's decisions reach the browser through the harness, the code between the model and the page. At every step, four things must go right: the model's reply must be parsed into the intended action, the action must take effect on the page, the result must be reported back accurately, and the model must be shown the information it needs. On real websites, many of the failures we observed occurred at one of these four stages rather than in the model's reasoning. A coordinate-space mismatch placed every click at 3/4 of its intended coordinates; actions on native dropdowns, inside iframes, and in text boxes failed silently; and self-generated chat-template tokens contaminated 4.9% of task episodes. WebFovea hardens each stage and surrounds the loop with guardrails that keep the agent within the rules and its budget. The four-stage view does not depend on the model, although some individual fixes do. Because we used the same model in all four submissions, the rise of our official hidden-set score from 31.0 to 57.0 reflects changes to the harness, up to run-to-run variance on live sites. We describe the design, the evidence for each component (including negative results), a failure analysis, the limitations, and a roadmap that includes routing different steps to different models.

Large language models are increasingly used as reasoning components in AI-driven materials Co-Scientists, yet the reliability of the resulting verification pipeline remains unclear. Metal-organic frameworks (MOFs) provide a particularly challenging setting because structures may appear under different identifiers, synthesis outcomes depend strongly on experimental conditions, evidence is distributed across heterogeneous sources, and some hypotheses require computation rather than literature alone. We introduce a diagnostic benchmark with four task families covering structural grounding, synthesis-condition verification, evidence-sufficiency verification, and MLIP-based computational verification. T-MOF-1-3 are evaluated under closed-book, retrieval-enabled, and oracle-evidence settings to localize failures in knowledge access, evidence acquisition, and reasoning, while T-MOF-4 separately evaluates computational verification. Guided by these diagnosed failure modes, we develop MOF-Verify, a failure-aware agentic harness that targets structural, literature, evidence-sufficiency, and computational bottlenecks before producing a final verdict. Across multiple backbone LLMs, MOF-Verify substantially improves hypothesis-verification performance over direct inference and retrieval-based baselines. Benchmark datasets are released at https://github.com/IMMS-Ewha/MOF-Verify-Benchmark.

An LLM agent can recognize uncertainty yet still choose the wrong next step: asking when action is already justified, or seeking clarification when the constraints must change. We formalize actionable indeterminacy: act when an accepted action is shared across all admissible preferences or objectives, clarify when each possibility is feasible but no action is shared, and propose a minimum-cost permitted constraint repair when the request is infeasible. We construct a solver-grounded benchmark spanning object allocation, meeting scheduling, apartment choice, and stable matching. Matched pairs retain the same source while changing whether intervention is necessary, and evaluation separates decision correctness, matched-pair reliability, and fully correct responses. Our findings reveal a recurring difficulty in recognizing when intervention is unnecessary: models can identify situations requiring clarification or repair yet still intervene when a justified action already exists. Correct decision labels also fail to guarantee usable actions, questions, or repairs. Crucially, response requirements shape not only how decisions are expressed but also which decisions are made. Making the required content explicit substantially improves fully correct responses and can change intervention decisions, even when outputs are already parseable. These findings highlight that reliable agency requires more than recognizing uncertainty: it requires intervening only when necessary and translating the chosen next step into a verifiable response.

GPU kernels generated by large language model (LLM) agents can remain less efficient than expert implementations, but runtime alone does not reveal how the gap relates to design discovery and implementation. We introduce D2K-Bench, a diagnostic benchmark of 26 tasks and 85 workloads that measures how effectively agents translate expert design guidance into efficient GPU kernels. The guidance covers L1: high-level algorithmic insights, L2: dataflow design, and L3: low-level optimization tricks, including dependencies among these levels. Pairwise runs with and without guidance share task descriptions, workloads, tools, hardware, and a 350-turn budget. Complementary assessments examine independently proposed designs and the design properties implemented in generated code. Across five models on NVIDIA B200 GPUs, guidance raises correctness over 130 model-task pairs from 93.1% to 98.5% and increases the Performance Score over all 26 tasks from 1.46 to 1.95. For the three frontier models with correct submissions on all 26 tasks in both runs (GPT-6-Astra, Claude-Opus-4.8, and GPT-5.6-Sol), geometric mean speedup increases from $1.69\times$ to $2.49\times$. Across all five models, the mean combined implementation score increases from 57 to 70 out of 100. These results show the value of expert design guidance while identifying design properties that remain unimplemented.

We understand little about how capabilities acquired in one language carry over to another, or what governs this transfer: evaluations rely on incomparable, saturation-prone datasets and rarely examine its determinants jointly. Identifying what predicts transfer would let us avoid exhaustive evaluation across all language pairs and let developers target the factors that limit performance in low-resource languages. To evaluate cross-lingual capability transfer, we introduce Multilingual GSM-Symbolic, an extensible multilingual mathematical dataset covering 30,000 item-matched question-answer pairs and spanning 15 languages. It utilises symbolic templates to prevent overfitting and ensure generalisation by allowing generation of millions of high-quality variations from a single sample. Using Multilingual GSM-Symbolic, we quantify the largest determinants of capability as model size ($\beta = 1.77$), language resource level ($\beta = 0.77$), reasoning ($\beta = 0.67$) and typological distance ($\beta = -0.25$). This joint estimation allows these determinants to be expressed in terms of one another: a 32B model evaluated in Marathi performs like a 10B model in English. Our findings have important implications for model developers, showing that model size and reasoning narrow the performance gap between low- and high-resource languages ($\beta = -0.27$ and $\beta = -0.20$, respectively), while similar levers have little or no effect on typologically distant languages. Overall, our analysis framework explains 92% of between-language variation, but only 23% of the model-by-language variation, and predicts a model's performance on an unseen language within 6.0pp (r=.96). Incorporating measurements from just 10 templates in the target language reduces this to 4.19pp, enabling reasonable estimates of performance with little or no downstream dataset.

Typed decision models return choices or distributions over answer options supplied at request time. Accuracy with complete options does not establish whether a model recognizes that a reference answer is missing or avoids rejecting valid candidates. We present a paired candidate-coverage benchmark protocol and an initial evaluation of Laya and Jev across AG News, DBpedia, Emotion, and TREC. The models receive identical frozen texts and requests: 300 calibration and 589 test texts yield 23,932 predictions per model. Present/absent pairs match ordinary candidate count, and name variants preserve descriptions, members, and order. Native rejection behavior differs sharply: at five TREC candidates with natural names, Laya detects 97.2% of missing-answer cases but falsely rejects 69.7% of present controls; Jev's rates are 24.8% and 0.0%. Calibration-only none-score thresholds change these rates to 33.9%/3.7% and 45.0%/1.8%, respectively. On DBpedia, Jev's high coverage-score AUROC supports a stronger operating point, whereas both models have weak complete-set accuracy on Emotion. Competence-conditioned analysis, probability-precision sensitivity, and interface audits show why classification, score ranking, and rejection policies need separate measurement. This initial benchmark is descriptive and limited to reference-label omission; it does not establish natural out-of-scope generalization, causal mechanisms, or a new rejection method.

