2026-10-07

Large language models (LLMs) have the potential to meet a key goal in economics: a quantitative model of household decision making, across a variety of settings. Yet existing evaluations cover few surveys and outcomes, and do not study how households adjust to changing economic conditions. We introduce a new evaluation, HouseholdBench, which unites 6 U.S. household surveys and 32 prediction tasks spanning numeric, categorical and probabilistic outcomes, related to consumption, income, labor, expectations, and housing. Using past behavior, demographics and macroeconomic conditions, the tasks test whether LLMs predict behavior, including how households adjust to changes in various policies. We evaluate 13 proprietary and open-weight LLMs against a no-change baseline and a gradient-boosted tree model. Most LLMs outperform the no-change baseline, including for policy response tasks -- with the best model lowering error for numeric outcomes by 12.2%. Across most tasks, gradient-boosted trees rank first; leading proprietary LLMs approach their performance, but open-weight models lag. LLMs exhibit systematic over- and underprediction across different tasks. We identify methods that enable a 4 billion parameter open-weight model to match proprietary models' performance: fine-tuning and aggregating 16 predictions per observation. Improvements generalize to policy-response tasks, which are excluded from fine-tuning. We release our datasets, code, and leaderboard on our website: https://jn-huang.github.io/householdbench

Aerospace electrical-design revisions can contain multiple genuine changes, although an engineering request may authorize only a subset. Propagating every detected difference can therefore produce overly broad impact reports. We present LOGIC, a controlled benchmark and evaluation framework in which locally deployable language models ground a request in a deterministic candidate-change inventory before selected changes are propagated through a typed electrical traceability graph. This separation permits candidate-selection errors to be distinguished from downstream propagation errors. LOGIC contains 168 scenarios, including 144 selection and 24 abstention cases. We evaluate three 7--8B models against intent-agnostic, lexical, and structured-evidence methods, with an oracle-root upper bound. On 96 explicitly anchored selection cases, gate-only structured evidence achieves candidate F1 of 1.0000, compared with 0.9677 for token-lexical matching. On 12 relational-paraphrase cases, token-lexical F1 is 0.1772 and gate-only F1 is 0.0000, compared with 0.5000--0.6400 for the large language models. Model grounding degrades as candidate inventories grow from 4 to 64 changes, while affected-element and typed-path accuracy remain comparatively stable when frozen selections are replayed over graphs of approximately 1K to 100K nodes. Strict evidence gating suppresses false positives but can remove correct semantic selections. An exploratory evidence-empty abstention policy raises strict abstention accuracy to 0.6667 for all three models and reduces unsafe-report rates to 0.1667, while decreasing answerable-case coverage by 16.0--27.1 percentage points. Four of six conflicting requests remain unsafe for each model. These findings support combining literal evidence and language-model reasoning with engineering review when intent cannot be established reliably.

Real-world reasoning rarely reduces to static question answering: agents must actively gather information from tools and sensors that are often noisy and unreliable. Yet most existing active reasoning benchmarks assume that environmental feedback is trustworthy, or introduce noise without exposing an explicit, calibrated uncertainty signal, leaving open how LLMs should reason when the evidence itself is uncertain. We introduce VisualNoiseQA, a novel benchmark for active reasoning under noisy visual feedback. A text-only LLM must solve VQA problems by iteratively querying a fixed, off-the-shelf VLM treated as a stochastic visual sensor. For each query, we draw multiple samples and expose an empirical uncertainty signal via self-consistency, enabling the reasoner to probe from different angles and decide what to ask next and when to stop. Our construction is automatic and scalable: starting from diverse VQA sources and two noisy VLMs, we retain only questions where the sensor is inconsistent yet human-solvable. We evaluate multiple LLM reasoners on 1,000 instances spanning perception, chart understanding, and knowledge-intensive reasoning. VisualNoiseQA thus provides a controlled playground to study how different LLMs exploit uncertainty signals for robust reasoning.

Large language model agents are increasingly deployed to perform complex tasks in real-world environments. However, the knowledge required for correct behavior in these environments is often implicit, undisclosed, and subject to change over time. Recent continual-learning harnesses seek to address this challenge by enabling agents to improve from serving experience. Yet the effectiveness and limitations of these methods are not yet well characterized. Existing benchmarks provide only partial coverage: some explicitly provide the target knowledge, others assume a static environment, and those that support continual adaptation remain limited in scale and knowledge diversity. To enable systematic evaluation, we formalize an evolving-environment streaming dataset (EESD), in which agents must infer, apply, and revise latent environment knowledge from interaction and outcome feedback as hidden policies evolve, and introduce ServeLearnBench, spanning retail support, banking, and sales-pitch generation with 53 environment windows and 7,718 tasks. We evaluate five learning harnesses (RAG, Mem0, SkillOpt, Continual Harness, and Prime) across six models (GPT-5.6 Terra, Opus 5, Kimi K3, GLM-5.3, DeepSeek V4.1 Flash, and GLM-5.3 Flash), covering 28 model-harness pairs and 252 learning runs. Our evaluation reveals three main findings: a substantial gap remains between task capability and learning from experience; continual adaptation is costly and can degrade already-correct behavior; and insufficient exploration emerges as a key bottleneck to effective adaptation. Overall, ServeLearnBench provides a controlled testbed for diagnosing these limitations and tracking progress toward agents that continually and reliably improve through serving experience.

Large language models (LLMs) are expected to answer questions faithfully based on the provided context, abstaining when the context information is insufficient to answer the questions. Existing faithfulness evaluations typically assess each question-context instance in isolation; however, such instance-level evaluation fails to capture a fundamental requirement of faithful behavior: the ability to adapt model responses to changes in available contexts. In particular, a model should provide correct answers when sufficient evidence is present and abstain when it is not. In this work, we propose a Pairwise Faithfulness Benchmark (PFaithBench) that evaluates whether a model can switch between answering and abstaining for the same question under supporting versus non-supporting contexts. Our evaluations across thirty-nine models with seven model families demonstrate that faithfulness fundamentally involves a trade-off between answering and abstaining, and that most current models exhibit a strong bias toward answering, with most faithfulness errors arising from over-answering, i.e., models tend to fabricate a response even when the provided context is insufficient. We further conduct a series of studies on faithfulness training under different data constructions. Our results show that training outcomes are highly sensitive to the specific composition of answering and abstaining data. Constructing answering and abstaining data from mismatched sources can cause models to rely on dataset-specific shortcuts rather than actual context sufficiency. Moreover, increasing answer-supervised data improves answering performance but exacerbates over-answering, while increasing abstaining data reduces hallucination but leads to over-abstention. The code and data are released at https://github.com/tmlr-group/PFaithBench.

As large language model (LLM) agents become increasingly autonomous, they may pursue task performance through deception, raising concerns about their reliable deployment. Existing evaluations show that LLM agents can deceive, but often examine isolated scenarios or narrowly defined conditions, limiting systematic understanding of when deception becomes more likely. To address this gap, we introduce DecepEval, a benchmark comprising 1,532 instances across 3 task families and 28 professional scenarios. Drawing on classical fraud theories, we propose the LLM Deception Diamond framework, which characterizes four external conditions that may induce deception: pressure, incentive, opportunity, and conflict. DecepEval pairs neutral and induced versions of each instance to measure condition-dependent changes in deception rates, while explicit task facts and observable agent behavior help distinguish deception from capability-related errors. Evaluations of nine frontier LLMs show that inducements increase deception across models and task families, even among models with low baseline deception rates. DecepEval makes these vulnerabilities measurable, providing a shared benchmark for progress toward trustworthy artificial intelligence.

Learning from experience is essential for LLM agents to adapt to unfamiliar and dynmaic environments. Evaluating this ability is therefore important for understanding how effectively agents acquire and use new knowledge. Existing benchmarks have sought to evaluate this ability, but they primarily evaluate tasks whose rules are provided in the instructions or already familiar to pretrained models, making it difficult to distinguish learning from interactions from reasoning with existing knowledge. To address this, we introduce Learn2Play Bench, a benchmark of newly designed text-based games, whose rules are novel or counterintuitive, requiring agents to acquire knowledge through interaction rather than rely solely on pretrained knowledge. These games provide reproducible feedback and automatic scoring, enabling controlled evaluation of learning across repeated attempts. We also vary game instances to test whether agents can apply what they have learned to new situations. Therefore, we evaluate how backbone models, self-evolving methods, and agent harnesses affect agents' learning ability, revealing three findings: (1) Experience retention: Retaining complete records of actions and feedback can support more effective learning than summarizing these experiences into rules or strategies. (2) Human agent gap: Top-performing human players achieve higher peak scores than the evaluated agents. Human explore more varied strategies, and repeat actions less. (3) Harness matters: With the backbone fixed, changing the harness can improve performance while reducing estimated inference cost. Together, these findings provide insights into how LLM agents learn from experience and suggest directions for future work to improve their learning ability. Project website: https://liushiliushi.github.io/learn2play-bench-website/

Croissant has emerged as a standard for machine-readable dataset metadata, yet populating its fields remains labor-intensive and requires careful reading of accompanying dataset documentation. We present the first benchmark enabling end-to-end evaluation of metadata extraction aligned with a community-standard schema. The benchmark comprises 602 papers, including 102 with human-validated gold annotations and 500 with LLM-generated silver annotations, covering the full Croissant schema with both core and Responsible AI (RAI) fields. Using this benchmark, we evaluate a range of extraction systems spanning frontier models, open-weight models, and agentic architectures, under a two-tier evaluation framework that combines rule-based scoring with an LLM judge selected via human audit. We find that single-pass extraction consistently outperforms the four agentic architectures we evaluate: across backbones, these decomposed variants achieve lower accuracy than a single full-context pass. The largest gap appears on long-form RAI fields, which require synthesizing and interpreting information scattered across a paper rather than copying it from a single location, a setting where current systems remain far from reliable. We release the benchmark, evaluation code, judge audit, a live demo, and a leaderboard open to new systems.

Field-effect transistor (FET) layouts on exfoliated two-dimensional flakes are typically drawn by hand for each flake, placing contacts and gates to match its position and outline in optical micrographs. To our knowledge, no executable benchmark tests whether language-model agents can perform this flake-specific construction reliably. We introduce 2D-FET-Bench V2, a benchmark of 128 layout tasks built from microscopy-derived flake contours, including hole-containing flakes and multi-flake tasks. Each task supplies a textual device specification and contour coordinates. An agent generates typed polygon and path operations rendered to GDSII. A deterministic verifier checks geometric and structural requirements, and a separate integrity check verifies that the supplied contours remain unchanged. Scripted reference layouts pass all 128 tasks, showing that every task is solvable. We evaluate six models and seven workflow and scaffold variants of GPT5.6-Luna, with five attempts per task. The best-performing configuration in the six-model panel, GPT5.6-Luna with ReAct-3, passes 62.3% of attempts and solves 80.5% of tasks at least once (coverage) and 43.8% in all five attempts (consistency). ReAct-3 exceeds the one-pass Plan-and-Execute by 27.0 pass@1 points at 2.46 times the tokens. An expert audit of one sampled verifier-passing layout per covered task, across five ReAct-3 configurations, accepts 56.4% to 63.5% of them. The benchmark evaluates geometric and structural FET layout construction.

We evaluate whether AI agents can match the reported findings of published biomedical studies using public data. Existing evaluations do not consistently separate analysis from prior knowledge or retrieval of the published answer. We introduce BioStudyBench, a benchmark of 25 long-horizon analysis tasks drawn from studies first published between July and September 2026, after the developer-reported knowledge cutoffs of the models we evaluate, semi-automatically filtered down from 404,019 PubMed records. In each task, the agent receives a neutral research question but no data files, so it must find and download the relevant public data, search the literature through tools that return only records dated before its cutoff, and report findings through data analysis. To measure gains over prior knowledge, we run every task both with and without access to data and tools. Across eight models, access to data and tools raises the pass rate by 47 percentage points on average over the no-data baseline. Open-weight models across sizes trail closed-weight models, with the best open-weight model passing 81.3% of tasks against 94.7% for the best closed-weight model.

Activation steering provides a lightweight and flexible way to control large language model (LLM) behavior. However, effective steering requires more than inducing the intended behavior: it should also limit unintended changes and remain robust across inputs and training data. Existing evaluations cover these dimensions only in fragments. As a result, the trade-offs between efficacy and side effects have not been systematically characterized. We introduce SteerScope, a two-axis, multi-dimensional evaluation suite that jointly characterizes steering outcomes and method properties through 15 metrics. We score target efficacy and side effects on language quality, task capabilities, and safety and reliability, and further assess generalization and data dependence through steering-specific metrics for sample efficiency and sample sensitivity. Rather than comparing methods at a single operating point, we characterize the trade-offs between efficacy and side effects. Under matched models, tasks, and evaluation protocols, we benchmark 23 methods spanning 4 families, including prompting, LoRA, and SFT as baseline methods, and release the suite as an extensible codebase. We find that current activation steering methods do not yet surpass the Prompt Steering baseline in their overall balance between steering efficacy and side effects: across both model scales, no evaluated activation steering method achieves higher efficacy without incurring greater composite side effects. We further uncover a consistent coupling between steering efficacy and side effects. Under OOD prompts, target efficacy is often preserved, whereas side effects tend to become more pronounced, particularly through declines in instruction relevance and fluency. Methods also exhibit sharply different sample-efficiency profiles.

Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand. We study where this evidence-to-action chain breaks as agents move from deciding whether to act to executing single actions and dependent workflows. Across ten model-harness configurations, strong static action assessment can coexist with much weaker interactive execution. Failures often begin before execution: agents stop with incomplete investigation or act before required evidence is established. Once required evidence is obtained, single-action execution is usually reliable, while multi-action workflows additionally expose unresolved prerequisites and incomplete execution. For this analysis, we introduce SafeActBench, comprising 656 cases across six operational domains and five protocols that progress from static action judgment and investigated non-action to single- and multi-action workflows. A provenance-bound Evidence Ledger and deterministic trajectory evaluator track what information was established, when actions occurred, and whether downstream dependencies were satisfied. These results show that failures arise not only from missing information, but also from how agents use established evidence when deciding and executing actions.

The rapid progress of LLM-based multi-agent systems (MAS) has shown that they largely outperform single agents on coding, math, and QA tasks, where executable tests provide a binary success signal. Whether this advantage transfers to real scientific data analysis remains untested. We introduce ST-Bench, a benchmark designed to answer two questions: whether MAS outperform single agents on complex scientific data analysis tasks, and if so, by how much and at what additional cost. ST-Bench contains 100 data science tasks adapted from published Earth science studies across hydrology, agriculture, and wetland methane research, expanded into 2,067 queries grounded in additional published studies and validated by domain experts. Using ST-Bench, we evaluate five recent MAS generation methods under two training protocols, against single-agent baselines on the same GPT-5 backbone. Nine of the ten MAS configurations exceed the cheapest single-agent baseline, with the strongest reaching nearly three times its composite score. This gain is primarily attributable to coverage: trained workflows produce realistic numerical metrics on a larger fraction of queries, while the quality of those metrics, conditional on producing realistic output, is comparable to that of the single-agent baseline. The strongest configuration requires approximately four times the single-agent inference time, whereas a more economical workflow captures the majority of the benefit at less than twice the cost. MAS specialization confers measurable benefit on scientific data analysis, but the benefit is conditional rather than universal.

As LLMs increasingly assist in moral reasoning, omission bias, the tendency to prefer inaction even when equivalent framings reverse substantive outcomes, poses a significant risk of skewed decision-making. Yet omission bias remains underexplored in LLM evaluation, with the few existing studies limited in scale and focused largely on utilitarian-deontological conflicts. To address this gap, we introduce OMIT, a benchmark consisting of 218 paired-frame scenarios across 10 conflict types, constructed by leveraging disagreement patterns from an LLM-based, five-perspective philosophical persona panel (utilitarianism, deontology, virtue ethics, care ethics, and contractualism). Evaluating eight LLMs, we find that omission bias is pervasive but inversely correlates with model size within families. We further evaluate four inference-time interventions and find that interventions encouraging models to consider moral principles before committing to a yes/no answer reduce omission bias and increase frame-consistent responses, although lower omission bias rates can also coincide with shifts toward action-biased responses. Ultimately, this work contributes not only the OMIT benchmark, but also a methodology for using diverse philosophical disagreement signals to evaluate framing-sensitive inaction preferences and the distributional effects of mitigation attempts in LLMs under complex moral conflicts.

Frontier LLM agents have been shown to be capable of solving increasingly complex tasks for which humans have measurable solutions. This begs the pertinent question of whether LLM agents can go beyond what humans have already solved. The ability to develop sophisticated strategies to tackle consequential problems becomes paramount as well-trodden, human-developed solutions become insufficient for problems for which we lack context or enough training data. We study agents' capability of such strategy formation through the communal practice of video game speedrunning. In speedrunning, practitioners compete to find the fastest way to complete a video game under certain conditions, and in so doing uncovering interesting unorthodox play styles that require a thorough understanding and mastery of the underlying game mechanics. We introduce SPEEDRUNBENCH, a benchmark that evaluates frontier LLM agents across 9 different games. To perform well in this benchmark, agents must repeatedly improve their strategy, reflect on their performance, exploit their gained knowledge, and reason across a long-horizon of actions to improve on an increasingly difficult problem: being faster than themselves and everyone else. Our experiments show that while frontier agents approach human world records in simple platformer games, they remain behind human performance on longer, more complex games under practical budgets. These results suggest that SPEEDRUNBENCH is a useful testbed for studying agents' strategy formation capabilities as well as being a saturation-resistant evaluation measure, as there is almost always a faster completion time waiting to be discovered.

LLM tool-use agents operate in dynamic environments where many actions carry operational risk. However, most safety mechanisms react only after errors manifest. Existing pre-emptive approaches either fine-tune the agent on chain-of-thought deliberation or compile natural-language guardrails into runtime checks, but they do so without exposing a structural, auditable verdict. We propose POLAR, a guardrail framework for small tool-calling agents that assesses reversibility through a structured two-layer ontology. POLAR assigns each action a graded reversibility score by deriving a candidate inverse sequence; calls failing a threshold are pruned before execution. Evaluated on $\tau^2$-bench across six agent models, POLAR improves mean task reward by 0.11 to 0.18 points on airline for four of six agents, but only eight of eighteen model--domain cells improve overall; retail and stronger agents often regress. POLAR provides an auditable structural check and characterizes its task-utility trade-offs. Reward is not a direct measure of prevented harm.

Conversational MLLM agents are increasingly expected to assist in professional workflows, from AI research and engineering design to product management and business operations. Yet this capability remains underexplored: existing benchmarks largely focus on informal, everyday interactions and personal-life scenarios featuring photographic natural images, isolated static artifacts, and recall-oriented questions. In contrast, professional scenarios often involve structured, information-heavy artifacts that undergo frequent revisions and authority updates, and compositional queries requiring reconciliation of many artifact versions while tracking state precisely. To address these challenges, we introduce DSV-Mem, a benchmark for evaluating Dense Stateful Visual Memory. DSV-Mem comprises expert-reviewed scenarios and 1,000 questions across five user-oriented categories (Current State, Past State, Derived State, Change History, and Conflict/Refusal). A Hartley-inspired criterion favors questions with broader visual-evidence inspection demands. We also introduce a generation harness that produces evaluation suites by decoupling state-transition synthesis from conversation filling. Evaluation over 27 configurations spanning frontier and open-weight models and memory management methods reveals that the strongest baseline scores below 45% on DSV-Mem. Analysis surfaces findings: 1) multimodality and information density both contribute to difficulty, but state evolution, particularly the number of governing updates, is the dominant tested factor. Raw conversation/haystack length, OCR, and arithmetic are not the primary bottlenecks; 2) models often fail to verify user premises against prior state updates before answering; 3) increased reasoning effort and memory management methods yield limited gains, whereas state-aware designs prove more effective. The benchmark and code will be publicly released.