Hallucination detection is particularly important for medical language models, but repeated-sampling approaches are expensive and existing uncertainty-head resources do not directly transfer to a new backbone and language. We adapt the LLM Uncertainty Head (LUH) framework to Aya-Expanse-8B-based Persian medical models, using Gaokerena-V and Gaokerena-R as two previously developed backbones. We first examine response variability on a 168-question Iranian medical entrance examination and observe substantially lower five-run consistency for Gaokerena-V than for Aya-Expanse-8B, whereas Gaokerena-R is comparable to Aya-Expanse-8B. We then construct two paired claim-level hallucination datasets directly in Persian, containing 1,600 responses for each backbone, and train lightweight claim-level heads on frozen backbone attention maps and token probabilities. On held-out test splits, the heads obtain PR-AUCs of 0.4820 and 0.4652, corresponding to 2.30 and 2.66 times their respective random baselines, and ROC-AUCs of 0.7852 and 0.7810. The heads require neither retrieval nor repeated sampling at inference time. These results provide an initial study of single-pass claim-level uncertainty estimation for Persian medical language models; the test splits are small and the labels are automatically generated.

Financial AI agents must do more than retrieve facts: investment workflows require correct quantitative execution, reliable use of procedural resources, and auditable structured outputs. We introduce FinSkillBench, an evaluation suite of 2,603 point in time episodes across 12 subtasks in portfolio construction, risk management, and fundamental analysis, with hidden regenerable ground truth and task specific deterministic verifiers. Executing 17,820 episodes across 9 models and 3 resource conditions, the paired analysis across 8 models shows that curated skill packages raise mean scores by +16.2 points (0.366 to 0.528), whereas skills generated within a single episode add only +0.5 points while consuming more tokens and turns. We then decompose the curated premium by granting human authored procedural documents and executable domain tools separately: documents alone add +5.6 points, tools alone add +19.5 points, and their combination is subadditive. The premium is strongly workflow dependent: executable tools dominate numerically intensive workflows, documentation matters more when procedural or output schema guidance is the bottleneck, and interpretive tasks benefit from both. The effects are sign stable across 10 scoring variants and cluster bootstrap analyses, and an independently implemented second harness reproduces the directional pattern while showing that effect magnitudes depend on how tools and data are exposed. Overall, a measured "skill premium" is a property of the full model, resource, and harness system rather than of the underlying model alone.

We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging. Each question targets a concise, publicly verifiable answer whose discovery requires locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources such as videos, scanned documents, images, or maps. Easier questions are filtered out by evaluating them with models without internet access to reduce the likelihood that they can be answered with parametric knowledge alone. We evaluate several models using provider-native search and a shared external retrieval harness under a common agent protocol. To contextualize model performance and effort, we also conduct a human evaluation on a sample of the questions. HyperBrowseComp provides a challenging testbed for persistent information seeking across languages and evidence modalities, with difficulty arising from discovering and connecting evidence on the open web.

Designing a scientific instrument tests whether language-model agents can do physics rather than recall it, provided the grading cannot be argued with. We introduce NeutronGym, to our knowledge the first executable environment for neutron instrument design: agents build instruments through validating tools, McStas ray-traces what they build, and a level-resolved ladder grades syntax, runtime, structure and science with no LLM judge. Procedural families supply unlimited instances of a fixed layout whose design parameters the agent must set, with held-out parameter regimes; a curated slice, McStasBench, adds 16 tasks from published instruments behind memorization probes and a sandbox. Seven models reproduce at most 7 of the 16, none retrieves a reference, and none meets an improvement target. The environment also trains. Reinforcement learning on its reward takes Qwen3-8B from 11% to 77% of held-out instances of a family whose targets come from a hidden design (69% at a second seed), past an untrained Qwen3-32B, and the recipe holds, at one seed each, on three further gated families. The analysis says what that gain is. Without the ladder's partial credit it collapses by 60 points. From reward alone the trained model reaches what a classical optimizer reaches, at the agent's simulation budget, only when handed the closed-form physics (77% against 81%, a gap that does not separate at this size), while frontier models still solve 98-99%. Getting a trustworthy result meant failing four task designs that no-model baselines could solve, and we release the probes that found them.

Electricity price forecasting (EPF) supports scheduling, bidding, and risk management in electricity markets, yet existing benchmarks focus mainly on numerical inputs, leaving the forecasting value of forecast-time textual context insufficiently evaluated. We introduce TRACE, a reproducible benchmark of 7,300 zone--day instances pairing prices from five zones in a major U.S. market with official operational text available at the forecast cutoff. TRACE reconstructs official operational text at each cutoff, preventing post-cutoff information leakage. We evaluate TRACE for semantic alignment and forecasting value. Semantic assessments align with central movement and both tail risks in ground-truth prices, most consistently for upper-tail price risk. Forecasting value is reflected in a median 7.4\% reduction in upper-tail pinball loss across time-series foundation models. A controlled cross-day text-mismatch ablation reverses the gains, falling below the no-text baseline.

Large language models (LLMs) rely on long-horizon tool invocation sequences for complex tasks, where each invocation can alter the task state and condition subsequent decisions. In long-horizon tool use, final-outcome rewards provide weak credit assignment over long interaction traces. Step-level rewards can offer more targeted feedback, but obtaining reliable step supervision often requires human or LLM judgment, or additional rollouts to estimate the downstream effect of an intermediate decision. In this paper, we argue that effective tool-use agents should estimate the long-horizon value of a possible next tool invocation before executing it. This objective requires comparative supervision over alternative invocations under the same context, while logged trajectories only contain the invocation that was actually taken. Therefore, we propose Comparative Inference for Tool-use Agents (CITA). CITA trains a Comparative Inference Model (CIM) from paired signals that combine observed tool behavior, scalable supervision from a Bayesian tool-graph simulator, and semantic judgments from LLM-based comparison. The resulting CIM learns to estimate how likely a possible next tool invocation is to support final task success under the current context. Across three tool-use benchmarks and multiple backbone LLMs, CITA consistently improves Tool F1 and task success. Additional analysis shows that CIM learns accurate step-level value estimates for comparative tool choices.