Large language models are increasingly used to simulate clients for counselor training and psychological counseling research, but reliable simulation requires clients to remain psychologically coherent across extended interactions. Existing role-playing methods largely rely on static profile prompts and may exhibit persona drift, unrealistic cooperativeness, or inconsistent psychological states, communicative actions, and emotions. Existing evaluations also lack a unified testbed for both stable client characteristics and evolving psychological dynamics. We propose MASC, a Multi-Agent Self-Calibration framework with latent construct alignment for consistent client role-playing in psychological counseling. MASC combines construct-guided generation, collaborative refinement, consistency verification, and memory-based revision in a closed calibration loop that detects and corrects inconsistencies as dialogue unfolds. We further introduce CRPC-Bench, a benchmark covering session-level profile information and Big-Five personality traits, as well as turn-level psychological state, communicative action, and emotion expression. CRPC-Bench contains 38 motivational interviewing client profiles augmented with personality and emotion annotations. Experiments show that MASC outperforms existing methods across profile, personality, receptivity, and turn-level consistency, with the heterogeneous configuration achieving the strongest overall performance. MASC and CRPC-Bench provide a unified foundation for developing and evaluating psychologically coherent client simulations for AI-assisted counseling research and training.

Large language models routinely answer questions that cannot be answered from the information given, and in dialogue they answer before enough has been said. Unanswerability is linearly decodable from hidden states, but it is unclear which of its forms share a representation and whether the signal is useful in dialogue. We contribute a turn-labeled multi-turn benchmark (423 conversations, 1,661 labeled turn-states) and an evaluation harness with a simulated user who answers clarifying questions, and use them with six datasets and six open-weight LLMs to test how far probes for unanswerability carry. Probes transfer robustly between datasets that share a ground of unanswerability: missing information in math (AUROC 0.77-0.97) and in a passage (SQuAD 2.0<->MuSiQue, 0.77-0.90). Probes for epistemic "known-unknowns" transfer poorly to math, but this separation weakens under lexical controls and changes with layer and coordinate system, so it remains unresolved. Single-turn probes fail zero-shot to detect when a conversation becomes answerable; in-structure probes recover it, but no better than a bag-of-words classifier. A gate on the calibrated probe, with no model fine-tuning, fires on underspecified turns far more precisely than chance, and its end-task success comes within 0.08 of a gate given the true labels. Yet across four models it does not reliably beat vanilla generation or prompted consolidation. The remaining gap lies mostly in how models use a clarification, not in detection.

LLMs are increasingly deployed as autonomous agents in social environments, making it critical to study their ability to faithfully simulate human interactions. Central to this is grounding agents in realistic user personas, yet existing datasets rely on fictional personas and are limited to a handful of languages, lacking the empirical grounding necessary to evaluate behavioral fidelity across diverse populations. We introduce Wiki-Talkie, a multilingual dataset of real-world conversations from Wikipedia Talk pages across five languages spanning two language families: Germanic (German, English) and Romance (Spanish, French, Italian), paired with personas derived from real user communities and encompassing sociodemographic attributes, self-descriptions, and behaviorally grounded interaction traits. Using Wiki-Talkie, we evaluate agent interactional behavior on a next-turn generation task across various persona conditioning strategies. Our evaluation assesses whether agents collectively reproduce the distributional behavioral patterns observed in human discussions. Results show that user's comment history exemplifying interaction behavior consistently outperforms explicit persona information. In addition, models systematically underproduce negative or extreme sentiments, while over producing references and suggestions, revealing biases toward agreeableness and positivity. Crucially, these patterns hold robustly across languages, with small cross-lingual differences.

As coding agents increasingly undertake real-world work autonomously, judging whether their risk treatments are warranted has become important. Existing work evaluates related agent behaviors from separate perspectives, but lacks a systematic framework for unifying these behaviors. To bridge this gap, we introduce ParanoiaEval, the first benchmark for unified evaluation of risk-treatment capabilities in coding agents. Grounded in the well-established Avoidance-Transfer-Mitigation-Acceptance framework in software engineering risk management, ParanoiaEval operationalizes its 4 fundamental treatments for coding-agent settings and contains 200 evidence-controlled repository-level task pairs, each differing only in treatment-defining evidence. We further introduce dedicated metrics for risk-treatment violations and evidence responsiveness, using a human-calibrated agentic judge for reliable evaluation. Large-scale experiments on 8 representative models and a post-hoc human study reveal that (I) unnecessary risk treatment occurs in 11.2%-58.7% of runs despite explicit evidence, with substantial variation across agent configurations; (II) stronger task capability does not ensure more appropriate risk treatment, while treatment violations substantially harm developers' experience, establishing risk treatment as an independent capability dimension; and (III) agents exhibit systematic patterns consistent with established risk-management findings, suggesting that knowledge from human practice can guide the diagnosis and improvement of this capability.

LLM-based agents are increasingly advancing scientific and engineering problem solving, with physics simulation emerging as a challenging yet practical testbed for reproducing complex physical phenomena with application in embodied AI, games and films. As the workhorse of such simulation, a solver computes how the state of a dynamic system evolves over time. Building such solvers requires physical understanding to identify appropriate models, mathematical reasoning to formulate the underlying dynamics, and software engineering to implement them as executable code, yet this capability of LLM agents remains underexplored. To this end, we introduce WorldSolver, a benchmark of 168 simulation tasks derived from physical phenomena in 61 classic computer graphics papers, spanning 7 physical domains. Each task contains a code scaffold that provides a fixed simulation environment for the scene, with the solver implementation left for the agent to complete. Specifically, we evaluate them along three dimensions: Execution Checks for successful execution, Visual Fidelity for reproducing the intended dynamic behavior in the rendered simulation, and Physical Plausibility for physics-grounded verification of the generated dynamics. Experiments on frontier agents reveal that producing executable solvers is difficult itself, and satisfying visual and physical correctness is even harder. GPT-5.6-Sol and Claude-Opus-5 perform comparatively better than the other evaluated agents, yet achieve overall scores of only 48.7% and 46.7%, respectively. WorldSolver is an early step toward agentic solver generation, and we hope it helps drive progress toward agents that can faithfully simulate the dynamic physical world. Code is available at https://github.com/sirujiang/WorldSolver.

Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Adversarial training lets the attacker adapt but keeps the tasks fixed, so a task stops teaching once the agent solves it. We introduce AdvSim2Real, which co-evolves a task curriculum, an injection adversary, and the agent inside a frozen web world model. The curriculum is rewarded for tasks the agent solves about half of the time, and the adversary only for a success flip, an injection that turns a judged success into a failure. Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. On 150 web tasks, AdvSim2Real raises completion under this unseen adversary by 33.6\% relative to the base agent.

Large language models (LLMs) can solve many narrow tasks, but querying them separately for millions of related instances can be prohibitively expensive. Can LLM agents autonomously create cheaper solutions for such workloads? We call this ability "bottling": the ability to turn general capabilities into task-specific solutions that balance answer quality and amortised cost. We introduce BOTTLED, a benchmark in which agents receive an entire unlabelled workload and must complete it under fixed time, compute and LLM API budgets. Agents choose their own approach, such as training a small model or writing a reusable program. Across ten models and three tasks, we find that strong zero-shot task performance does not reliably translate into strong bottling capabilities. Models with similar zero-shot scores can differ substantially after bottling, and 48 of 60 bottling runs score below the lower bound of the 95% confidence interval of their model's zero-shot performance. Moreover, 31 of 60 runs underperform the stronger of two small-model distillation baselines with the same token budget. Nevertheless, bottling can yield substantial savings: on query-product relevance classification, Opus 5 retains about 82% of its zero-shot macro-F1 at roughly 657 times lower reported cost. Bottling is also competitive with Jev, a "system one" model built especially for cheap, repetitive inference: Opus 5 on the same task recovers about 94% of Jev's macro-F1 at a quarter of Jev's projected full-workload cost. BOTTLED provides a basis for evaluating and improving agents' ability to invest limited resources in reusable solutions for large, repetitive workloads.

We study the application of large language models (LLMs) to the visual exploration of textual corpora. We introduce zero-shot visualization (ZSV), a task in which users specify concepts in natural language and documents are mapped onto the corresponding concept axes for visualization. Building a ZSV system of practical value is non-trivial, as it requires choices at the intersection of feature functions, efficient implementation tradeoffs, and pre/post-processing decisions affecting visualization quality. To that end, we establish a benchmark that compares methods spanning embedding similarity, direct semantic judgments, and conditional likelihood estimation in this setting. Across multiple datasets and use cases we evaluate the properties of different scoring methods and design choices in terms of semantic faithfulness, score fidelity, and computational cost. Our results identify that scoring based on next-token probabilities offers the strongest practical trade-off among the evaluated methods. We further apply this approach to unlabeled corpora to examine its behavior in realistic exploratory settings. These experiments highlight additional design considerations, including the use of graded axes together with binary relevance filtering, and reveal a compositional sentiment bias in off-topic documents. Based on these findings, we provide practical guidelines for constructing end-to-end ZSV baselines.

Recent advances in Large Language Models (LLMs) have demonstrated remarkable capabilities in web front-end execution, with browser-based game generation emerging as a particularly prominent frontier. While previous efforts frequently rely on complex multi-turn workflows or focus on static game evaluation benchmarks, this work targets direct end-to-end real-world game synthesis driven by coding agents. However, generating complex games directly from sparse user queries often forces coding agents to make underspecified assumptions, yielding incomplete mechanics, disconnected gameplay flows, and limited visual aesthetics. To resolve this issue, this paper presents GameGo, a scalable framework that systematically transforms brief game seeds into comprehensive Product Requirements Documents grounded in industry game-development practices. To retain core gameplay constraints without restricting design exploration, GameGo uses task-specific dynamic compression to maximize information density while preserving instruction following. Based on this pipeline, GameGoData is constructed with 55,060 development trajectories across 2D, 2.5D, and 3D games, alongside GameGoBench, a benchmark comprising 124 diverse game queries. Training GameGoCoder on GameGoData yields a model that outperforms matched baselines and is comparable to frontier models across gamedev benchmarks. All code, datasets, and models will be made publicly available.

Block diffusion language models keep a large key-value (KV) cache throughout generation and attend to it at every denoising step, limiting both memory capacity and generation speed. Reducing these costs requires deciding which past tokens to use for denoising the current block (selection) and which to keep in memory for future blocks (eviction). We propose MaskAhead, a training-free method that solves both tasks with a single mask-query-based ranking mechanism. Current-block masks guide selection, while probes of upcoming masked blocks guide eviction. Both rank KV entries by their estimated contribution to the attention output. Our quantized variant, Q-MaskAhead, computes selection and attention directly from low-bit KV, largely preserving the selected entries. Experiments on Fast-dLLM-v2, DreamReasoner, and LLaDA2.0-mini cover long-generation reasoning, long-prompt question answering, and needle-in-a-haystack retrieval. On long-prompt QA, MaskAhead reduces KV memory by $9.5\times$ on average with a 1.2-point mean F1 loss relative to dense inference. Q-MaskAhead increases the reduction to $20.1\times$ with a 2.3-point mean F1 loss. In a batch-32 systems profile, MaskAhead achieves $1.23\times$ end-to-end and $1.68\times$ decode-stage speedups over dense inference.

The Offline AI Modules workstream enables practical, low-power, and community-accessible deployment of voice-first AI systems that operate fully offline. Designed for African language communities where speech is the dominant mode of interaction and internet connectivity is unreliable or absent, the workstream delivers three reinforcing components: a modular voice-first offline architecture, a low-cost hardware reference bill of materials, and a reproducible quantization and a reproducible quantization and benchmarking pipeline for instruction-tuned language models in the 2-5B parameter class. This paper presents the first end-to-end benchmark evaluation of the stack across two hardware tiers: an NVIDIA Jetson Orin NX (TierB) and a Raspberry Pi5 (TierA). Three instruction-tuned models are evaluated across four quantization formats, assessed for deployment metrics (decode throughput, chat latency, memory, power) and multilingual quality (topic classification accuracy on MasakhaNEWS across English, Hausa, Igbo, Nigerian Pidgin, and Yoruba; per-language perplexity drift). Speech recognition is evaluated using Ethio-ASR on Amharic and Oromo across both tiers. The principal finding is that Q4_K_M quantization represents the best size-to-quality trade-off for deployment on both tiers: gemma-4-E2B-it achieves 28.8t/s decode throughput and 89.2% topic classification accuracy at Q4_K_M on TierB, while all three models run within the 16GB memory budget on TierA.

Modern language-model agents increasingly interact with external environments over long-horizon, multi-step trajectories, where the accumulated interaction history can quickly exceed practical context budgets. To ensure reliability, agents must maintain factual information over long horizons, remember execution errors and corrective feedback, and track progress across actions. Several approaches have been proposed to achieve this without the need for maintaining the entire execution history in context, such as using the reasoning and action history, learning to maintain a fixed-size memory through an overwrite mechanism, and periodic summarization. Although overwrite memory can in principle retain anything an append-only memory can, it must learn to carry each fact through every subsequent rewrite, which is difficult to learn from sparse outcome rewards; for interactive applications like web agents, we find that trained overwrite memories delete key information required by the trajectory, as well as corrective feedback received from the environment. We introduce AMBER (Append-only Memory Bank for Evidence Retention) - a simple and scalable framework where an agent jointly learns to reason, act, and write free-form memory, while an append-only rule guarantees retention by construction. This allows AMBER to be trained end-to-end with reinforcement learning from outcome rewards without the need for extensive curated SFT data. On WebArena Lite, AMBER improves average success over overwrite-based memory by 4.09 percentage points, increases the fraction of tasks solved in five repeated runs by 4.8 percentage points, and matches an overwrite baseline trained on substantially more expensive curated supervision. AMBER achieves these improvements while maintaining a practical token budget, providing a strong balance between context efficiency, task performance, and reliable long-horizon execution.

A team of coding agents can look fine agent by agent yet fail as a team: each passes its own tests while the merged result is broken, and single-agent evaluation never catches it. As teams run several LLM coding agents in parallel on one codebase, the agents collide: two rewrite the same function, one codes against a contract a teammate just changed, and integration fails after the work is done. Most coordination tools react (watch for a conflict, then warn), but at agent speed the warning arrives after the wasted edit. We recast the problem as scheduling: take each work item's declared scope, partition the work into disjoint scopes, and order merges along the producer->consumer graph, all up front. We build this planner into Nerveplane and evaluate it with NP-Bench, an environment-grounded three-arm benchmark (no coordination; reactive detection; proactive planning) that verifies integration off a real git merge, both in a deterministic simulation and with live agents. The planner lifts clean-integration from 1/9 to 9/9 scenarios and cuts merge conflicts from 13 to 0, with a gap that grows in the number of agents. On a live breaking contract change it rescues an outcome both baselines miss on every seed: the clean-integration rate rises from 0 (no coordination and reactive detection) to 1.0 on a frontier model and 0.6 on a small one, while agents respect assigned scopes (0/5 leakage). A cross-session memory drops the repeated-mistake rate from 1.00 to 0.00 on strong and weak models alike. We also report a negative result: routing facts to agents does not rescue long-context accuracy at window-fitting scales; its value is cost and capacity, not attention. Across two capability tiers and two vendors, the benefit did not shrink as models got stronger, because it comes from how work is allocated, not model reasoning.

Modern voice assistants may be shared by multiple users and should be able to answer questions about earlier conversations such as "When did I originally plan to leave?" or adapt their behavior to individual users based on past interactions. This requires more than retrieving a topically similar passage: the assistant must identify the current speaker, recover the relevant past state, and distinguish it from later revisions. We present PERSIST, a persistent memory system for multi-session, multi-speaker spoken dialogue that explicitly models Who, What, and When. PERSIST structures cross-session histories into readable event records and retrieves them with a 3W joint scoring mechanism that combines semantic content, acoustic speaker identity, and temporal state. For real-time full-duplex interaction, PERSIST further reuses intermediate representations from the dialogue backbone, avoiding query-audio re-encoding and reducing retrieval latency from 578.42 ms to 7.03 ms. We also introduce SpokenTrace, a diagnostic benchmark that factorizes evaluation along memory tasks and speaker-query types, exposing failures in recall, speaker attribution, and temporal-state tracking. On SpokenTrace, PERSIST achieves 85.08% end-to-end task accuracy and improves all-support EM@3 from 49.01% with BGE-large to 82.10%.

Speculative decoding accelerates autoregressive generation by using a lightweight draft model to propose multiple tokens that are verified by a target model in parallel. However, the standard acceptance rule focuses on exact distribution correction and rejects tokens that remain highly plausible under the target model when the draft model assigns excess probability. This conservative verification limits the number of draft tokens retained after each verification forward pass. We introduce Nucleus Speculative Decoding (NSD), a relaxed verification method that incorporates target-model plausibility into speculative decoding. NSD accepts a draft token if it satisfies the standard acceptance rule or belongs to the target model's nucleus. We theoretically characterize the distributional deviation introduced by our method and show that the single-step error is exactly determined by the draft model's excess probability within the target nucleus. We further derive sequence-level fidelity bounds that quantify how local deviations accumulate over autoregressive decoding. Experiments across multiple target models and proposal mechanisms demonstrate that NSD consistently improves speculative decoding efficiency while maintaining competitive task performance. Our method achieves throughput speedups of up to $5.16\times$ over autoregressive decoding and up to $3.15\times$ over standard speculative decoding. These improvements coincide with longer accepted lengths, allowing more output tokens to share the cost of each target verification pass. Analysis shows that plausibility-aware verification provides an effective approach for relaxed verification and speculative decoding efficiency. Our code is available at https://github.com/EIT-NLP/Nucleus-Speculative-Decoding.

Long-horizon LLM agents act on an append-only interaction history that is re-sent to the model at every step, so the context and its cost grow with steps until the sessions exceed the context window. Existing methods manage the context through context requirement prediction, relying on additional model calls, heuristic rules, or trained policies. However, these predictive approaches introduce runtime overhead, invalidate prefix caches, and permanently discard content with no guarantee of recovery. To overcome these limitations, we introduce ReFold: a training-free rendering layer that preserves the underlying interaction history while compressing only the model's rendered context. It removes two kinds of inter-turn redundancy without an auxiliary predictor: content an earlier turn already displayed, replaced by a stub, and turns the agent itself reports finished, folded into a one-line note. Both operators use chunked rendering, rewriting the cached prefix once every few steps rather than at every step. Every removal is strictly reversible, a wrong removal costs one restore from the history rather than permanent content loss. Because it operates at the rendering layer, ReFold is plug-and-play across standard ReAct-style harnesses. Evaluations across five long-horizon benchmarks and two frontier LLMs demonstrate that ReFold reduces token consumption by up to 2.5x and halves the KV-cache memory per session without degrading task success rates. Under capped context budgets, it avoids up to 92% of forced compactions. Under concurrent serving workloads, it reduces request queuing delays by up to 100%, accelerating inference by up to 1.7x, while cutting inference costs by up to 3.4x.