Interactive world models are increasingly capable of generating environments and acting within them, yet deliberately editing an existing executable world remains underexplored. We formulate world editing as intervening on an existing world while preserving properties that should remain unchanged, and introduce intervention depth as an axis describing how strongly an edit couples world entities, dynamics, and systems. We instantiate this capability through industry-grade game modding and introduce IGMWorld, together with IGMBench, a benchmark of 110 tasks and over 1.1K executable state and behavioral criteria across Minecraft and Terraria. The tasks span property, entity, dynamics, and system interventions and are evaluated through deterministic executability, behavioral, preservation, and visual checks. Frontier coding agents already exhibit substantial world-editing capability: the strongest configuration solves 78.2% of tasks under a strict task-level criterion, while criterion-level performance reaches 94.8%. Reliability generally decreases with intervention depth, and this pattern persists even among tasks with similar numbers of evaluation criteria. Most failed edits still build and load successfully, suggesting that the main difficulty is making the edited world behave as requested. Visual consistency remains a separate weakness, with all evaluated configurations below 50% joint visual pass rate. These results show that world editing is a distinct capability from world generation and interaction, and that executable games provide a practical testbed for studying it.

Benchmarks for agent-generated data work grade a pipeline by running it once against a fixed snapshot. ArrivalBench instead re-executes the pipeline an agent leaves behind under adversarial but replayable delivery schedules (late, duplicated, out-of-order and retried records) and requires its final state to equal a batch recomputation of the complete log. Because the oracle recomputes rather than classifies, a wrong table and a crash are distinct verdicts: a crash is visible to monitoring a team already runs, and a wrong table is not. On 40 tasks we built, our reimplementation of single-execution grading certifies 86-100% of the pipelines eleven models produce; re-executing the same artifacts finds 7.0-79.2% of the certified ones silently wrong. The gap is not produced by the repair loop: within the same model and task, pipelines repaired against the snapshot test fail replay about as often as those that passed it first time. In every model, idempotency hazards fail more often than ordering hazards. Separating a wrong answer from a crash also changes how interventions read: a hazard warning cuts one model's silent failure from 48.2% to 10.5% while raising its crash rate from 9.0% to 37.0%, so all-in failure moves only from 51.0% to 44.0%. All eleven arms were independently re-run, and rates moved by at most 5.9 points.

Large language models often solve a theorem forward yet fail to disprove a closely related false one: a falsification gap that supervised fine-tuning does not close and can actively worsen. We frame counterexample generation as constrained witness emission against a deterministic per-theorem Python verifier, and release SymCE, a corpus of 4,707 false undergraduate-algebra and real-analysis conjectures, each paired with executable verifiers. The verifier also serves as the reward function, making SymCE a training environment. Training Qwen3-4B with SFT followed by GRPO under this oracle reveals an imitation trap: counterexample-only SFT collapses true-theorem recognition from 0.27 to 0.00, while RLVR with a sparse outcome-only reward repairs this and exceeds the base, to 0.66. The collapse replicates across four seeds and on Gemma-3-4B. Sparse and dense rewards yield statistically indistinguishable in-domain success yet diverge by 33 points on a held-out calibration probe, a dissociation we trace to the partial-credit term. Our 4B model outperforms every evaluated 7B open-weights math specialist, remains competitive with six frontier commercial APIs, and transfers under unchanged prompting to GSM8K, MATH-500 and MMLU-college-math. A human audit of 177 verifier decisions finds 97.7% accuracy. Code, data, verifier modules and annotations: https://github.com/ce-rlvr/SymCE.

LLM judges are increasingly used to assess whether AI outputs meet workplace requirements, but agreement on response rankings does not establish agreement on acceptance rates or occupational aggregates. We introduce O*NET-BENCH, an audit suite derived from an existing survey of 45,796 worker ratings, and evaluate 33 pre-existing judge configurations across six model families on 4,501 test ratings. Twenty-five configurations achieve tie-aware pair accuracy of at least 0.60, although a train-fitted response-only TF-IDF baseline nearly matches the strongest judge. Despite this ordering agreement, judges estimate that 3.0%-97.9% of responses are acceptable, compared with 61.1% for occupation-matched workers. In one fine-tuned lineage, changing from pointwise scoring to a bundled few-shot/listwise protocol improves response ordering while reducing agreement with worker means at the task and occupation levels; this reversal replicates on a task- and worker-disjoint validation split under prespecified criteria. Cross-validated calibration largely removes mean bias, but calibrated scores explain at most 8.5% of individual worker-rating variance. Prediction-assisted estimation yields at most small precision gains at the studied label budgets. These results show that ranking agreement alone is insufficient for occupational measurement. Judges should be validated against the acceptance rates and aggregates their scores will be used to estimate.

Deploying LLMs on consumer-grade GPUs with insufficient memory to hold their weights can result in prohibitively slow inference, because decoding repeatedly transfers offloaded weights from system RAM or flash storage into GPU at much lower bandwidth than local GPU-memory access. Activation sparsity reduces these transfers by skipping weights associated with zero or near-zero activations. However, as more activation contributions are omitted, model quality eventually degrades rapidly, indicating that weights associated with small-magnitude activations collectively influence model quality sharply. In this work, we improve the trade-off between model quality and decoding performance when exploiting activation sparsity. Our key idea is to replace the binary choice of whether or not to read a weight with three options: fully retain it, approximate it using a compressed weight representation, or omit it entirely. SpAx skips weights associated with activations closest to zero, reads approximate weights for smaller-magnitude activations, and reads original weights for the largest-magnitude activations. Smaller-magnitude activations attenuate the errors introduced by approximate weights, while compressed weight representations require fewer bytes to be transferred. With weights offloaded to CPU memory, SpAx speeds up decoding by 3.86X on average (up to 5.57X) with 16-bit weights and 2.06X (up to 2.74X) with 4-bit weights, at a WikiText-2 perplexity increase of at most 10%. With weights offloaded to flash storage, the speedups are 3.31X on average (up to 4.81X) and 1.54X (up to 2.03X).

Audio-Visual Large Language Models (AV-LLMs) remain prone to cross-modal hallucinations, where one modality incorrectly affects predictions about another. Although contrastive decoding reduces hallucinations in vision-language models, its direct extension to AV-LLMs overlooks a key challenge: different questions require different perceptual evidence, including audio, video, or their interaction. Notably, we observe that joint audio-visual inference can weaken the prediction even when a model can recover the correct answer from a single informative modality. For example, when asked which instrument is heard, a model may correctly predict violin from the audio alone. Once a video showing a guitar is added, its confidence in violin may drop. In this paper, we introduce Relevant Evidence Decoding (RED), a training-free method that identifies question-relevant evidence and selectively strengthens its contribution. RED uses pointwise mutual information to quantify the predictive support provided by audio and video beyond the question alone. It decomposes their joint contribution into audio, video, and residual interaction components. A question-only inference pass determines the required evidence type, after which the model augments the original audio-visual prediction with the corresponding PMI contribution. Across three audio-visual hallucination benchmarks and three AV-LLMs, RED improves accuracy over standard decoding by up to 7% on CMM, 6.3% on AVHBench, and 3.8% on SVHalluc, with an average relative time to first token of 1.5x standard decoding.