Ambient documentation systems are rapidly gaining adoption, yet their impact on clinical note quality remains poorly characterized. We introduce MedConv, a multilingual dataset of 300 clinical encounters in English, Danish, and German, and use it alongside the Ambient Clinical Intelligence benchmark (ACI-BENCH) to compare Corti, a clinical AI platform, with two leading, accessible ambient scribe software applications built on general-purpose AI. We present a controlled clinical evaluation framework that combines entailment metrics with LLM-judged pairwise comparisons across eight dimensions adopted from PDSQI-9. Results show that Corti's API-based text-generation infrastructure is on par with or outperforms leading commercial scribes. We further show that Corti's configurable API provides the flexibility necessary to fine-tune quality dimensions for specific documentation use cases. We present the evaluation methodology and release a dataset to support future reproducible comparison of ambient documentation systems.

Large language model agents are accelerating scientific automation, yet verified executions rarely become persistent program-level improvements, and existing evaluations do not examine this process across sequential tasks in both the natural and social sciences. We formalize ScienceClaw as fixed-parameter program self-evolution that unifies task solving, scientific verification, and program updates. ScienceClaw-Eval spans 23 disciplines and measures scientific correctness, evolutionary gain, retention, cross-dataset transfer, and evolution cost through sequential streams and independent reset evaluation. Our framework repairs executable workflows through multi-turn interaction, converts re-execution-verified failure--success trajectories into linked Skill and Operator candidates, and retains an update only when source-task replay reproduces the repair and independent scientific tasks improve. Code is available at https://github.com/beita6969/ScienceClaw.

A common approach to measuring bias in Large Language Models is to compare the log-likelihoods of two contrastive stereotype sentences. We argue that such single-pair comparisons are often unreliable: simply rewriting the same stereotype with an alternative attribute can yield logically inconsistent preferences. To address this, we propose a dual minimal pair setup that introduces two axes of comparison for robust stereotype evaluation. First, we present a data-augmentation framework that fills critical gaps in existing stereotype datasets by generating paraphrases and alternate attributes. We apply our framework on a set of English, Russian, Spanish and Chinese stereotypes. Second, we introduce two evaluation metrics tailored to the dual minimal pair setup. One of these metrics provides a new perspective on bias by modeling the mutual information (MI) between social groups and stereotyped attributes. This MI-based metric is better suited for aggregation and enables more robust comparisons of stereotype strength across different languages and models. Our code is available at https://github.com/stepanat/missing-minimal-pair/.

Injecting skills into a frozen language model currently costs a million parameters and a reinforcement-learning pipeline. We introduce \method{}, a System-1 decision operator trained by behavior cloning that lowers this cost by roughly two orders of magnitude. The default operator uses 330K parameters to match a 1.33M-parameter operator trained with reinforcement learning, exceeds or achieve comparable performance, while collapsing 3,685-token deliberation into a 6-token decision with no loss in accuracy. A rank-4 variant with 23K parameters, 1/58 of the strongest published skill operator, suffices for SearchQA and near-suffices for LiveMath, where higher rank still helps; the same recipe transfers across five tasks and three backbones, with out-of-distribution gains persisting on LiveMath problems released months after training. The gap to prior work is trainability, and it is set jointly by initialization and architecture: the initialization of prior operators zeroes the gradient of both large factor matrices at the first optimization step, whereas our zero-initialized output projection inside a shared low-rank backbone receives a gradient immediately, which a gradient-flow probe confirms directly. The gain isn't chain-of-thought compression: 23 of 57 LiveMath points beat the base model's best-of-8 sampling, and a logit-lens probe shows the operator amplifies the answer along the model's existing late-layer pathway, not writing it earlier. Gains track the base model's headroom across 13 base--task pairs, and skills compose as approximately linear operators that can be added, interpolated, and hot-swapped at inference time. Code on https://github.com/rlisml/decisionsteer.

Large-scale pretrained transformer models have achieved state-of-the-art performance across diverse machine translation tasks, including multilingual settings. Knowledge distillation has emerged as a sustainable approach for model compression, transferring knowledge from large teacher models to smaller, more efficient student models. Similarly, quantization, which reduces the numerical precision of model weights and activations (e.g., from 32-bit to 8-bit representations) is widely used to accelerate inference, enabling models to run several times faster during deployment. However, both techniques face limitations when applied to specialized domain data, particularly under low-resource conditions. In knowledge distillation, the effectiveness of transfer is often constrained by the scarcity of domain-specific parallel data, while quantization can lead to performance degradation as bit precision decreases. In this work, we investigate the combined application of knowledge distillation and quantization for French-to-English biomedical translation, a domain characterized by specialized terminology and limited parallel resources. We develop and compare multiple fine-tuning strategies to adapt compressed student models to this challenging setting. Our experiments demonstrate that a collaboratively distilled and quantized student model achieves a 69% reduction in size, a 98.21% increase in inference speed, and a 98.46% reduction in CO2 emissions compared to the original baseline all without sacrificing translation quality. These results indicate that jointly optimized compression techniques can yield efficient, high-performance models suitable for translation service providers operating under resource constraints.

LLM agents interact with external systems by generating structured tool calls. Given a user request, conversational context, and a catalog of tool schemas, a tool-calling model must select tools and generate their arguments, potentially producing multiple calls in a single response. Standard autoregressive decoding generates these calls token by token, incurring substantial latency for requests involving multiple calls or many argument fields. The explicit argument structure offers opportunities for parallel generation, but later argument values may depend on preceding fields and calls, so independently generated values can differ from the target model's output. We present SchemaFill, a framework for efficient LLM tool calling through slot-parallel speculative decoding. SchemaFill generates future slot values concurrently as candidates, without requiring advance knowledge of the actual call sequence or argument values. Candidates spanning multiple fields and calls are concatenated for verification by the target model under the actual output prefix. Only verified tokens are committed, and the target supplies corrections when candidates disagree. This applies target verification while exploiting parallelism across slots and calls. On Glaive and BFCL, SchemaFill achieves up to a 4.05$\times$ improvement in end-to-end throughput over autoregressive decoding. Code is available at https://github.com/Czzzk/SchemaFill.

Diffusion large language models dLLMs) have emerged as a promising alternative to autoregressive language models through bidirectional diffusion-based token generation. However, their growing model sizes and high inference costs make efficient deployment challenging: full-sequence denoising repeatedly invokes compute-intensive forward passes, while block-diffusion models additionally introduce a memory-intensive KV-cache. Low-bit weight-activation quantization is therefore attractive, yet existing dLLM post-training quantization methods rely on calibration data despite activation distributions shifting across masking states and denoising steps. We present SoloQ, a calibration-free quantization framework that maps weights and activations into a normalized rotated basis with a predictable marginal distribution, enabling data-independent quantization. SoloQ combines a structured K-RPBH rotation with a lightweight rescaling correction for calibration-free quantization. Its predictable post-rotation distribution supports both distribution-matched codebooks and hardware-native NVFP4. For block-diffusion models, SoloQ further applies commit-time KV-cache quantization to compress persistent states without perturbing the actively denoised block. Across full-sequence dLLMs (LLaDA and Dream) and block-diffusion dLLMs(Fast-dLLM v2 and Nemotron-Labs-Diffusion), SoloQ retains accuracy under 4-bit quantization and outperforms calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, SoloQ reduces peak memory by up to 2.61X and accelerates end-to-end inference by up to 2.24X.

Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are insufficiently trained for the assigned inputs. We propose Distributionally Robust MoE Training (DRMoET), a drop-in objective that treats layer-wise experts as endogenous robustness groups and optimizes high-loss routing outcomes rather than merely equalizing traffic. DRMoET updates a per-layer expert distribution by an entropy-regularized softmax rule on EMA-smoothed, activation-weighted expert losses, strengthening plausible non-top routing paths while preserving standard MoE computation. Under the FLAME-MoE recipe at 746M-total and 10.3B-total scales, DRMoET improves downstream averages over both standard FLAME-MoE and auxiliary-loss-free balancing. At 10.3B total parameters and 67B training tokens, DRMoET improves the seven-task average from 0.6625 to 0.6767, while the auxiliary-loss-free baseline achieves 0.6431. Mechanistic analyses show lower expert-loss variance with nearly unchanged mean loss, 4.3% lower excess loss under forced mid-$k$ misrouting, and improved domain-expert specialization. These results position routing robustness-not only utilization balance-as a practical objective for reliable sparse MoE scaling. Project page and code are available at: https://drmoet.github.io/.

Deterministic benchmark scores show that an agent received credit, but not whether that credit was earned, reported honestly, or would hold on a second run. We introduce a Trust Layer for Agent Evaluation, an additive post-hoc framework that reports, beside each recorded score, whether it should be believed. It verifies four properties: whether the result is supported by the benchmark's own grading logic, whether a passing answer was earned through traceable computation, whether the agent's completion claim matches what occurred, and whether the result is stable under repeated execution. The first three use only saved artifacts; the fourth re-runs the agent. Model judgments only label evidence under majority voting; all verdicts follow deterministic rules and never modify the recorded score. Applied to five agent configurations on 108 tasks from Agents' Last Exam, every model shows passing runs with no traceable computation (at rates varying tenfold), confirmed false completion claims, and unstable results: 18-46% of tasks do not stay in one score band over five runs. Only 22.6% of recorded passes clear all four checks (95% CI 15.0-32.6, n=84). Measuring what an agent can do and verifying that it did it are different problems, and current benchmarks address only the first.

We evaluate large language models (LLMs) in multi-turn interactions through their time-to-event: the number of interaction steps required to produce an event of interest, such as a successful jailbreak or agentic task completion. Under limited compute, interactions may be terminated before the event occurs, so that event times are only partially observed (censored). Existing allocation methods for calibrating time-to-event bounds satisfy the budget only in expectation and can exceed the available budget on a particular evaluation run. Enforcing a hard constraint is particularly challenging as the cost of a trajectory is initially unknown. We introduce Hard-budget Allocation with Reflow for Predictive calibration (HARP), a budget allocation that satisfies hard resource constraints and adaptively reallocates unused budget. We show how to use HARP to construct lower predictive bounds (LPBs) on the time-to-event and to estimate evaluation metrics such as the jailbreak rate on a fixed benchmark. Although HARP induces dependence in acquisition decisions across different trajectories, we prove that HARP never exceeds the target budget, that its LPBs have finite-sample coverage guarantees, and that its metric estimates are unbiased. Experiments on agentic task success, LLM jailbreaks, toxic content generation, and RAG hallucinations show that HARP achieves coverage close to the nominal level with low variance, while never exceeding the given budget.

As LLMs increasingly assist scientific writing and peer review, detecting who wrote the text is no longer sufficient: we need to determine who contributed the underlying insight. We introduce Insight Provenance, the task of identifying whether a review insight originates from a human, an LLM, or their hybrid contribution. We construct InsightProv-v0 from 4,057 scientific papers and 12,660 human reviews, simulating different levels of LLM involvement with GPT-4o, Gemini, and DeepSeek and annotating provenance at the sentence level. We show that strong performance on raw data can be misleading, as models exploit linguistic and textual-authorship shortcuts that degrade substantially under progressively debiased evaluation. We therefore propose a two-stage adversarial framework that suppresses shortcut signals while preserving provenance-relevant information. Beyond detection, extensive analyses reveal what makes intellectual authorship identifiable: paper grounding and neighboring review context provide complementary provenance signals, while human, hybrid, and AI insights systematically differ in their information sources and failure modes. Most strikingly, AI insights predominantly remain close to generic or paper-provided information, whereas human insights more often introduce external knowledge and independent judgment. These findings suggest that while wording can be rewritten by an LLM, the provenance of an idea leaves a deeper and more persistent signal.

Long-term memory allows Large Language Models (LLMs) to maintain personalized context across interactions, but retrieved user history can induce memory-induced sycophancy, causing models to favor stored user beliefs over objective evidence. Existing defenses primarily operate on retrieved context and are rarely evaluated jointly with internal behavioral bias. We introduce a $2 \times 2$ defense-in-depth framework separating internal activation steering from external memory handling. We extract sycophancy steering directions from 100 paired prompts and evaluate four open-weight models across 10 steering coefficients and five memory-defense configurations on MemSyco-Bench (answers for all 1,550 items; defense conditions judged on a fixed 250-item subsample), with three LLM judges. Three of the five configurations are new (rewriting every memory, a Router Gate that keeps, rewrites, or drops each memory, and dropping all memory); the other two are MemSyco's baselines. Selective Router Gate filtering preserves substantially more of MemSyco's average accuracy than complete memory removal, and this separation persists when the models are steered toward sycophancy. On Llama 3.1 8B with Router Gate, mild inverse steering ($\alpha = -1.5$) lowers judge-averaged sycophancy from 35.80% to 31.32% while average accuracy moves from 43.99% to 43.31%; this reduction has the same direction under all three judges but is not statistically significant (paired $p = 0.08$ to $0.63$ on 149 items). External memory filtering is the part of the design that holds up; our data do not show that inverse steering adds to it.

Fine-grained mixed-precision quantization promises efficient large language model inference, but local precision choices can conflict with regular GPU storage and computation units. This precision-boundary mismatch limits the translation of compression into practical acceleration. We introduce AlignQuant, a post-training quantization method that uses GPU-compatible two-dimensional weight tiles as the common unit of precision allocation, compact storage, and execution. This shared partition lets precision follow sensitivity within output channels. Joint prefill/decode calibration scores precision reductions using projection-output perturbations weighted by language-model loss gradients under quantized activations. Phase-normalized scores prioritize higher precision for tiles important to either phase under a model-wide weight-storage budget. Each tile stores one selected representation, while phase-specialized kernels reuse the packed model and expand lower-bit weights for INT8 computation with 8-bit activations. Across four LLMs spanning 3B to 14B parameters, AlignQuant achieves up to $2.50\times$ generation speedup over BF16 while preserving model quality. Evaluations further cover three GPUs and contexts up to 64K tokens. These results show that local precision flexibility and regular GPU execution can coexist through a shared tile unit. The implementation is available at https://github.com/HanzhiZhang-Ulrica/AlignQuant.

Safety auditing of post-trained large language models typically relies on model behavior, requiring model execution and depending on the coverage of available evaluations. This work asks a different question: Do the target behaviors optimized during supervised fine-tuning (SFT) leave readable evidence directly in checkpoint updates? We find that harmful-compliance SFT induces a continuous, objective-dependent ordering in checkpoint-update space. Using a reference geometry defined by pure harmful-compliance, safety-targeted, and benign-utility SFT, we find that a checkpoint-level coordinate s_H tracks controlled harmful-objective composition with Spearman correlations of 0.986-0.992 across four 7-8B backbones, with the same ordering persisting at larger model scales. Matched compliance-versus-refusal controls show that this checkpoint trace reflects the SFT objective rather than harmful-input exposure, while additional controls rule out simple explanations based on harmful-example count or generic training intensity. Building on this structure, we introduce TRACE, a weights-only auditing method that localizes an unknown checkpoint update relative to frozen harmful and non-harmful reference prototypes and converts this geometry into a continuous harmful-objective score. TRACE requires neither model queries nor access to the unknown SFT data, and can be evaluated directly from checkpoint updates. Across distribution shifts, unseen data, different SFT configurations, partial checkpoint access, and LoRA/full-parameter fine-tuning, the trace remains stable and is positively associated with independently measured attack success rates. TRACE remains informative even at low harmful-objective proportions, providing a complementary auditing signal when behavioral evaluation is unavailable or incomplete. Code is available at https://anonymous.4open.science/r/Code4TRACE-54D3.

Modern recommendation is shifting from platform-centric personalization toward user-governed personalization, where a personal LLM agent can act on the user's behalf across services. We formalize this emerging paradigm as Personal-Agent Mediated Recommendation: a platform recommender ranks a candidate set using platform-local information, and a personal agent uses user-authorized cross-platform history to mediate the resulting ranking and produce the final top-K slate. Such mediation is nontrivial: the platform ranking can encode strong population evidence that the personal agent cannot observe, so effective mediation must therefore balance beneficial rescues against harmful overrides. To study this trade-off, we introduce MediateRec, a benchmark that includes scalable proxy cross-platform environments and a real cross-platform test under a controlled platform-agent information boundary. To train the agent to use cross-platform history effectively, we further propose Personal Attribution Mediation Optimization (PAMO), which counterfactually masks that history to estimate personal mediation support and reallocates rank-aware advantage mass under a platform-relative value floor. We theoretically prove that PAMO preserves cutoff-level advantage mass and is locally optimal among first-order reallocations that preserve this mass without lowering average platform-relative value. Experiments on MediateRec show that personal-agent mediation enables meaningful platform corrections, yet even strong proprietary LLMs introduce non-negligible harmful overrides. PAMO consistently improves over matched outcome-only RL across seen and unseen target platforms and on the real cross-platform test, while achieving a better rescue-harm balance.

Speculative decoding accelerates autoregressive generation in large language models. In each drafting stage, a lightweight draft model proposes tokens that the target model subsequently verifies. With increasingly capable draft models, we find that the target model frequently accepts all tokens produced in a drafting stage. A verification nevertheless follows each drafting stage, resulting in unnecessary target-model forward passes even when drafting could have continued. Adaptive draft length methods decide during decoding how many draft tokens precede a verification, but they raise the speedup only for autoregressive draft models. For a parallel draft model, drafting further requires target-model hidden states for draft tokens that have not been verified. We propose DLoop, a looped form of speculative decoding that adaptively performs multiple drafting stages before verification. DLoop continues drafting while the draft model remains confident and verifies all accumulated draft tokens together. Loop-aware training keeps the draft model reliable in the additional drafting stages by exposing it to its own hidden states for unverified draft tokens. By spending additional draft-model forward passes, DLoop reduces the number of target-model forward passes required for verification. Across diverse speculative decoding methods including EAGLE-3, DFlash, Domino, DSpark, and multi-token prediction modules, DLoop improves the wall-clock speedup by 5 to 41 percent while preserving lossless decoding. Code will be available at https://github.com/naver-ai/DLoop.

Speculative decoding reduces large language model inference latency by drafting multiple tokens before target-model verification, but its effectiveness depends on both the proposal mechanism and draft depth. Fixed configurations cannot respond to changes in predictability, repetition, and acceptance during generation, so deeper drafting can increase wasted computation without proportional speedup. We introduce APEX, a learned controller that balances decoding speed and draft-token waste through request-level expert selection and block-level depth adaptation. APEX-Router selects among EAGLE-3, n-gram, and draft-model speculation for each request, while APEX-Depth adjusts draft length at each verification block using causal decoding signals and recent verifier feedback. APEX models accepted draft length as censored survival feedback, learning position-wise rejection hazards, block execution costs, and an action utility that balances throughput, accepted progress, and wasted tokens. This allows the controller to adapt speculation while retaining the target model's verification procedure. We integrate APEX into vLLM and evaluate it with Qwen3-8B across six workloads, achieving up to 5.24X speedup over autoregressive decoding. Across the aggregate evaluation, APEX-S achieves 4.27X speedup, while APEX-B achieves 3.27X speedup with a 41.0% relative reduction in wasted-token percentage compared with fixed n-gram speculation at k=16, providing distinct operating points for balancing acceleration and draft-token utilization.

Post-training quantization reduces the cost of deploying language-model agents, but its effect on recovery from temporary tool failures can depend on how recovery is evaluated. We compare 8-bit and 4-bit variants of Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct on twenty deterministic tool-use tasks and five prompts. The 8-bit-4-bit recovery comparison changes direction across prompts and evaluation targets. On tasks that both variants complete without faults under the same prompt, the difference ranges from 0 to +20.2 percentage points for Llama and from -50.0 to +35.0 points for Qwen. Full-pipeline point estimates favor 8-bit Llama under all five prompts, whereas the Qwen comparison changes direction across prompts. The evaluation target can also reverse the result. For Llama under one prompt, scoring each variant only on its own clean-passing tasks favors 4-bit by 17.5 points; scoring the same tasks for both variants gives no difference, while scoring the full pipeline favors 8-bit by 28.3 points. Executor leniency is a third such choice. Rescoring the same logs with strict output parsing, which 8-bit Llama violates far more often than 4-bit Llama under that prompt, turns that +28.3 into -15.0 while leaving Qwen essentially unchanged. These findings show that one prompt, one screened task set, and one scoring policy do not establish a stable conclusion about quantized-agent robustness. Evaluations should compare variants on matched tasks, report full-pipeline success for deployment decisions, state the scoring policy, and quantify uncertainty across tasks rather than injected fault sites.

Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key-value (KV) cache by T. The larger cache limits how many sequences a GPU can decode at once and slows each decoding step, which reads the whole cache. We propose Hybrid Latent Attention (HLA), which keeps exact keys and values within a sliding window of W recent tokens and stores each older token as a compact latent that the query of each loop reads directly, without reconstructing keys and values. We uptrain HLA on Ouro looped models (T=4) with 1.4B and 2.6B parameters, keeping the pretrained weights frozen and training only the added parameters to reproduce the original attention. The cache shrinks by 10.7x per token, fitting 4.0-8.8x as many concurrent sequences per GPU, and decoding throughput improves by 2.5x at 1K-token contexts and by up to 7.4x at 16K. HLA retains over 97% of the original accuracy on math, knowledge and reasoning benchmarks, and 96-100% on long-context retrieval up to 16K tokens. After supervised fine-tuning, it performs on par with the fine-tuned original model on competition-level math.

AI agents are increasingly being programmed to automate knowledge work over large collections of unstructured data. Such automation requires repeatability: when the underlying evidence is unchanged, the agent's categories, priorities, and counts should not shift materially between runs, even if each individual answer appears plausible. We introduce a repeat-run evaluation framework that aligns semantically equivalent categories and focuses on two operating metrics: theme churn, the normalized change in the returned category set, and volume disagreement, the change in counts for categories that persist. We evaluate three recurring customer-feedback tasks across eight frontier models, corpus sizes from 100 to 5,000 records, multiple prompts, and three execution designs: raw generation, taxonomy-free hierarchical decomposition, and a taxonomy-grounded agent (TGA) using persistent themes, subthemes, and record-level predictions. With Claude Opus 4.8 and the 1,000-record corpus fixed, TGA reduces theme churn by 86--88% relative to both raw generation and hierarchical decomposition, while matched-theme volumes have zero disagreement. The taxonomy-grounded agent is more stable than every raw model in the screen, remains more stable at each corpus size, and keeps this advantage when theme matching is made stricter or looser. Although evaluated on customer feedback, the framework targets repeated synthesis of unstructured corpora more broadly, including financial reports, legal documents, incident records, and scientific literature. Overall, these results show that taxonomy grounding produces more consistent and repeatable outputs for recurring knowledge work.

Multimodal large language models (MLLMs) advance rapidly, while conventional benchmark development lags behind, delaying investigation of newly observed capability gaps. Such investigation requires an expressive task format and an on-demand construction process: information-rich charts make chart question answering (Chart QA) suitable for probing coupled perception and reasoning. Automated Chart QA construction is intended to shorten the benchmark-development cycle by turning identified gaps into targeted samples on demand. Current methods, however, commonly separate target guidance from scratch generation: target-guided systems often require prepared data, charts, or templates, while scratch-generation systems primarily ensure artifact validity, without explicitly controlling whether newly synthesized requirements and content remain aligned with an externally specified diagnostic target. We introduce ChartBmkAgent, which turns an identified capability gap into targeted diagnostic evidence by constructing complete Chart QA samples from sparse error-taxonomy specifications. Throughout construction, a central harness governs specialized agents, requires stage-specific evidence of alignment with the original error category, and records the basis for each acceptance decision. On 300 taxonomy-wide samples, MLLM accuracies ranged from 32.7% to 84.3% with distinct category profiles, showing that generated samples reveal capability differences. Across three source-model comparisons, targeted follow-ups scored 50.0% versus 82.2% on matched controls ($p=8.96\times10^{-6}$); all six cross-model comparisons had the same direction, demonstrating targeted validation and diagnostic-data generation. Multiple evaluator models assessed whether each sample tested its specified error category; 86.4% met this criterion, providing empirical evidence of target preservation.

Multiple-choice question answering (MCQA) is commonly used to evaluate large language models under the assumption that one of the provided options is correct, typically using answer-selection accuracy. However, in real deployments, users or retrieval systems may provide invalid option sets in which none of the listed choices is correct, and selecting one of them may incur downstream cost. We study this setting as penalty-framed no-valid-option MCQA. Using the mathematics subset of MMLU-Pro, we remove the labeled correct option, allow models to either choose a remaining option or output ABSTAIN, and penalize invalid forced-choice responses. We further introduce correct-conditioned analysis, evaluating abstention only on instances that the model originally answered correctly. Experiments show that high MCQA accuracy does not fully guarantee abstention reliability: even under explicit no-valid-option-aware instructions and penalty-based scoring, models still produce invalid forced-choice responses for a subset of originally correct instances. These results show that penalty-framed no-valid-option MCQA reveals an aspect of model reliability not captured by standard answer-selection accuracy.

Large Language Models (LLMs) require substantial computational resources, limiting their deployment on resource-constrained hardware. Ternary LLMs mitigate these demands through weight quantization via ternary values, achieving significant compression often with 50-90% sparsity. However, existing approaches have limitations: methods optimized for ternary weights, such as BitNet, redundant segment reduction (RSR), and its improved version RSR++, do not exploit sparsity structures, while conventional sparse formats neglect ternary characteristics, foregoing dual optimization opportunities. In this paper, we introduce Sparse Segment Reduction (SSR), a ternary matrix multiplication method designed to accelerate the inference of ternary LLMs and general Ternary Weight Networks (TWNs). SSR has a dedicated optimized ternary data format and an algorithm that systematically exploits sparsity patterns through computation trees that scale with the sparsity. SSR provides theoretical gains with asymptotically faster inference than RSR++ for sparsity above 50%, while practical evaluations reveal performance improvements across all sparsity levels. Evaluation results show that SSR achieves 2.1-11.3x speedup over RSR++ on ternary GEMM with 45-95% sparsity. Furthermore, SSR achieves 3.5-6.3x end-to-end speedup and 4.9% of memory saving over RSR++ on the Llama-3 1B model inference.

Language models increasingly act as agents. An agent that says an action is wrong and then takes it anyway is a different failure from one that does not know better, and evaluations of stated values cannot see it. We build a pre-registered panel of 248 scenarios across five kinds of pressure. Each scenario is posed twice to the same model, once as the agent choosing what to do and once in the third person asking which option is right, so the model's own judgment is the reference. Every scenario has a twin with the pressure removed, and every model gets a positive control in which its operator orders the violating action, so that a missing gap can be told apart from a blind instrument. On OLMo-3-7B-Instruct, the model takes the action it judged wrong on about one in five pressuring scenarios, more often than on the same scenarios with the pressure removed. Across four instruct models the gap depends on the post-training recipe: OLMo-3 and Meta's Llama-3.1-8B-Instruct carry it; Tulu 3 shows none on the whole panel (above about 0.01 in probability) or on its own most-pressuring scenarios; Qwen2.5-7B-Instruct shows none on the whole panel (above about 0.02) and is unresolved on its own (0.083, -0.028 to 0.195). Meta's recipe and Ai2's Tulu 3 start from the same Llama-3.1 weights, and only Meta's carries the gap. Reading a chat model outside its chat template reverses the sign of its gap with nothing at stake (-0.038 against +0.055 under the template on OLMo-3), a distortion present on two of three recipes. On both models that carry it, reasoning about the stakes before acting moves the choice back toward the model's own judgment, against a same-length non-moral task, with or without the pressure; on OLMo-3, naming the norm at stake does about a third of that. The gap is a measurable target for post-training recipes, not a fixed property of pretrained weights.

Diffusion Language Models (DLMs) hold the promise of order-agnostic, parallel text generation. Recently, continuous diffusion and flow matching models have seen substantial gains, driven by carefully crafted token representations and diffusion/flow spaces. In this work, we introduce Hierarchical Continuous Diffusion Language Models (H-CDLMs), a simple framework that further improves continuous DLMs with minimal compute and parameter overhead. Drawing on the discrete DLM and continuous image diffusion literature on joint diffusion, we diffuse multiple modalities in parallel. These modalities represent tokens at different semantic granularities: in our instantiation, the tokens themselves and coarser clusters obtained by clustering pretrained token embeddings. We propose a general setup that allows per-modality samplers and schedules to enhance the interplay between modalities. Applied to CoBit, this yields H-CoBit, which delivers large empirical gains across benchmarks. At dataset entropy, H-CoBit improves MAUVE and reaches a generative perplexity (GenPPL) of 49.4 on LM1B and 50.4 on OWT, improving on the baseline by 24.2 and 20.7 points and surpassing even discrete DLMs of comparable size. On GSM8K, it reaches 27.4% accuracy, outperforming prior continuous diffusion and flow-based models. We further apply H-CDLM to the flow matching model FLM, obtaining consistent gains with H-FLM and demonstrating that the framework generalizes across continuous generative paradigms. Our code will be made publicly available at https://github.com/matol-16/HCDLM.git .

Large language models (LLMs) have become increasingly capable problem solvers, but being able to solve a problem is not the same as being able to teach it. Existing approaches to training LLMs as teachers rely on demonstrations, preference data, or predefined pedagogical criteria that specify what good teaching looks like. However, these signals are often not grounded in individual student learning outcomes, where effective teaching strategies can vary substantially across learners. To address this, we introduce Sherpa, a multi-turn reinforcement learning framework that instantiates multiple student archetypes with LLMs conditioned on distinct learning preferences and trains a teacher model to adapt its instruction by directly maximizing their learning outcomes. Teacher LLMs trained with Sherpa improve instructed students' performance across all archetypes by an average of 20.5 percentage points. Under MathTutorBench's evaluation, Sherpa raises the overall pedagogy score from 52.5% to 79.2%, indicating better teaching responses. Our human studies show that the trained teacher is preferred over the base model in 79.6% of pairwise comparisons. Together, Sherpa trains LLM teachers to adapt to diverse simulated students and become better aligned with human teachers, paving the road towards AI tutors teaching real students.

Large speech language models have demonstrated strong capabilities in unified cross-modal understanding and generation, yet paralinguistic cues, especially emotion, remain difficult to preserve. Existing systems typically rely on entangled acoustic representations, which allow the underlying language model to depend excessively on recovered lexical content instead of grounding its behavior in acoustic-prosodic evidence. We address this limitation with EMODE, an emotion-aware speech language model built around \textbf{Dynamic Para-Semantic Experts (DPSE)}. DPSE decomposes continuous speech features into semantic and paralinguistic pathways, routes them dynamically, and fuses them before integration into the language model. To turn this structural decomposition into functional specialization, EMODE is trained with a three-stage curriculum consisting of semantic warm-up, paralinguistic activation, and joint refinement, guided by Orthogonal Expert Guidance (OEG), Semantic-to-Acoustic Alignment (SAA), and Gating Diversity Regularization (GDR). Experiments on SER test, empathetic response evaluation, and the newly constructed bilingual MEPA benchmark show that EMODE improves the balance between lexical fidelity and emotional sensitivity, strengthens affect-grounded response generation, and exposes the value of explicit para-semantic factorization for robust cross-corpus emotion understanding.

Evolution Strategies (ES) enable memory efficient full parameter fine-tuning of large language models (LLMs) using only forward computation. However, standard ES uniformly averages rewards across problems and compresses problem level population feedback into a single scalar, making it difficult to capture how the learning value of each problem changes with model capability. To address this limitation, we propose Difficulty-Aware Reweighting and Targeted Replay for Evolution Strategies (DART-ES). DART-ES estimates the local solvability of each problem from its pass rate across the perturbation population and aggregates historical observations to construct a dynamic difficulty state. This shared state jointly guides continuous difficulty reweighting and rare solvable sample replay, thereby improving perturbation direction evaluation and training data allocation without introducing an additional difficulty model or backpropagation. Extensive experiments show that DART-ES achieves good fine-tuning performance. DART-ES outperforms ES on all five base models and improves the average accuracy from 72.07\% to 73.53\%, exceeding the 73.26\% achieved by GRPO on GSM8K. Across five challenging mathematical reasoning benchmarks, DART-ES achieves an average accuracy of 49.20\%, compared with 48.34\% for ES and remains competitive with strong 7B models trained with RL. Further experiments show consistent gains in instruction tuning, code generation and the Countdown task with a 14B model, demonstrating strong generalization across tasks and scalability to larger models. Beyond performance gains, DART-ES also shows clear advantages in system efficiency. It reduces runtime per step by 15.2\%--50.2\% and peak memory usage per GPU by 21.1\%--51.1\% compared with GRPO. Despite performing full parameter fine-tuning, DART-ES also requires less runtime and GPU memory than GRPO+LoRA.

Large Language Models (LLMs) have achieved remarkable capabilities but remain vulnerable to jailbreak attacks that elicit harmful or unsafe outputs. Existing safety alignment approaches, including Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), often require substantial attack-specific supervision and computational resources, while remaining susceptible to shallow safety alignment and over-refusal. To address these challenges, we introduce SSRFT(Supervised Safe-Role Fine-Tuning), the first framework that reformulates safety alignment as the internalization of a predefined safe role. SSRFT constructs a Safe-Role Question-Answer (SRQA) dataset from psychometric questions, limited jailbreak prompts, and a safe-role description. Role-consistent responses are synthesized, validated, and expanded into diverse scenarios, enabling models to internalize safety-oriented values and principles rather than explicit refusal patterns. Experiments across multiple Base and Instruct models show that SSRFT achieves more robust and generalizable safety alignment than standard SFT. SSRFT shows substantially greater robustness to prefilling attacks and better generalization to unseen jailbreak domains, while reducing over-refusal on benign queries and preserving the model's general capabilities. These results establish safe-role internalization as an effective alternative to refusal-centric safety alignment. Warning: This paper contains examples of harmful and toxic language.

Large language models are increasingly used to simulate how individuals respond to new situations, yet the behavioral reasoning behind these responses is either inherited from pretraining or learned from individual-level annotations, which offer limited behavioral diversity and little supervision of the reasoning itself. We propose to learn behavioral reasoning from prediction markets, whose price trajectories record how populations respond to real-world events at scale. We introduce macro2mind, which trains a language model with GRPO using market signals. A social behavioral decomposition makes behavioral reasoning an explicit step of forecasting: the model infers representative groups of market participants, predicts how each interprets the news and updates its beliefs, reasons about their interactions, and aggregates these responses into a price. A hindsight-regret curriculum with difficulty-aware sampling focuses training on transitions where hindsight-identified groups substantially improve the forecast while prioritizing examples that remain learnable for the current policy. The learned reasoning applies to user simulation without further training. On SWM-Bench, macro2mind achieves state-of-the-art directional accuracy and correlation on Polymarket. Trained on market data, it transfers zero-shot to four user-simulation benchmarks (Humanual, OvertonBench, PRISM, and CAD) and has competitive performance among zero-shot methods. Used as a data generator, macro2mind also raises a downstream simulator's accuracy on unseen users by 15.5 points, outperforming data generated by its backbone by 13.2 points.

Large language models have shown strong reasoning capabilities, but their high inference costs make knowledge distillation an important approach for transferring such capabilities to compact models in resource-constrained scenarios. On-policy self-distillation further reduces the reliance on external large teacher models while improving the reasoning ability of compact language models. However, existing methods typically either distill all token positions uniformly or select tokens using fixed heuristic criteria, assigning the same distillation strength to the selected positions rather than adaptively learning which tokens are most beneficial for distillation. To address these limitations, we propose BiToK-SD (Bilevel Top-K Token Selection for Self-Distillation), a bilevel-optimization-based token selection method that learns where distillation should be applied during on-policy self-distillation. Specifically, BiToK-SD is formulated as a bilevel optimization problem, where the lower-level problem models Top-K token selection as a differentiable threshold-based relaxation, allowing the selected positions to adapt as the student policy evolves, while the upper-level problem performs knowledge distillation on the selected positions. Experiments on mathematical reasoning benchmarks show that BiToK-SD achieves the best average performance among all compared methods while requiring only lightweight additional computation.

Direct Preference Optimization (DPO) has become a standard reward-model-free approach for aligning language models with preference data. However, as the scaled preference margin grows during training, the logistic DPO loss becomes progressively less sensitive to further changes. We study DPO from a loss-level geometric perspective and identify the sigmoid factor as a learning signal that characterizes the local sensitivity of the objective. Based on this view, we propose Learning-Signal-Controlled Direct Preference Optimization (LSC-DPO), which dynamically regulates the learning signal near a target regime. A log-space analysis establishes conditions for stable tracking of the target learning-signal regime. Experiments on AlpacaEval 2, MT-Bench, and Anthropic-HH show that LSC-DPO consistently improves over DPO and strong preference-optimization baselines. We further find that different coefficient initializations induce distinct transient learning-signal trajectories even when their later signal levels become similar. Based on this observation, we derive a signal-budget compensation rule that adjusts the target learning signal to compensate for these transient differences. The resulting compensation substantially reduces performance variation across coefficient initializations.

Repository-level software engineering (SWE) is a challenging long-horizon setting: agents must reason over extended interactions, use tools, and adapt to stateful environments. Recent work trains SWE agents with reinforcement learning methods such as Group Relative Policy Optimization (GRPO), which independently sample multiple trajectories per issue, test the resulting patches, and compare terminal rewards within a fixed group. However, this training setup does not reuse verifier feedback from failed patches as context for subsequent attempts, even though this feedback contains valuable diagnostic information about what went wrong. Training on recovery trajectories is challenging because the preceding outcome determines whether the next trajectory is generated, while the failed execution determines its conditioning context. We introduce FC-SWE, a failure-conditioned RL framework that incorporates recovery attempts into policy training. After a patch fails verification, FC-SWE restores the repository to its original task state and uses the failed patch and verifier feedback as context for a recovery trajectory. FC-SWE adapts GRPO to these chains of complete, multi-turn tool-use trajectories through two mechanisms. Trajectory-local rewards preserve each attempt's verifier outcome, preventing recovery success from rewarding an earlier failed patch. Active-set advantage estimation forms a comparison group from all initial and recovery trajectories actually executed for the same issue, so failed attempts remain in the group while unexecuted attempts are excluded. On all 500 SWE-bench Verified tasks under a verifier-assisted protocol, FC-SWE with Qwen3.5-4B and SWE-agent achieves 41.7% Resolved@1 and 52.8% Resolved@2, compared with 38.9% and 48.5% for GRPO. Although trained with at most two attempts per chain, FC-SWE reaches 70.7% Resolved@11 under an eleven-attempt test-time budget.