Scientific discovery often begins when scattered clues call for a new way of describing the world. Such abductive exploration can require constructing the representation in which an explanation is stated, when the world is epistemically open, and composing evidence scattered across contexts, when it is structurally interconnected. Existing benchmarks rarely separate these two demands or control them independently. We introduce ULTRADISCOVERY, an interactive world of five domains in which an agent revises an initially successful theory and predicts the outcome of an unseen cross-domain intervention. A $2 \times 2$ design leaves the representation open or discloses it, and leaves the evidence distributed or aligns it, with the latent dynamics fixed. With the representation open, agents across eleven models often retract the axiom they were taught, and none introduces the unobserved entity or rewrites the variables that a replacement requires. Disclosure triples intervention requests and adds about one of the eighteen findings the world affords, and alignment adds less. Two vendor-harness systems carry discovery into more domains, and one of them rewrites the variables in Open episodes. No system makes the exact prediction within 200 paid actions. At larger budgets one exact prediction appears with both aids, while every Open episode remains inexact. The results locate the difficulty in the step from accumulating evidence to composing it into a representation that transfers.

Accident, defect and outage investigations end with a decision that ordinary question answering never faces: whether the evidence gathered so far is enough to close the case. We study this decision for LLM investigators, which request evidence from a case file, revise their hypotheses, and either close the case with a conclusion grounded in what they read or leave it open and name what is missing. This judgment does not come with capability: an untrained 9B model overstates its evidence in 97% of its answers, and a frontier model that identifies the right cause in 84% of cases still overstates in 91% and closes 17 of the 41 cases whose official finding is "cause undetermined". Measuring it is also non-trivial: the source of a case largely predicts its label, and a rule that reads only the source reaches 83.0 balanced accuracy on our test cases. We therefore evaluate closure with three tests: closure accuracy, reported against this rule and within each source; evidence dependence, which removes the grounds of a conclusion and checks whether the model stops closing; and conclusion and gap quality, a judged checklist of what the model asserts and what it says is missing. We build Nautil, 731 audited cases from aviation, rail, maritime, chemical-safety and vehicle-defect reports and production server incidents, with teacher trajectories, an out-of-distribution test set and counterfactual evidence versions. Fine-tuning a 9B model on these trajectories makes its closures follow the evidence: removing the grounds lowers its closure rate by 26 points relative to a matched control, overstatement falls from 97% to 35%, and correct, non-overstated conclusions rise from 3% to 43%. Reinforcement learning that rewards only the closure decision then raises balanced accuracy from 69.2 to 83.3, on par with the teacher, and within-source accuracy from 60.4 to 74.1, at some cost in evidence dependence.

Tool-equipped AI agents use tool calls to access data and act on external systems. Horizontal growth of agentic systems increases the number of these interactions, and further motivates the need for automated, per-call oversight that can operate at low latency and/or on-prem. Conventional authorization schemes can determine whether an agent is allowed to invoke a tool, but cannot assess the agent's underlying cognition, specifically, whether the tool selection represents a logical, relevant step toward satisfying the intent of the task or not. Consequently, an allowed call may still deviate from the task's intent: a rogue agent might deviate the calls or nudge other agents to make a combination of calls that would not align with the intent of the task. Therefore, every call needs to be verified. In this study we investigate the applicability of Small Language Models (SLMs) to this purpose: an SLM functions as a task-tool relevance classifier that evaluates every selected tool independently against the assigned task and returns a relevance signal for downstream enforcement. Equipped with a novel dataset with multi-tool tasks whose required tools span distinct Model Context Protocol (MCP) servers, we used prompt-optimization, supervised fine-tuning, and reinforcement learning through GRPO to optimize and specialize SLMs.

A reasoning model asked whether a causal effect is recoverable from observational data can fail in two ways: it refuses an identifiable query or answers a nonidentifiable one. The latter is more consequential, as no observational data can validate the claimed formula. Measuring this failure requires queries that are provably non-identifiable, which prior evaluations lack, and grading that accepts correct formulas in any equivalent form, which string matching cannot provide. We build CERTID, a formal identification pipeline that addresses both limitations. CERTID uses the sound and complete causal identification algorithm ID to certify whether an effect is identifiable from a given graph and query, and verifies returned formulas against structural causal models whose interventional distributions are known exactly. CERTID further develops theoretical results to mitigate structural leakage, repair non-identifiable queries, and establish grading guarantees. We evaluate three frontier reasoning models (Gemini Flash, Gemini Pro, and GPT5.5) on 1,200 certified instances spanning 4 to 50 vertices. Accuracy proves a poor proxy for soundness: on identical instances, the false-claim rate on non-identifiable queries varies by seventeen-fold across models. We also find that models decide identifiability with 97-100% accuracy on graphs generated after the strongest model's training snapshot. Instances, the certification procedure, the verifier, and per-instance records are available at https://anonymous.4open.science/r/certid-D718.

We describe the Writerslogic team's participation in the CLEF 2026 SimpleText shared task, addressing Task 1 (text simplification) and Task 2 (complexity spotting). For Task 1, we develop a multi-candidate generation pipeline using GPT-4o-mini that produces five simplification candidates per sentence at varying temperatures, then selects the best candidate using a reference-free scoring heuristic that rewards compression, source word retention, Cochrane Plain Language Summary vocabulary usage, and lexical simplicity. On Task 1.1 (sentence-level simplification), our Claude Sonnet 4 submission achieves SARI 47.43 and BLEU 14.21, the top-ranked sentence-level system (3rd on the combined Task 1 leaderboard, behind two document-level submissions). For Task 2, we fine-tune a DeBERTa-v3-large NLI model on 350K labeled (source, sentence) pairs, framing hallucination detection as natural language inference. The model reads the most relevant source sentence as premise and the candidate as hypothesis, directly learning to distinguish grounded from hallucinated content. On Task 2.1 (binary overgeneration identification), our fine-tuned DeBERTa system achieves 0.8081 document-level macro F1 (0.8085 in our best ensemble), the top-ranked entry within the identification track and 2nd among teams overall, behind AIIR Lab (0.8197). On Task 2.2 (multi-class error classification), our best submission reaches 0.804 multiclass accuracy, ranking 2nd among unique teams behind AIIR Lab (0.827). We evaluate both tasks on English and multilingual biomedical text from Cochrane systematic reviews.

Understanding and assessing natural hazards is essential for disaster preparedness and risk reduction. Recent advances in large language models have spurred growing interest in AI agents for hazard analysis, particularly their ability to integrate scientific data, models, and tools into automated workflows. However, effective automation requires agents to determine which scientific methods are appropriate for a given event and executable with the available data and tools. As new evidence and execution results become available, these conditions can change, requiring agents to reconsider their choices. We formulate this problem as state-dependent scientific route selection and introduce HazardWeaver. Specifically, HazardWeaver first leverages the Hazard Knowledge Compiler to extract evidence-linked conditions governing scientific applicability, then its Hazard Capability Graph represents executable scientific capabilities and checks compatibility between their inputs and outputs. Using these complementary representations, the Hazard Weaver Agent component selects applicable and executable routes, carries out their workflows, and revises its decisions as the analysis state changes. To evaluate both the scientific outputs and the decisions that produce them, we introduce the Hazard Weaver Benchmark, comprising 141 instances across seven single-hazard domains and four multi-hazard interaction classes. The benchmark accommodates multiple valid scientific routes and evaluates output correctness, route validity, and justified abstention. Extensive experiments on this benchmark show that HazardWeaver outperforms existing agent systems, with the largest gains on tasks with multiple eligible scientific routes. Our code is publicly available at https://github.com/LabRAI/HazardWeaver.

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations. We introduce Queen, a 4B-parameter chess-language model that can explain its moves and plans while playing at the level of a typical Grandmaster. Our novel framework enables domain-specific reasoning through complementary components: an encoder-decoder architecture and an iterative distillation algorithm. This architecture integrates a silent expert chess encoder with an instruction-tuned LM through cross-attention, which we train via a question-answering curriculum to extract chess concepts from the encoder's representations. Building on this domain-adapted model, we iteratively improve its explanations with a natural-language analog of the Bellman update: the model analyzes the positions after its top candidate moves and consolidates them into an explanation of the current position, which is then distilled back into the model. Over seven iterations, our model gains over 900 Elo points (1782 to 2697), substantially surpassing all frontier models on both playing strength and puzzle accuracy, despite containing three orders of magnitude fewer parameters. Furthermore, LM-based evaluations show that our explanations are fluent and approach GPT-5.6-Sol (high) in coherence. The generality of our architecture and training procedure suggests a recipe for applying language models to domains where silent expert encoders are available, like games, robotics, and computer use.

A typed decision model answers a fixed question about an input by returning a probability for each of several caller-defined options. Each option carries a short label and a written definition, which is where a developer states the rule the model should apply. Jev introduced this interface for routing, moderation and triage, open implementations followed, and the same operation occurs whenever a language model is used as a classifier by scoring label strings. We study the open implementations, whose weights we can inspect and patch, and ask whether the probability follows the definitions or the labels. A preference for the label we call option-label bias. Across four open-weight typed decision models, three ways of reading an answer from a Qwen2.5 backbone, eleven classification tasks and PolicyBench, a synthetic routing suite we introduce in which the rule appears only in the definitions, the answer is mostly the labels. Deleting every definition leaves accuracy unchanged (laya-td: 0.8559 against 0.8487), although those definitions support 0.7971 on their own, and renaming the options to A and B raises accuracy by +0.1511 [+0.1377, +0.1646]. One system, von, is unaffected, and the two code bases differ in one expression: laya writes each option as "{label}: {definition}", while von writes only the definition. Changing that expression in both directions, with no weight changed, makes all three laya checkpoints exactly invariant (+0.0000 [+0.0000, +0.0000]) and creates the effect in von, whose accuracy falls from 0.8511 to 0.2281 when a label contradicts its definition. Earlier work attributed this failure to the constrained decision head these models use in place of a text decoder; our results locate it in the prompt rendering. We give a two-call test that tells a practitioner which case applies to their model, and measure what four mitigations are worth.

Despite the success of discrete diffusion language models (dLMs) for fast parallel decoding, their non-smooth, high-dimensional space hinders trajectory steering for reasoning and inference acceleration. To overcome this, we present Sigma, the first large-scale (3B/8B) continuous dLM built on steerable, low-dimensional ODE/SDE latent trajectories. Trained blockwise via likelihood optimization, Sigma jointly denoises Gaussian-corrupted token embeddings while learning an optimal embedding geometry. To accelerate training, Sigma leverages pre-trained weights from autoregressive (AR) models for warm-starting. During inference, we identify classifier-free guidance and score temperature as essential for high-fidelity reasoning and coding. Across comprehensive math reasoning and coding evaluations against state-of-the-art discrete counterparts (masked dLMs and AR baselines), Sigma achieves competitive performance with discrete models on standard benchmarks (e.g., GSM8K, Minerva, HumanEval, MBPP) after pre-training and on challenging reasoning tasks (e.g., MATH-500, AIME) after supervised fine-tuning. Beyond performance parity, we uncover key structural properties unique to continuous dLMs: (i) embedding-space steering effectively governs the quality-diversity trade-off, yielding strong pass@k performance and (ii) continuous trajectories enable graceful degradation for low NFEs and efficient distillation. These establish continuous dLMs as a promising paradigm for efficient language generation.

Egocentric videos capture how people carry out everyday activities, yet testing an agent requires evaluating the consequences of actions it chooses itself. We introduce Ego2World, a benchmark that turns annotated cooking activities into executable planning environments under partial observation. Its compiler links source steps and objects to symbolic action rules, persistent world states, and explicit task conditions, so researchers can execute an agent's proposed actions and check their outcomes. World state and agent belief are maintained separately, enabling controlled studies of planning and information reuse across continuing tasks. Evaluating six planners on 105 tasks shows that accepted operations often leave task goals unmet. Execution traces and condition checks distinguish interrupted runs, partial attainment, and completed execution without goal attainment. In a separate paired Qwen-Plus study, persistent belief improves action validity by 4.15 percentage points and reduces visual-query attempts by 90.27%, with higher token use and no detected completion gain. Ego2World provides a reusable testbed for tracing how planning and memory choices affect execution, observation demand, and task attainment, connecting recorded human activity to the development and evaluation of interactive agents.

Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose BitNest, a bit-nested speculative decoding framework that embeds a low-precision draft directly into the higher-precision target representation. Instead of deriving a draft from a predefined target, BitNest first constructs a strong low-precision base and then recovers the higher-precision target through residual refinement, enabling both models to share a single physical weight representation. BitNest further extends this progressive-precision design to the KV cache for long-context inference. Across multiple 7B--8B edge-friendly LLMs and diverse workloads, BitNest achieves an average speculative acceptance rate of 95.2% while closely preserving higher-precision model quality, and delivers 1.48--1.61x end-to-end speedup over FP16 autoregressive decoding. On the LLaMA models supported by all representative self-speculative baselines, BitNest also achieves consistently competitive or higher decoding speedup.

Current analyses of LLMs' parametric knowledge are largely output-centric, drawing conclusions about what a model knows without verifying what it was actually trained on. This leaves fundamental questions, such as whether a correct response reflects genuine generalization or rote memorization, grounded in speculation rather than evidence. To resolve these ambiguities, we introduce LUMOS, a diagnostic framework that traces knowledge along the causal chain from training-data exposure to behavioral output, leveraging OLMo 2 with its fully transparent training corpus. By grounding analysis in verified exposure, we reveal that models internally encode rare facts with high separability (84%) yet fail to express them behaviorally (54%), though this retrieval gap narrows with scale. Furthermore, when models are asked to self-reflect on their own answers, they perform reliably on trained content (83%) but drop to random-baseline levels (49%) on unseen content. This collapse persists even under chain-of-thought prompting, which inflates confidence signals rather than improving calibration. Collectively, these findings demonstrate that incorporating the training-data axis into LLM evaluation transforms speculative diagnoses into verifiable claims, and we advocate that this axis should be a standard component of knowledge assessment in LLMs.

Long-context LLM serving is increasingly bottlenecked by KV-cache memory, especially in resource-constrained scenarios. Among existing KV-cache compression strategies, token-wise methods reduce cached states but risk information loss through eviction or condensation, while feature-wise methods reduce per-token KV dimensions but can require full-dimensional reconstruction to apply positional embedding, limiting decoding speedups. We introduce SlimKV, a question-agnostic joint token-feature KV-cache compression method. SlimKV uses low-rank-aware training to compress long contexts into beacon memory states with latent KV representations, together with layer-adaptive rank allocation. We further uncover a positional asymmetry: removing key-side RoPE affects beacon and raw tokens differently, with much smaller degradation for beacon tokens. Exploiting this asymmetry, SlimKV trains beacon KV projections under a K-RoPE-free constraint and enables latent-space attention during decoding, mitigating reconstruction latency. On LongBench, SlimKV outperforms baselines at 16x/32x compression and remains leading at 4x/8x, where it retains over 96% of the uncompressed model's score. Needle-in-a-Haystack confirms robustness across evidence positions, and efficiency evaluation shows up to 7.34x attention speedup and 3.38x end-to-end decoding speedup over the uncompressed model at 128K length.

The key-value (KV) cache becomes a major memory bottleneck in long-context LLM inference, placing substantial pressure on memory capacity and bandwidth. To mitigate this bottleneck, vector quantization (VQ) has emerged as a promising approach for aggressive KV cache compression. However, existing VQ methods degrade substantially in the 1-bit regime. At such extreme compression, each codebook must represent a larger group of channels with a limited set of centroids, making effective use of its capacity increasingly challenging. To address this, we introduce $\textbf{TaSQ}$, which tailors the VQ target space by combining query-guided channel weighting, cross-head normalization, and covariance-aware channel grouping to better reflect the error sensitivity and statistical structure of cached activations. Since these transforms are RoPE-compatible and can be easily merged into projection weights and codebooks, TaSQ preserves the conventional VQ lookup structure and adds negligible serving overhead. Across general, long-chain-of-thought reasoning, and long-context retrieval benchmarks, TaSQ consistently outperforms existing low-bit KV cache VQ baselines while preserving reasoning stability. On a single RTX 6000 Ada GPU, its SGLang implementation supports up to $14\times$ larger batch sizes and achieves $1.87\times$ higher peak throughput compared to the BF16 baseline.

Reasoning traces improve large language models (LLMs), but current models are trained to reason mostly in English. It has been shown that forcing a model to reason in another language degrades accuracy, even when the reasoning language matches the language of the prompt -- but only for a setting where the model reasons over a short prompt. Here, we ask whether the same holds for retrieval-augmented generation (RAG), where the model must read and integrate a large amount of retrieved evidence in the target language. To study this, we build a fully monolingual German RAG question-answering testbed over the fictional world of the tabletop role-playing game The Dark Eye, a domain that is richly documented in German but too niche for the model to answer from memory, so that it has to rely on retrieval. Varying the forced reasoning language of an agentic RAG system on this testbed, we find that aligning the reasoning language with the language of the query and the retrieved documents helps. Forced German reasoning outperforms forced French, although the model benchmarks higher in French, so the benefit comes from alignment and not from language proficiency. The advantage grows when the retrieved context is richer and structure-aware. However, forced German only reaches the level of the model's native, unconstrained English reasoning without surpassing it, showing that native multilingual reasoning is needed. We publicly release the testbed and QA benchmark.

We study hindsight-guided distillation for rare disease diagnosis on ZebraMap: a 1.5B student is fine-tuned on chain-of-thought traces from a 8B teacher that observes the ground-truth diagnosis during generation. Absolute accuracy remains low for all models - the task is hard at this scale - but within this ceiling a filtered variant (StudentF) achieves a small, statistically significant accuracy advantage over the teacher (p < 0.001), concentrated in better-represented diseases. The unfiltered student does not significantly outperform the teacher (p = 0.129), establishing that contamination filtering - not hindsight distillation alone - drives the gain. The gap traces to an artifact we term GT hallucination. Label-visible generation causes the teacher to embed "ground truth is X" phrases in its reasoning chain; SFT copies the pattern. At inference, the unfiltered student reproduces the phrase in 33.9% of cases, with severe accuracy degradation when the hallucinated label is wrong. A regex filter removing these slots reduces contamination to near-zero, producing the observed gain - though the effect remains small. We precisely quantify this gain-cost tradeoff, document frequency-dependent knowledge transfer absent from the RL-trained teacher, and characterize a calibration gap that SFT does not close - identifying both as directions for future work.

Agent harnesses make many small, typed decisions per task: which model to call, which tool to use, whether retrieved text is relevant, whether an input carries an injection. System-1 decision models answer such questions in a single forward pass with class probabilities, promising large cost and latency savings over LLM calls. We present a paired evaluation of an open-weight (Laya) and a hosted (Jev) System-1 model on 11 agent decision points built from 18 public sources: 7,283 base cases plus 6,640 robustness variants, with byte-identical inputs, paired tests, and cross-hardware and cross-day reproducibility checks. Jev is significantly more accurate on 9 of 11 decision points (+10.8 to +46.0 pp). Neither model beats chance on zero-shot model routing, and they tie on RAG relevance gating. Laya changes 30% of its answers when the option order is reversed and degrades sharply with many or similar candidates (31% at 50 nearest-neighbour tools, vs. 98% for Jev on items with a unique correct tool). We also audit our own pipeline. Three analysis errors and one design confound distorted headline deployment claims: an omitted pre-screen cost (reported 23.9% saving, actual 4.3%), gate accuracy reported as end-to-end quality (58% vs. 98%), in-sample thresholds (5% target, up to 17% held-out misses), and a "channel effect" on injection false positives that vanishes with channel-native content. Two other suspected confounds did not change the conclusions. All cases, raw outputs and analysis code are available at https://github.com/David-DL-Space/sys1-eval.

Looped Transformers apply the same layers several times per token, adding compute to improve quality without more parameters. Each recursion, however, writes its own key-value cache, so memory still grows with compute. Inference-time techniques can shrink this cache at a cost in quality. We pretrain looped language models to share memory: only the first recursion writes a cache, and later recursions read it while keeping a short window of their own. Surprisingly, we find that sharing memory does not cost quality and instead improves it. At 150M-1B parameters, our Looped Prediction Transformer (LPT) and its hybrid variant set a new quality-memory frontier for looped models: with five recursions, the hybrid lowers validation perplexity on FineWeb-Edu by 1.12-1.82 relative to a same-size standard Transformer while using 76-79% less context memory. Through an extensive analysis, we investigate why memory sharing helps. Shared and local memory develop different representations, and later recursions attend mostly to the shared memory, which also acts as a gradient highway to the first recursion.

On-policy distillation (OPD) trains a student on the responses it generates. Existing LLM multi-teacher OPD transfers what specialists predict through their output distributions. We introduce Latent-MOPD, to our knowledge the first representation-level multi-teacher OPD method for LLMs. It integrates existing specialists through both their predictions and the hidden states used to compute them, without additional teacher training. To coordinate representation supervision from multiple specialists, we select late-layer targets according to the teacher-student relationship, bridge unequal hidden widths with a shared projection, and group updates by domain. Each teacher's supervision gradually shifts from hidden states to token predictions, with both channels using the same routed specialist. In our main same-family setting, Latent-MOPD outperforms the token-only, representation-only and uniform-averaging baselines on all nine benchmarks across math, code and logic. With the same parameter count as each teacher, the student also surpasses the per-benchmark best teacher on a majority of these benchmarks. With larger, separately developed cross-family teachers, Latent-MOPD outperforms both single-channel baselines on all benchmarks. A same-family all-layer representation-only control remains stable with domain-pure updates but collapses when teacher domains are interleaved within an update. Our results show that a single student can integrate capabilities from several specialists through both their output distributions and internal representations.

On-policy distillation (OPD) improves large language model reasoning by training students on their own rollouts with dense token-wise supervision from the teacher. However, token-wise OPD does not explicitly provide a coherent alternative reasoning step showing how the student's step could be revised to improve subsequent reasoning. Furthermore, this paradigm can become less effective when the student produces a degenerate reasoning prefix, as subsequent teacher supervision remains conditioned on that prefix and may reinforce poor reasoning patterns. In this work, we focus on learning reasoning revision with segment-wise OPD to rework intermediate reasoning steps and better support subsequent reasoning. Through controlled reasoning interventions, we find that replacing student segments with teacher redrafts improves subsequent reasoning accuracy. Therefore, we address the problem of turning teacher redrafts into explicit supervision for learning to revise reasoning. We propose Segment-wise On-Policy Distillation (Seg-OPD), which selects student segments based on an uncertainty metric and obtains corresponding teacher redrafts. Seg-OPD trains the student to prefer teacher redrafts over their paired student segments while retaining dense token-wise OPD supervision. Extensive experiments on mathematical reasoning and competitive programming tasks show that Seg-OPD-trained students achieve higher revision success rates than baselines. Seg-OPD consistently outperforms the compared state-of-the-art baselines in reasoning accuracy with an average relative improvement of 5.22% across diverse models and tasks. Code is available at https://anonymous.4open.science/r/Seg-OPD.

Language model agents are deployed with a harness, the software around the model that manages its context, tools, and feedback. When such an agent is distilled into a smaller one, the harness stays in place, so the student mainly needs the teacher-specific abilities that the harness cannot provide, such as acting correctly on harness information. Standard distillation, however, imitates the teacher's full outputs and treats the harness as part of the input. We propose Harness-Aware Distillation (HAD), which focuses distillation on what the teacher adds beyond the harness. HAD complements on-policy distillation with two components: an action preference that contrasts the same teacher's actions with and without the harness information, scored after the student's own reasoning, and a validity check that drops preference pairs whose preferred action contradicts the harness records. We show that the contrast gives the student information that imitating the teacher alone cannot provide, and HAD needs no task rewards, success labels, or future information. Across multiple long-horizon agent benchmarks and models, HAD outperforms on-policy distillation baselines with the same fixed harness. Our analysis shows that HAD enters fewer unproductive loops and recovers from errors more often than the baselines, and suggests that it adaptively keeps learnable feedback in its weights while reading state information from the harness.

Large language models (LLMs) can support rubric-based screening of CVs, but applying a high-capability model to every candidate and criterion is costly. We present CreateScore, a domain-theory-informed Bayesian network for criterion-level LLM routing. A hand-specified directed acyclic graph with Dirichlet-multinomial conditional probability tables converts CV evidence into posterior uncertainty; low-uncertainty decisions are resolved by a local 8B model and uncertain ones are escalated to a 120B reference model. The graph is causally motivated, but the system performs standard Bayesian conditioning, not causal inference. The escalation threshold is calibrated on a training fold (target: 70% resolved locally) and then frozen. On 200 synthetic Data Science CVs (139 training and 61 test candidates, five criteria), 77.7% of criterion decisions were resolved locally (237 of 305). Relative to a reference condition in which the 120B model adjudicated every criterion, routed escalation reduced token use by 65.2% and raised exact score agreement from 32.8% (8B alone) to 42.6% (95% CI 31.0-55.1%); at n = 61 the gain was not statistically distinguishable. The uncertainty signal did not, however, identify the decisions on which the 8B model erred: disagreement with the reference was 16.2% among escalated and 19.4% among locally resolved decisions (AUROC 0.47, 95% CI 0.39-0.56), no better than random selection. We also document how an earlier evaluation was invalidated when truncated reasoning-model outputs were silently replaced by local labels, and we recommend safeguards for cascade evaluation. CreateScore is supported as an auditable cost-reduction mechanism, not yet as a targeted error detector, and is not an autonomous hiring system.

What makes a short story gripping; a news article newsworthy; or a math proof elegant? These constructs resist articulation or verification; their meaning is at least partially tacit. However, modern AI models are improved primarily via articulated constitutions, rubrics and verifiers (i.e. in RLAIF and RLVR); tacit components of preferences are typically understudied. We introduce a large, labeled preference dataset CreativePreferences, containing 2.8M texts labeled by 317M human preference judgments across 7 creative domains, with 42 benchmark tasks. We model these labels with executable programs, rubric banks and densely trained models (V, A and VAT, respectively). We observe robust articulability gaps, VAT-VA; and verifiability gaps, VAT-V; we estimate upper and lower bounds for each gap with a novel measurement approach that discovers articulable and verifiable metrics, identifies spurious variables and estimates the value of undiscovered metrics using capture-recapture. These gaps occur across all domains, even in domains traditionally treated as fully verifiable: correctness-centered domains (i.e. mathematics and software engineering) and claim- and novelty-centric domains (i.e. news, patents, peer review). The size of the gap varies based on domain (e.g. peer review and creative writing have the largest articulability gaps) and widens as more people take part in the judgment, consistent with Collins' collective tacit knowledge. We show two consequences: (1) on human generations, the full model more closely matches human preferences, often in disagreement with articulated criteria, and (2) in an analogy to Goodhart's law, articulating preference shifts it away from the tacit dimension. Articulability and verifiability gaps are consequential; we give recommendations on when tasks can be prompted; how learning mechanisms might improve; and when to leave judgments with humans.

Complex reasoning queries can be decomposed into directed acyclic task graphs and distributed across heterogeneous LLMs, reducing latency through parallelism and enabling smaller models to solve complex tasks. In practice, however, the suitability of an LLM for a given subtask may be a priori unknown, and execution alone does not reveal output correctness. We propose JOVE, an online framework that jointly assigns executor LLMs and selects intermediate outputs for paid verification. Verification runs asynchronously and is used to improve future allocations, so the system must balance spending on execution now against learning for later. We study how to optimize this trade-off under a long-term budget and a per-query latency constraint, with stochastic, initially unknown LLM service quality, invocation costs, and execution times. JOVE makes execution and verification decisions by solving a sequence of per-query mixed-integer linear programs. Online learning updates task-dependent estimates of LLM quality based on verification feedback, while an information-gain bonus incorporates the value of learning into allocation decisions. Under a natural set of assumptions, we establish sublinear quality-learning regret for JOVE. Across four reasoning benchmarks, JOVE achieves competitive accuracy against standard inference baselines while reducing average cost and latency by at least 3.17 times.

Trajectory evaluation is essential for improving the reliability of LLM-based agents, but production use makes it expensive to run repeatedly. Modern agents generate long traces containing tool calls, observations, retries, and external outputs, while not all raw tokens are equally useful for diagnosis. We present \textit{LiteTrajEval}, a lightweight architecture for budget-bounded trajectory evaluation. LiteTrajEval derives compact domain-specific rule profiles offline, then preprocesses each trajectory online, marks heuristic failure signals, serializes it under a fixed global budget, and invokes a single rubric-guided LLM judge to produce structured diagnostic reports. Evaluated on public Magentic-One-style and $\tau$-bench-style trajectory datasets, LiteTrajEval improves failure-localization alignment with human annotations by roughly 20--35 percentage points on Magentic-One and up to 23 percentage points on $\tau$-retail compared with AgentRx, while reducing cost by about 6$\times$ and evaluation time by more than 8$\times$. This solution has also been deployed in our enterprise agentic platform.

Critic-free reinforcement fine-tuning (RFT) for agentic large language models is often done through GRPO-style methods, which compute a group baseline over repeated rollouts to reduce target variance. However, this setup is ill-suited to agents acting in stateful environments such as live services or security sandboxes, where repeated rollouts are impractical to obtain and aggressive updates entrench the noise of long, sparsely verified trajectories. We propose \textit{Follow the Winners} (FTW), a critic-free policy-learning algorithm that adapts the cross-entropy method to RFT, replacing group rollouts with an ordinal filter on replay-buffer samples that yields polynomial concentration in the order statistic of returns. We derive FTW through a control-as-inference lens, which also recovers GRPO and DPO as specific modelling choices, identifying GRPO as risk-neutral while DPO and FTW share a bounded risk-seeking offset that FTW controls. We identify this offset as an inherent trade-off of variance reduction through ordinal filters on samples, whereas a critic model induces a different trade-off between bias and variance. Scaled to agentic LLM post-training, FTW matches GRPO and PPO on Sokoban and Search-R1 baselines, showing a viable trade-off from a value model or group rollouts to CPU memory.

Large language models (LLMs) have reached expert-level performance on competition mathematics largely through the volume of search placed around them: candidate solutions are sampled in quantity and retained only when an external criterion accepts them. Such a procedure improves the outcome that survives it while leaving untouched what the model represents. We examine that question where no external criterion exists: translating statements between the dialects of neighbouring subfields, where fidelity turns on the level of generality at which content is asserted. The source leaves that level implicit in its vocabulary, so a faithful translation must recover it from the relation between the theories. We introduce an instrument that codes truth, content and scope in separate blind queues, with a judge-free measure of whether a rewrite states the hypothesis implicit in its source, and establish its sensitivity with a planted-positive control. Across seven models from four families, translating towards the general framing widens the domain of quantification in 60.6% of rewrites and narrows it in none; translating towards the concrete framing narrows it in 28.3% and widens it in 0.3%. The hypothesis that would prevent it is stated in 21.6% of model rewrites and 4.2% of human statements. Capability does not govern the asymmetry: it appears in every model tested, and the most capable widens least. It replicates on the half of the benchmark held out by a pre-registered rule, and on statements written by mathematicians. Instructing a model to state every hypothesis it requires raises that rate but not its sensitivity to direction. We argue that these systems have acquired an object-level correspondence between subfield vocabularies without the constraint under which a translation between theories carries hypotheses to hypotheses.

Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and shape much of the response. Most post-training recipes for dLMs do not use this signal to decide which tokens to train on: they typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that shape the response. We introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments (pivots). Pivot-SD selects pivots using an information-gain metric measuring uncertainty reduction over the remaining masked positions. Pivots from successful trajectories are trained with cross-entropy, and pivots from failed trajectories with targeted unlikelihood, leaving the rest of the failed trajectory untouched. Using only 200 questions and four rollouts each, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks.