ai development teams

Abstract: Can AI agents organize scientific investigations that lead to well-supported conclusions? To evaluate this capability, we introduce PhysClosureBench, a benchmark of 50 expert-verified, goal-specified research tasks spanning strong-field, attosecond, and neighboring physics areas. Each task specifies the research objective and essential scientific constraints, while leaving the agent to organize the investigation and determine how to generate and validate the evidence needed to answer the question. A 20-criterion Atomic Score measures local scientific accomplishment, while Research Closure (RC0--RC4) assesses how fully the investigation resolves the specified research question based on whether its indispensable evidence requirements and applicable scope conditions are satisfied. We evaluate seven frontier LLMs, where Claude-Opus-5 achieves the highest mean Atomic Score (6.29/10) but reaches RC4 in only 32.7% of attempts. Across the benchmark, 36.2% of attempts scoring at least 8/10 still fall short of RC4, while all seven models reach RC4 on substantially fewer tasks in all three runs than in at least one. Case-level analysis further illustrates how misinterpretation of generated results and unresolved physical applicability limits can lead to incomplete closure. Together, these results reveal a persistent gap between AI agents' local scientific accomplishments and their ability to achieve complete research closure.

Abstract: Language-model agents can accumulate interaction feedback, but reliably turning that feedback into reusable experience remains challenging. Existing methods either optimize external experience scaffolds or train experience-related behavior on specific task distributions, leaving open whether experience learning itself can become a transferable model capability. Training such a capability calls for supervision beyond isolated tasks, which mainly support in-episode problem solving rather than revealing what knowledge remains useful across episodes. In this work, we introduce MetaExp, a scalable framework that trains this capability through verifiable task families with controlled cross-task structure. Each family preserves a reusable factor while varying task-specific realizations, enabling transfer without direct solution reuse. We propose two complementary designs: Shared-Tool families keep black-box tools fixed across tasks to elicit operational experience, while Shared-Workflow families preserve latent procedures across changing realizations to elicit procedural experience. We scale environments under both designs, collect and filter sequential interaction trajectories, and fine-tune Qwen3.5-122B-A10B on 4,956 trajectories from 1,000 task families. With parameters frozen during downstream evolution, MetaExp improves by +8.5 points on SpreadsheetBench, peaks at +2.3 on APEX-Agents, and improves by +14.8 on AutomationBench. The gains persist under domain-excluded training, supporting experience learning as a transferable capability for continued non-parametric improvement after deployment.

Abstract: Automated ideation systems are often evaluated on the novelty of the ideas they produce, and that judgment is increasingly delegated to large language models. Such judges are typically built ad hoc and validated, if at all, on human-authored papers rather than on the generated ideas they are meant to score. So, how do novelty judges perform? Not well. We present a systematic controlled study of novelty evaluation design choices. We first build an evaluation set automatically, mining OpenReview for passages where reviewers explicitly affirm or dispute a paper's originality and keeping only submissions with unanimous agreement at the extremes of their research area; we pair these with ideas from a vanilla LLM generator. Across six judges, we find that small prompt design choices have large consequences; e.g., simply telling the judge that reviewers found one idea novel and the other not can change its verdict on more than half of the identical idea pairs it is shown, shifting pairwise accuracy by over 50 points and occasionally pushing it below chance. The same change helps one judge and hurts another. Retrieval and larger reasoning budgets help little, and two purpose-built novelty evaluators are outperformed by our cheapest prompted baseline. These results raise questions about reported novelty gains of automated ideation systems, and call for robust novelty evaluation methods.

Abstract: Verifier-free reinforcement learning with probability-based rewards offers a promising way to train LLMs on general reasoning tasks where external verifiers are unavailable. Yet the reliability of these rewards, especially in long-horizon reasoning, remains underexplored. This work identifies a length-dependent failure mode of probability rewards, which we call the Posterior Concentration Phenomenon (PCP). We show that the probability of a reference answer conditioned on a reasoning trace often collapses to a low-variance interval as the trace becomes lengthy. This phenomenon results in nearly indistinguishable rewards, which, under GRPO-based settings, makes probability-based policy optimization unstable and inefficient. Motivated by this, we propose Reinforcement Learning with Concentration-aware Posterior Rewards (RLCPR), a verifier-free RL framework to explicitly account for PCP for better optimization stability and token efficiency. It has two components: uncertainty-aware data sampling, which reduces concentration-prone rollouts before generation, and concentration-aware regularization, which penalizes unnecessarily long traces when posterior rewards collapse. Extensive experiments show that, alongside higher token efficiency, RLCPR outperforms the state-of-the-art verifier-free RL baseline by up to 4.0% on six of seven benchmarks, including general-domain and mathematical reasoning challenges.

Abstract: An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.

Abstract: Black-box distillation is a practical route for transferring capabilities from API-accessible large language models that expose only text outputs into smaller student models. Recent on-policy adversarial methods such as GAD improve over SeqKD by forming an adversarial loop between a critic and a student, where the critic provides rewards for GRPO-based student policy optimization over the student's sampled responses. However, GRPO computes advantages from the within-group relative rewards of student samples for the same prompt, whereas the critic is trained primarily to distinguish teacher responses from student responses. This objective mismatch can produce reward groups with collapsed scale or fragile margins, leading to brittle grouped optimization signals. We propose Groupwise Reward Geometry Conditioning (GRGC), a two-stage framework that improves advantage construction by shaping student-side reward groups during both critic training and policy optimization. To improve critic-side conditioning, Gaussian groupwise Optimal Transport calibration regularizes the critic during training to produce reward groups with non-collapsed spread and smooth rank-wise gaps by matching sorted prompt-wise rewards to group-centered Gaussian quantiles. Building on this conditioned reward geometry, policy-side group power modulation reshapes the prompt-wise reward groups before they are converted into advantages, preserving the critic-induced ordering while increasing optimization-relevant margin separability. Extensive experiments across diverse teachers, student model families and scales, and training datasets demonstrate the effectiveness of GRGC on both in-distribution and out-of-distribution evaluations, while introducing negligible overhead over GAD. The code is available at https://github.com/2018cx/GRGC.

Abstract: During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine-tuning data mixtures and the difficulty of earning rewards through reward designing, enabling systematic comparisons of hacking dynamics and interventions. Experiments show that CATCH can produce diverse RL training trajectories with clear reward hacking, and analyses demonstrate that both initial models and reward difficulties shape the emergence of reward hacking. We further evaluate the effectiveness of different reward hacking detection and mitigation methods. A key finding is that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy model learn to mislead the monitor with code comments. This highlights the need to evaluate hacking mitigations throughout training with CATCH. The source code and resources are publicly released at https://github.com/THUAIS-Lab/CATCH.

Abstract: Long-horizon agent interactions generate useful but noisy experience, and retraining models to absorb it is expensive. Context-evolving agents therefore need memory extraction methods that improve with more test-time compute without relying on gold labels. We propose RefCon, which combines sequential self-refinement with parallel self-contrast to extract higher-quality memories without gold labels. Evaluated on AppWorld and BFCL-V3 across multiple context-evolving agent frameworks, RefCon delivers strong and consistent gains, including relative improvements of 21.6% on ACE and 16.6% on ReMe over no-scaling baselines, while a diversity-focused variant (DivCon) achieves a 35.5% gain on ReasoningBank. RefCon consistently outperforms existing baselines without ground-truth labels, and generalizes across model scales and to software engineering tasks, where it surpasses even ground-truth baselines. We further analyze the accuracy-token trade-off and scaling behavior, showing RefCon maintains favorable efficiency and continues to improve as more trajectories are used, unlike diversity-only scaling which saturates earlier.

Abstract: Frontier LLMs are increasingly used to automate scientific research through iterative search. We distinguish idea-driven search from solution-driven search and identify three core challenges: organizing evolving research ideas, selecting promising directions, and maintaining alignment between ideas and their implementations. To address these challenges, we introduce the Agentic Idea Manager (AIM), a fully autonomous framework for managing and exploring research directions in idea-driven automated research. Inspired by Bayesian optimization, AIM uses an Agentic Surrogate and an Agentic Acquisition mechanism to organize discovered ideas and guide their selection. A Solution Auditor maintains idea-solution integrity, while a Resource Planner adaptively allocates the remaining experimental budget across parallel search branches. Experiments on 10 AutoLab benchmark tasks show that AIM surpasses the strongest baseline by 1.6 percentage points on System Optimization tasks and 4.9 percentage points on long-horizon Model Development & CUDA tasks. Notably, AIM reaches the best baseline performance up to 3.1x faster in wall-clock time. We further provide a theoretical analysis of when searching over ideas becomes beneficial. Our analysis shows that explicit idea-level allocation makes semantic coverage directly controllable, and that broader coverage becomes increasingly valuable when competitive research directions are sparse among many plausible alternatives. Project Page: https://imhgchoi.github.io/agentic-idea-manager/

Abstract: Modern agentic systems combine an AI model with a harness that controls execution and environmental interactions. Harness design strongly affects long-horizon performance, yet its combinatorial search space demands substantial human effort that must be repeated as models change. Existing automated methods explore this space narrowly, optimizing only components such as prompts or skills or becoming trapped by fixed, exploitative search strategies. We introduce MILO (Meta-evolutionary Island Orchestration), a framework that co-evolves agent harnesses and the strategy used to discover them. MILO combines: (i) hierarchical lineage memory over island-based trees, using rejected mutations as negative evidence; (ii) per-island mutator agents that rewrite complete harnesses using global search history and parent-specific feedback; and (iii) an orchestrator that adapts search through lineage grafting and speciation, mutator reassignment and curriculum revision. Across Terminal-Bench 2.1, PaperBench, and DeepSWE, MILO-discovered harnesses outperform eight state-of-the-art harnesses and six search methods using frontier (Opus 4.8) and open-weight (gpt-oss-120b) models. With Opus 4.8, MILO improves resolution over its initial harness by $+12.0\%$, $+28.3\%$, and $+10.3\%$, respectively, compared with best prior-search gains of $+4.5\%$, $+18.3\%$, and $0\%$. On Terminal-Bench 2.1, it achieves $86.1 \pm 2.0\%$, exceeding the official leaderboard's top entry ($83.8 \pm 2.3\%$) while using 26\% fewer tokens than its initial harness. On EinsteinArena open problems, MILO improves best-known upper bounds for Erdős minimum-overlap ($0.3808586 \to 0.3808568$) and the first and third autocorrelation inequalities ($1.50274365 \to 1.50274360$; $1.45081 \to 1.44889$).

Abstract: Interactive agent benchmarks and multi-turn reinforcement learning increasingly place a second language model in the role of the user. This simulated user controls what information the agent receives and when, yet current benchmarks score only the agent and do not directly measure whether the user correctly executed its assigned role. We introduce UserProxyBench, an evaluation layer over the tau-bench family, and the User Fidelity Score (UFS), which measures adherence to the benchmark's private user instructions using task-grounded rubric criteria scored independently of agent success. Holding the agent fixed at GPT-5.5 and varying only the user proxy across 375 enterprise tasks changes mean task reward by 15.2 points, while 24.4% of successful episodes contain a user-specification violation. The dominant failure is premature disclosure: users provide information before it is requested. This behavior has little effect on task reward, yet among successful episodes it causes the agent to make 1.06 fewer tool calls on average, changing the interaction being evaluated while preserving the reward. Finally, across seven proxies we identify an empirical cost-fidelity frontier, enabling practitioners to select the least expensive simulator that satisfies a required fidelity level.

Abstract: Machine unlearning has emerged as a critical post-hoc safety measure to erase sensitive concepts from Text-to-Image (T2I) models without prohibitive retraining. However, we reveal that current state-of-the-art (SOTA) approaches are brittle due to a severe lack of robustness to noise initialization. We call this phenomenon ``probabilistic forgetting'': suppressed concepts re-emerge under specific random initial noise conditions, despite appearing unlearned on other initializations. We trace this failure to the misalignment between standard Gaussian sampling during unlearning and the unlearning objective. Since the target concept manifests only in specific initial noise regions throughout the unlearning phase, uniform random sampling yields sparse, uninformative gradient updates that fail to drive robust erasure. To overcome this issue, we propose an adaptive, concept-conditioned sampling strategy that dynamically concentrates gradient updates on regions where the target concept manifests, down-weighting uninformative areas. We integrate our framework with six distinct SOTA unlearning methods across four diffusion backbones and evaluate it across safety, object, and artistic-style unlearning, as well as under black-box and white-box adversarial attacks. Our method reduces the conditional nudity re-emergence rate across random initializations by 67.2% on average over four baselines and lowers attack success rates across both adversarial evaluations. Across concept domains, Adaptive Noise Sampling strengthens adversarial robustness and non-target retention while preserving competitive generative quality and target-erasure performance.

Abstract: Thorough evaluation of vision-language models (VLMs) has become prohibitively expensive, as benchmarks span an ever-broader spectrum of capabilities and new models arrive at a relentless pace. Benchmark compression methods that preserve model rankings at a fraction of the cost are well studied for language models, but for VLMs the question remains under-explored. We present PRIMEBench (Pruning Redundant Items for Multimodal Evaluation), a vision-aware hierarchical benchmark compression framework that substantially reduces evaluation cost while preserving model rankings. This hierarchical framework operates in four stages: data cleaning to remove items answerable without the image and all-correct items, category representative selection to pick one benchmark per capability category, item pruning with Vision-Aware Variance (VAW), and category-count pruning. VAW combines inter-model variance with a vision-dependence score computed from multimodal embeddings alone, while encouraging coverage of diverse items within each benchmark. On models held out from item selection, it has the highest mean fidelity at the released 5% retention. The hierarchical design lets practitioners stop at any stage to match their compute budget; the released suite removes over 97% of items while preserving model rankings. Beyond compression, our analyses show how VLM evaluation behaves as model panels grow and evolve, providing guidance for designing future benchmarks that are more efficient, robust to model turnover, and explicit about the limits of evaluation-side pruning.

Abstract: Rubric-based evaluation is widely used to assess LLM-based systems by decomposing response quality into task-specific scoring criteria. However, automatically generating rubrics that reliably capture task-specific quality requirements remains challenging. We introduce Mubric, a mutation testing-guided approach to rubric generation. Mutation testing, a classic software testing methodology, evaluates a test suite by injecting faults into programs and checking whether the tests detect them. We draw an analogy between test suites and rubrics: if a rubric captures an important quality requirement, introducing a corresponding defect into an otherwise high-quality response should reduce its score. Mubric first mines common defects from real pairs of preferred and dispreferred responses and abstracts these defects into reusable mutation operators, each specifying how to introduce a particular type of response defect. For a new task, it applies relevant operators to a reference response, checks whether the injected defects reduce response quality, and uses insufficiently penalized defects to refine the rubric. We evaluate Mubric on 703 tasks across four representative domains against six advanced rubric generation methods. Mubric achieves the highest overall evaluation accuracy, outperforming the strongest baseline by 7.48 percentage points.

Abstract: Tool-using agents are entering settings where a wrong action carries real cost, and the benchmarks certifying them grade what each simulated tool call reports having done, assuming the tool did what its interface advertises. The audit taxonomies we survey publish no category for that assumption, and a defect beneath a score is present on every rerun. We treat a tool's advertised surfaces as an executable contract, check the implementation against it, and trace each score's provenance through the task files and evaluator code to the verdicts that derive from state a defective tool should have written. Across 34 audited mutating tools in four benchmarks we confirm seven tool defects and one evaluator property at pinned commits. On injected defects the checker raised no false positive in 25 flags, flagged 2 of 5 negative controls, and missed most: in 29 of 33 scored misses a clause covered the defect but no probe revealed it. The checker's own static half, run alone, flags 14 of 17 confirmed sites, so on these findings the dynamic half confirms and traces rather than discovers. Twelve further AgentDojo tools, with six held-out tools and the seven audited first, complete its 25-tool mutating surface, on which at least 5 tools diverge from their advertised surface as our contracts read it, a rate for AgentDojo alone. No gold trajectory reaches either tau2-bench defect; on 1,120 paths built to isolate the telecom defect, a number fixed by construction, the evaluator rewards a refuel of a suspended line and fails the repaired tool. The clearest case is a clinical benchmark whose tool tells the agent each write executed under a documented no-write design its interface does not disclose; its grader takes that message as evidence, so its action success rate records whether a request carried the expected payload, not whether any record changed.

Abstract: Large language models (LLMs) are increasingly used for software engineering tasks, yet their stochastic behavior challenges the validity, reproducibility, and comparability of their evaluations. Conventional practices such as reporting a single output, an average score, best-of-N, or pass@k performance can obscure variability and estimation uncertainty, potentially leading to misleading conclusions about system reliability. This article presents SafeLLM4SE, a practical methodology and reporting standard for statistically principled evaluation of LLM-based software engineering systems. Rather than treating generated outputs as deterministic artifacts, SafeLLM4SE treats them as realizations of a stochastic process and distinguishes quality, stability, and estimation uncertainty. It combines adaptive sampling with confidence intervals, distribution-aware statistical comparisons, effect sizes, and a minimum reporting standard covering model configuration, reproducibility, evaluation procedures, and resource usage. SafeLLM4SE is also provided as an open-source software package available on PyPI, enabling researchers and practitioners to reproduce and extend the methodology. We illustrate its application by comparing two LLMs on HumanEval, a benchmark of programming problems assessed through functional tests.

Abstract: While Large Vision-Language Models (LVLMs) achieve remarkable success, hallucinations remain a significant barrier to their reliable deployment. Recent studies primarily attribute these issues to cross-modal attention imbalances; most solutions therefore focus on reweighting visual tokens or suppressing language priors. However, such approaches often overlook the spectral characteristics of the visual information flow and frequently rely on Contrastive Decoding (CD), which doubles inference time. Instead of following conventional approaches, we identify two distinct hallucination patterns-Perceptual-Semantic Dissociation and Localized Fixation-and propose FLASH (Frequency-Localized Attention SHaping), a training-free and CD-free framework. FLASH utilizes a Spectral Vortex Score to detect vision heads within multi-head attention layers and applies adaptive spectral modulation to rectify the visual information flow during decoding. Empirical results demonstrate that FLASH achieves a superior balance between performance and efficiency compared to SOTA methods.

Abstract: LLM-as-a-Judge is increasingly used to evaluate policy responses on open-ended tasks that lack ground-truth answers. Existing work often directly converts the resulting judgments into reward signals for policy training, paying limited attention to intrinsic judgment quality and largely restricting the use of Judges to training-time supervision. We systematically investigate judgment quality and downstream utility by examining both how judgments are elicited and how they are used. For judgment elicitation, we vary the Judge protocol along three dimensions: verdict granularity, critique usage, and evaluation batching. For judgment usage, beyond policy training, we extend Judge to test-time inference through Best-of-N selection, Judge-guided revision, and beam search. We find that, (i) Surprisingly, judgment quality and downstream utility do not always align. (ii) Judge protocol design substantially affects both intrinsic judgment quality and downstream utility. (iii) Judge guidance effectively converts test-time compute into performance gains, with benefits varying across inference strategies. Our results call for a multifaceted evaluation of LLM Judges on open-ended tasks, encompassing intrinsic judgment quality, and downstream utility.

Abstract: Skill evolution improves the capabilities of large language models by analyzing trajectories generated under a given skill and modifying the skill accordingly. Existing approaches typically generate a single trajectory per question. However, this provides insufficient optimization signals since it requires inferring effective skill edits from a solitary path. It is difficult to pinpoint which actions caused the failure in a failed trajectory, or to determine which actions in a successful one should be incorporated into the skill. Furthermore, they rely on a local batch of trajectories for analysis, making the optimization direction susceptible to noisy evidence. To address these, we propose SkillCome, a Skill-evolution method based on group Contrast optimization with dual memory. For each question, SkillCome generates trajectories and performs group contrast analysis to precisely identify key behavioral divergences between successful and failed trajectories, offering reliable optimization signals. The dual memory system further accumulates evidence from historical steps to track patterns shared across different groups, leading to more generalized optimization directions. Together, SkillCome builds a systematic optimization process that transforms experience from observed successful trajectories into reusable skills. Extensive experiments on six benchmarks spanning question answering, reasoning, and agentic tasks demonstrate the effectiveness of our method. SkillCome consistently outperforms baselines across five models of varying families and scales, with gains up to +5.69 points.

Abstract: Reproducibility is essential for scientific research, yet prior work shows that LLM outputs vary with hardware and batching. We identify an overlooked factor: the hidden injection of the current date into system prompts, which users cannot control and which changes every day. Across 9 recent LLMs and 6 datasets spanning multiple-choice QA (MCQA), math reasoning, code generation, and machine translation, performance varies solely with the current date, with deltas of up to 6% on MCQA, 14% on math reasoning, 7% on code generation, and 2.84 BLEU on machine translation. Model rankings also shift, affecting leaderboards. This date effect exceeds other sources of non-determinism, such as batch size and numerical precision. Standard prompting techniques -- chain-of-thought and few-shot prompting -- do not reduce the sensitivity; chain-of-thought even amplifies it. Our findings underscore the need for careful evaluation protocols to ensure reproducibility and fair comparisons in LLM research.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

Abstract: Self-evolution lets large language models (LLMs) improve iteratively using their own generated data, but often suffers from self-evolution degeneration: performance improves, plateaus, then declines. Existing methods address this issue at the component level, targeting either the Questioner or the Solver, and overlook that self-evolution is a tightly coupled system. We propose a holistic framework based on learnable information gain, which measures how much novel, parameterizable information a round provides relative to the previous round. Theoretically, this gain equals the Kullback-Leibler divergence between the two rounds' data distributions plus their entropy change. Practically, it is estimated by fitting a small language model to the previous round and scoring new data via negative log-likelihood. Based on this diagnostic, we propose ATRI (Adaptive Training Regulation via Information-gain), which reweights samples within a round and halts training across rounds when information gain remains low. Experiments on popular datasets demonstrate the superiority of our proposal.

Abstract: Test-time scaling improves software engineering agents by generating multiple candidate trajectories and selecting the best one. Verifying and selecting among these long interactions can consume as many tokens as generation itself. Existing hybrid workflows first apply an LLM-based execution-free (EF) verifier to filter candidates before running tests, which adds another model pass over every trajectory. We introduce LatentSift, a token-free and execution-free filter that replaces this first stage with hidden states the policy already produces while generating the candidates. It represents each candidate through its reasoning, observation, and function-call states, compares them with positive and negative banks of such states collected from successful and unsuccessful trajectories during policy training, and fuses the resulting distance scores with a learned linear score to retain promising candidates for the execution-based stages. On SWE-bench Verified, across three agents and two policy sizes, LatentSift cuts EF-verifier tokens by 66.6--81.0% and total verification tokens, which include test generation, by 49.1--62.1% at K=16, while hybrid Best@16 matches or improves on each agent's reference workflow, rising from 59.26% to 60.06% on DeepSWE-Preview.

Abstract: Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable reasoning. Grounded in cognitive science, we model creativity along three axes: action (what to do and how creatively), process (when to explore versus exploit), and outcome (the novelty and usefulness of the resulting idea). We use these axes to train models with GRPO, exposing them to varying degrees and forms of creativity throughout training. This produces substantially more diverse scientific proposals, expanding the range of research directions by 27.8% and contribution types by 14.9% over the base model. It also improves predicted citation impact by up to 32.0 percentage points and originality by 66.2 points. These gains cannot be reproduced by simply increasing decoding temperature; instead, we find that semantic guidance specifying what kind of creativity to pursue is critical. Overall, our results suggest that creativity is a learnable, multi-level ability that can be shaped to help researchers reach ideas beyond those typically explored by LLMs.

Abstract: Reinforcement learning can turn one language model into several specialists, each excellent at a single skill such as mathematics, coding or following instructions, but users need one model with all of these skills. Multi-teacher on-policy distillation (MOPD) merges them by letting the specialists teach one student: the student answers each prompt, and the specialist for that prompt's domain gives feedback on every token. This routing decides which specialist teaches, but not how strongly its feedback moves the shared student. In Qwen3.5 models at three sizes, we find that MOPD's student does not beat one taught by the best single specialist and gains little of the mathematics specialist's advantage. The feedback is unbalanced: instruction-following feedback is several times more spread out than mathematics feedback and dominates the student's updates. We propose Domain-Normalized MOPD (DN-MOPD), which keeps the routing and rescales each domain's feedback by its measured spread. On six public benchmarks, DN-MOPD improves the average score over MOPD at every size, across three random seeds and under two answer-length limits, and recovers most of the lost mathematics gain. Controls with fixed domain weights show that the gain comes mainly from turning down instruction-following feedback rather than turning up mathematics alone, and that fixed weights close to those DN-MOPD measures perform comparably. Combining specialists therefore requires deciding not only which one teaches, but also how strongly its feedback counts.

Abstract: Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present COEVOLVE, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, COEVOLVE enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, COEVOLVE attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers

Abstract: Advances in multimodal understanding, reasoning, and tool use enable agents to tackle increasingly complex visual reasoning tasks. By distilling past execution experience into reusable skills, agents can transfer lessons from both successes and failures into future reasoning, reducing repeated errors and improving capabilities. However, limited experience may produce unreliable, poorly generalizable skills, while static datasets may lack the targeted and diverse practice needed for refinement. To address this gap, we introduce V-Gym, an autonomous framework that iteratively co-evolves procedural skills and multimodal practice data from execution trajectories. During skill evolution, V-Gym analyzes trajectories to distill and refine hierarchical skills, updating procedural guidance and applicability conditions while retaining an update only if it improves validation performance. During data evolution, V-Gym selects generation seeds by balancing data utility and exploration, then translates trajectory-identified bottlenecks into diverse, targeted practice data that expand the data bank after quality checks. The resulting practice outcomes feed back into subsequent skill updates, closing the loop for continual skill refinement. Experiments across diverse multimodal reasoning benchmarks show substantial improvements over baselines with multiple backbone models. Its evolved skills generalize across domains and models, while evolved data support more effective skill refinement, enabling autonomous diagnosis, targeted practice, and continual self-improvement.

Abstract: Large language model (LLM)-based scientific agents are increasingly used for scientific problem solving, yet their robustness to imperfections arising during multi-turn interactions remains poorly understood. We introduce \textsc{SciARP} (\textbf{Sci}entific \textbf{A}gent \textbf{R}obustness to \textbf{P}erturbations), a benchmark for evaluating scientific agents under scientifically plausible perturbations throughout multi-turn problem solving. \textsc{SciARP} transforms 620 scientific problems into interdependent tasks of 3--13 turns and defines 13 perturbation types spanning problem understanding, evidence processing, reasoning, and conclusion formation. Clean and perturbed versions of each task are independently executed under matched settings, producing paired live trajectories for evaluating both task success and process reliability. Experiments across eight LLMs from four model families reveal three key robustness characteristics. First, different classes of scientific perturbations exhibit distinct robustness profiles and can decouple task progression from scientific reliability: agents may continue advancing through the task even after their information or reasoning has become unreliable. Second, stronger clean-task performance does not necessarily translate into stronger robustness, as models with higher clean-task accuracy can exhibit larger degradation under perturbation. Third, perturbation effects exhibit strong temporal dynamics: they may remain latent for multiple turns before emerging and subsequently propagate through downstream dependencies. Together, these findings show that current scientific agents remain insufficiently robust to scientifically plausible perturbations, with failures often remaining undetected, propagating, and resisting recovery.

Abstract: Large Language Model (LLM) based Multi-Agent Debate (MAD) is one of the most effective test time scaling techniques. Through multi-round communication, agents complement each other in knowledge and reasoning and solve tasks that no single member can solve. However, existing MAD frameworks fail to beat strong Single Agent and Consistency-based baselines under the same strict cost limit, which shakes the foundation of the MAD field. We propose Conditional Progressive Pruning (CPP), a lightweight pruning framework that fully exploits multi-round MAD. CPP outperforms all existing MAD frameworks on multiple dominated benchmarks. It is also the first to fully outperform consistency methods. Our code, detailed agent interaction records will be released soon.

Abstract: Rapid progress in Large Language Models (LLMs) is saturating static benchmarks faster than they can be designed. While existing automated evolution frameworks attempt to generate harder questions by perturbing individual tasks, they remain constrained by rigid, hard-coded generation rules. Moving beyond the evolution of isolated tasks, we propose to optimize the benchmark generation workflow itself end to end with MetaBench-Harness, a dual-loop search framework. Specifically, the inner loop utilizes a benchmark harness to generate a new benchmark in each round, while the outer meta-harness orchestration layer iteratively refines and searches over harness implementations based on historical evolution trajectories. By applying MetaBench-Harness to the competitive programming CodeContests and Olympiad mathematics AIME-2024 datasets, we demonstrate that the evolved benchmarks are challenging and discriminative for frontier models. Trajectory and quality analyses verify that MetaBench-Harness enables multi-dimensional evolution, steadily improving evolution reasonableness, benchmark competency, and evaluator robustness across successive rounds. Furthermore, case studies reveal its effective utilization of diverse difficulty levers to reframe problems and elevate required capabilities. Ultimately, this work provides a solution to the pressing challenge of benchmark saturation.

Abstract: LLM agents require reliable step-level evaluation to compare candidate branches and allocate computation effectively. However, lightweight evaluation remains challenging. External verifiers introduce additional inference cost, while agent-produced confidence or self-evaluation scores can be miscalibrated, especially when candidates are generated by heterogeneous agents. We propose \textbf{G}eneralized \textbf{L}ayer-wise \textbf{I}ntrinsic \textbf{D}istributional \textbf{E}valuation (\textbf{GLIDE}) for LLM agents. \textsc{GLIDE} derives intrinsic step evidence from layer-wise residual coherence, which measures whether local residual updates consistently support the global residual change induced by a candidate step. It calibrates this evidence against the recent score distribution of the generating agent and converts it into a pessimistic reward that jointly accounts for absolute residual evidence and agent-relative standing. The reward provides a cross-agent value signal for MCTS branch selection, while normalized predictive uncertainty guides adaptive branching. Experiments on multi-hop reasoning, sequential decision making, and symbolic logic show that \textsc{GLIDE} improves task performance, step-level ranking quality, and computational efficiency without external verifiers or task-specific supervision.

Abstract: Human judgement is the reference measure for evaluating generative models, yet the software used to collect it lags behing the methodology. Researchers adapt listening-test frameworks designed for perceptual protocols such as MUSHRA, rely on closed commercial survey platforms, or implement single-use web applications. Live arenas such as Chatbot Arena and Music Arena rank publicly deployed systems at scale, but do not support controlled comparisons of a laboratory's own models with its own participants. We present PANEL, an open-source, self-hosted platform for such studies. A study is authored in the browser and distributed as a single link, with audio, video, image, and text stimuli, seven question types, and screening and skip logic. The platform reports per-question summaries, across-condition significance tests, pairwise win rates and Bradley--Terry scores, and supports power analysis from pilot data. Consent versioning, self-service withdrawal, retention enforcement, and audit logging support GDPR-compliant operation. Each study exports as a machine-readable specification. PANEL is available at https://github.com/matteospanio/panel.

Abstract: The rise of autonomous AI agents equipped with tools has introduced significant security risks, ranging from unintended tool misuse to adversarial manipulation through Indirect Prompt Injection (IPI) attacks. In practice, deployed agent systems such as OpenAI Codex and Claude Code protect tool invocations through a combination of coarse-grained permission rules and LLM-based judgments about individual proposed actions. Both components, however, have important limitations: static policies must anticipate possible user intents and therefore do not scale to open-ended tasks, while LLM-driven authorization supports dynamic decisions but produces inconsistent outcomes and remains vulnerable to targeted IPI attacks. To provide scalable and more consistent authorization, we propose MetaPermit, a policy-based tool access-control framework that decouples semantic inference from security enforcement. By analyzing agent-user interactions, we derive a compact, task-independent set of meta-attributes that capture the relationships among the user's intent, the execution context, and the proposed tool call. These meta-attributes allow MetaPermit to authorize tool use without enumerating user intents. At runtime, an LLM infers the meta-attribute values for each proposed tool call, while a fixed policy evaluates these values to allow or deny the call, making each decision auditable through the inferred values and the applied policy rule. We evaluate MetaPermit on the AgentDojo and AgentDyn benchmarks, across seven task suites and five attack methods, using two widely deployed open-weight LLMs. The results show that MetaPermit produces 31% more consistent authorization decisions than LLM-driven authorization and outperforms the state-of-the-art defenses CaMeL and IPIGuard in both task completion, with improvements of up to 109%, and robustness to IPI attacks, with no malicious tool calls executed.

Abstract: Language-model agents increasingly improve by converting execution experience into reusable external skills. Yet repeated skill updates form a learning process of their own: locally useful edits can accumulate into redundant or task-specific instructions, while new updates can disrupt behavior that previously worked. We study this problem as skill-evolution overfitting and introduce SkillEvoReg, a general regularization framework for skill evolution inspired by anti-overfitting techniques in neural-network training. SkillEvoReg combines training-time skill dropout, which perturbs update generation, and complexity-aware local regularization, which controls unnecessary structural growth, with causal counterexample validation (CCV), which provides targeted behavioral validation of candidate-specific regressions. We instantiate the framework across heterogeneous skill-evolution systems while retaining each system's native skill evolver and task evaluator. Across SkillOpt, SkillEvolBench, and ContinualSkillBench, SkillEvoReg consistently controls skill-state growth while preserving competitive downstream capability, improves several transfer and later-stage evolution outcomes, and identifies update-level regressions that structural metrics alone cannot reveal. These results suggest that explicit regularization is a useful complement to increasingly capable skill updaters.

Abstract: Evaluating explainable Artificial Intelligence (XAI) methods is a challenging task due to the lack of reliable evaluation procedures and, in particular, the absence of ground truth explanations. In the literature, existing evaluation approaches typically assess explanations by measuring their fidelity with respect to the predictions of a black-box model. However, such evaluation strategies only quantify the degree to which an explanation reproduces the model's output, without ensuring that the explanation correctly reflects the underlying decision process. As a consequence, different explanations may achieve similar fidelity scores while providing inconsistent or misleading interpretations of the model behavior. In this paper, we propose a framework for the evaluation of XAI methods based on synthetic ground truth. The proposed approach relies on controlled interventions to generate synthetic datasets in which the importance of input components can be determined by design. This enables the construction of ground truth explanations that are directly aligned with the behavior of the model under analysis. The framework is instantiated across three data domains, namely binary images, tabular data, and time series, allowing a comprehensive assessment of explanation methods in heterogeneous settings. Experimental results obtained by evaluating nine widely used XAI methods show significant limitations in current techniques and highlight the importance of synthetic, intervention-based benchmarks for a reliable assessment of explanation quality.

Abstract: Bias in AI systems is typically framed as a flaw to be minimized, yet it also serves as a critical indicator of underlying weaknesses in data, modeling assumptions, and system design. Existing approaches often treat bias as an isolated problem rather than as evidence that can strengthen verification and governance across the AI lifecycle. This paper aims to reconceptualize bias as a diagnostic tool that supports rigorous AI verification. We seek to develop a multidimensional framework to analyze bias, demonstrate how biases emerge in both Traditional and Generative AI, and provide a structured pathway for verification-driven mitigation. We present a multidimensional framework analyzing bias across four dimensions: origin sources, emergence points throughout the AI modeling lifecycle, technical and methodological causes, and validation approaches for detection and mitigation. Through a comprehensive typology spanning traditional and generative AI systems, we demonstrate how biases manifest and propagate across development stages. Our analysis encompasses 30 distinct bias types, 16 verification methods, and 20 countermeasures, providing an actionable roadmap for practitioners. We introduce a hierarchical evidence framework that distinguishes internal validity (mechanistic integrity of AI systems) from external validity (contextual reliability in deployment environments). The framework reveals how biases manifest and propagate across modeling stages, enabling systematic mapping between bias types, verification techniques, and effective countermeasures. The proposed evidence hierarchy clarifies how different verification strategies contribute to mechanistic integrity and contextual reliability. We advocate for ''Ethics by Design'' principles that integrate bias verification throughout the development lifecycle, enabling the construction of fairer, more robust, and trustworthy AI systems.

Abstract: Rufus-Air is an open and reproducible post-training recipe on GLM-4.5-Air-Base (106B-A12B), organized as a serial pipeline of eight stages: SFT, Reasoning RL, Coding RL, Instruction-Following RL, General Agent, Coding Agent, Search Agent, and RLHF. We document the data, reward design, infrastructure, stage order, and stagewise results needed to reproduce the recipe. Stages progress from basic to advanced capabilities and from hard, verifiable rewards to softer judge-based signals. Training builds on open-source components and public data, much of it used as released, without new human annotation or an in-house distillation teacher. Our main findings are that (i) diverse, high-quality SFT establishes a strong capability floor; (ii) difficulty filtering keeps RL prompts within a productive learning range; (iii) reward reliability provides a practical principle for ordering stages; and (iv) infrastructure and engineering choices are part of the recipe, not just an implementation detail. Rufus-Air improves over the official GLM-4.5-Air post-trained release and is competitive with similarly sized open models.

Abstract: Multi-agent Reinforcement Learning (MARL) trains a team of agents that share one environment and learn their policies together. Training maximizes the team return, and a high return does not imply that the rewards are shared fairly among the agents in every episode. Testing is an established way to discover the failures of deep reinforcement learning, yet few methods address the fairness of MARL. In this work, we propose FairTest, a search-based testing approach that seeks the unfair executions of a MARL policy. The design combines search guidance with test prioritization. The guidance scores each candidate with three fitness functions. One measures the fairness of the runs already performed, another predicts the fairness from abstract states and fairness features, and the third reads the decision uncertainty from the policy. Crossover and mutation derive further candidates from the observed executions. The prioritization ranks the candidates by the predicted fairness and the decision uncertainty, so that the runs reach the candidates where failures are expected. FairTest is evaluated on three environments and two MARL algorithms, and four baselines are given the same budget. It detects the most fairness failures compared to three baselines with statistical significance and large effect sizes. The failure count exceeds that of the strongest baseline by 221% on average and coverage improves by an average of 23%.

Abstract: Scientific weak signals are early, low-visibility research directions that later become central to mature scientific topics, yet existing resources such as trend tracking, citation forecasting, and foresight reports rarely provide validated reference sets that link concrete early precursors to later paradigms. We introduce BackTrend, a retrospective benchmark in which, given a mature target topic and a temporal evidence constraint, systems must recover two types of precursors: problem-space signals, underrecognized research problems, and solution-space signals, emerging methods for known problems. BackTrend contains 25 mature target topics in artificial intelligence and machine learning and 66 human-validated weak signals, reconstructed from large-scale literature by grounding each candidate in its 2019-2024 publication-frequency trajectory. We evaluate frontier LLMs, RAG systems, and agentic research systems using semantic matching and coverage-based metrics. Current systems often generate plausible but misaligned precursors, exhibiting topic drift, granularity mismatch, near-miss matching, and incomplete coverage; the strongest system achieves only 10.1% F1, while Coverage10 reaches at most 18.5% of the reference signals. Our budget analyses show that additional retrieval and web-search evidence can improve performance up to a moderate budget, but does not by itself close the substantial performance gap.

Abstract: In recent years, large language models (LLMs) have emerged as a popular alternative for evaluation. Often referred to as LLMs as judges (LLJs), these systems have been widely adopted by researchers and practitioners across a broad range of measurement tasks, driven by their strong performance, scalability, and cost-effectiveness relative to human judgment. However, a growing body of work has shown that the use of LLJs raise concerns about their validity and reliability as evaluators. Existing efforts to address these challenges have largely focused on developing bias-mitigation techniques and refining prompting strategies. While these approaches represent an important step forward, they primarily offer technical fixes and leave a more fundamental challenge unaddressed: the lack of standardized, transparent, and reproducible evaluation practices. In this paper, we introduce LLJ Cards, a framework that synthesizes best practices from measurement theory, natural language generation, and machine learning literature into practical guidelines for LLJ-based evaluations. While LLJs offer a promising path toward scalable evaluation, their effective use requires grounding in rigorous evaluation principles to ensure validity, reliability, and reproducibility. LLJ Cards addresses this need by providing a structured framework for applying these principles in the design and reporting of automated evaluations.

Abstract: Many knowledge-work benchmarks are constructed around individual tasks, with the context needed for each task selected together with or after the task has been specified. This design measures performance on workplace-like tasks in an environment assembled for the task. When task specification guides which context is selected, the evaluation can encode task information into the environment and pre-complete part of the information-localization work that workplace performance normally requires. We introduce WorkWorlds, an evaluation infrastructure that separates organizational state from task specification. A world first fixes a revision, date, and employee seat and materializes the organizational state that employee can access; tasks are introduced only afterward. We implement WorkWorlds in a primary synthetic pharmaceutical company with 8 measured tasks across 6 employee seats, and construct additional organizational worlds. Across 192 matched evaluations, moving from task-curated context to the full role-visible workplace reduced evidence access from 90.4% to 74.5% and criterion pass from 79.4% to 68.2%, while pass conditional on evidence access remained nearly unchanged; most of the measured difference occurred before the agent reached sufficient evidence.

Abstract: A RAG system can hand you the right answer and cite a source it did not actually use. Models output these unfaithful citations via post-rationalization: they write the answer first and then attach a citation to whatever passage looks close enough. Search agents are now trained with reinforcement learning from verifiable rewards (RLVR), which pays them for getting the answer right. We asked whether that training also teaches them to cite honestly. Improving an existing methodology with a required control, we compared an instruction-tuned model against three RLVR agents trained from it, on four question-answering datasets, using only free-tier Kaggle GPUs. Post-rationalization is everywhere: on Wikipedia-based questions roughly one citation in seven is unfaithful. RLVR does not fix it. The agents post-rationalize at their base model's rate, and one lands slightly worse. Rewarding correct answers buys nothing in citation faithfulness, so faithfulness has to be trained and measured on its own terms.

Abstract: Collective intelligence depends not only on what team members know, but also on how they organize their work. When the structure of a solution is unknown, useful roles and divisions of labor cannot be specified in advance; teams must learn from experience how to organize reasoning as it unfolds. Human teams routinely adapt this way, while existing AI agent teams rely on fixed protocols, explicit task decomposition, or routing. We introduce Self-Organizing Agent Teams (SAT), fixed teams of AI agents that learn reusable strategies from prior collaborations to organize roles, conversational phases, participation, and information flow. These strategies enable what we call collaborative computation: agents exchange, challenge, repair, and synthesize partial reasoning into solutions no member produced independently. In two independent settings, we learn teamwork strategies that transfer unchanged to unseen benchmarks, using only 15 mathematics and 25 graduate-level knowledge problems. Across five mathematics and physics benchmarks, self-organizing teams average 66.7% accuracy, versus 48.8% for their strongest member, 58.7% for compute-matched inference by that agent, and 59.0% for a perfect router over members' independent answers; on AIME 2026, they exceed this router by 13.4 points. Because gains vary across benchmarks, we ask when self-organizing collaboration helps. Across eight benchmarks, demonstrability (the organizational-psychology construct of whether a team can distinguish correct from incorrect reasoning) strongly tracks improvement over the strongest member (Spearman $ρ=0.90$, $p=0.005$): teams benefit most when correct reasoning can be recognized once it appears. More broadly, these results suggest that organization itself can become an agent capability: agent teams can learn how to reason together and produce solutions their members could not reach independently.

Abstract: Multi-turn agents trained with reinforcement learning (RL) receive a single scalar reward per trajectory, which motivates self on-policy distillation (OPD) to supply dense token-level supervision from a self-teacher with privileged task skills, letting a skill-free student internalize them. This recipe, however, is undermined by two findings in agentic tasks: privileged information alone does not always make a teacher reliable, and the benefit of teacher supervision is stage-dependent. We therefore propose RetireOPD (Self-Retiring On-Policy Distillation), which first optimizes a decoupled, skill-conditioned teacher with environment rewards and then trains a skill-free student jointly with RL and OPD. Rather than following a predefined distillation schedule, RetireOPD adopts Adaptive Retirement: the student drops the teacher on its own once their discrepancy stops shrinking and it reaches a target fraction of the teacher's success rate, after which training proceeds with RL alone. Across Qwen2.5 models from 1.5B to 7B, RetireOPD improves ALFWorld success rate over RL baseline by 14.1% to 18.8% and WebShop accuracy by 11.8% to 19.0%, and surpasses its own skill-conditioned teacher in every setting.

Abstract: A memory can answer a current query correctly while discarding distinctions required by a later update. We investigate this failure with a paired-history audit: two histories have the same current answer, receive a shared future update, and require different subsequent answers. A pilot evaluates 24 history pairs across six synthetic mechanisms, 12 memory conditions, two repeats, and two model backends. A deterministic frontier selector obtains strict reveal accuracy of 96/96 on DeepSeek and 82/96 on GLM; a structured writer obtains 62 successes with one unresolved outcome and 56/96. The configured four-outcome joint contrast has finite-sample identification intervals of [0.521, 0.542] and [0.292, 0.313], not confidence intervals. A record-level audit distinguishes retained-state adequacy, response delivery, and answer-schema compliance without changing those original scores. It finds 26 and 25 well-formed but semantically wrong structured reveal memories, while all 14 GLM frontier reveal failures contain correct values in the wrong wrapper. Tombstone removal produces 16/16 exact replay failures in the targeted mechanism. Identifier renaming then exposes a separate flaw: original frontier late-reference adequacy falls from 8/8 to 94/320 transformed instances. We provide and test a label-equivariant repair, but it preserves only 2/8 original late-reference answers: eliminating a naming shortcut does not solve unknown future relevance. These results support a scoped evaluation methodology and reproducible failure analysis, not general superiority of the repaired algorithm. Paid pilot evidence, retrospective diagnostics, and new offline tests are reported separately; no independent held-out or natural-task validation is claimed.

Abstract: Reinforcement learning now trains language-model agents that act over dozens of steps in live environments. The gains are large, and they are read as better decision-making. An agent in a closed loop writes its own inputs. Each observation follows from its own earlier actions, so the states it meets late in an episode are partly of its own making. An SFT checkpoint and an RL checkpoint are then scored from different states, even on identical tasks. Endpoint success mixes two changes: where the agent arrives, and what it does once it is there. Restricting the comparison to states both policies reach does not separate them. That restriction selects on an outcome, and in our data it flips the sign of the effect. We introduce checkpoint handoff, an evaluation protocol that clones a state one released checkpoint reached and hands it to another, with no retraining. Crossing a reacher role and a solver role over SFT and RL splits an endpoint gain into REACH and SOLVE. REACH is how often a policy arrives at a state the environment confirms is a fixed number of actions from success. SOLVE is how often it finishes from an identical cloned state. Across two benchmarks and two independently released pipelines, the reacher by solver interaction is positive in all five conditions. An RL history is worth more to an RL solver than the same history is to an SFT solver. On ALFWorld, RL improves both terms, and the SFT solver never succeeds where the RL solver fails. Independent REACH and SOLVE gaps predict the aggregate interaction. Handoff asks only that one checkpoint's history can be replayed under another, so long-horizon evaluation can report arrival and completion beside endpoint success.

Abstract: Agent benchmarks are substantially more costly to evaluate than conventional LLM benchmarks. Benchmark compression is therefore a natural solution, yet existing methods primarily model redundancy in task--model final-score distributions, which is important in agentic evaluation. To address this limitation, we analyze large-scale trajectories and identify six complementary process signals that are systematically associated with final agent performance. To disentangle agent performance redundancy from a complete perspective, we propose DualViewEval, an agent benchmark compression method that jointly exploits outcome and process relations to learn an exact-size miniset and predict the full-benchmark scores. Across five agent benchmarks and five representative baselines, DualViewEval achieves the best results in all datasets. With only 20 tasks, it achieves $24\times$--$40\times$ compression on APEX-Agents and BFCL, reducing mean absolute error (MAE) by $14.5\%$--$28.2\%$ over the strongest competitors while improving Kendall's $τ$ by up to $7.2\%$ relative to EssenceBench on SWE-bench Verified. The selected minisets further reveal capability differences among different agents, providing compact and diagnostic feedback for efficient agentic model development.

Abstract: Test-time reinforcement learning (TTRL) enables models to improve their reasoning without relying on labeled training data, but existing approaches typically optimize a large fraction of the model parameters. This raises a natural question: can effective test-time adaptation emerge when both the reward signal and the optimization space are severely restricted? We answer this question with label-free bias-only TTRL, which uses majority-vote pseudo-labels as rewards and optimizes only approximately 100K bias parameters while keeping the pretrained backbone frozen. On MATH-500, our approach reaches 76.67% accuracy, slightly exceeding our own labeled bias-steering reproduction while optimizing 76,000x fewer parameters than full-parameter TTRL. The same training procedure improves performance across vision-language and audio reasoning tasks, including MathVista, AI2D, LogicVista, and MMAU. We further show that the learned steering vectors transfer to 4,500 held-out MATH problems, indicating that the adaptation is not limited to the problems used during test-time optimization. Finally, we analyze why this highly restricted adaptation can work, showing that majority-vote reliability improves with rollout consensus and that bias subspaces with greater accessible gradient energy exhibit stronger downstream trainability. These results demonstrate that substantial test-time adaptation can emerge from optimizing a tiny bias-only subspace using entirely label-free rewards.

Abstract: Agent skills, reusable procedural documents that extend LLM agents beyond their parametric memory, have become an important interface for deploying agents on real-world tasks. Community-maintained skill libraries built around this interface are growing rapidly. However, this ecosystem remains deeply English-centric: our audit finds that low-resource languages such as Swahili and Hindi have no in-language skill content, so retrieval often returns a skill written in a different language than the query, degrading accuracy and recall. A practical solution is to synthesize in-language skills for retrieval but the quality can be unreliable, so relevance in this setting alone often surfaces a related but unusable candidate. To address this, we propose M-SQE, a post-retrieval Multilingual Skill Quality Estimation framework that scores candidates via a Theory view for intrinsic quality and an Action view for task-grounded utility, unified into a domain-conditioned final score. We evaluate M-SQE across three skill-use domains: general, tool-use, and cultural tasks. Empirically, we build three-layer candidate skill pools mirroring today's ecosystem, where M-SQE's task success exceeds existing baseline's average by at least +3.5 points across three different retrievers. Particularly, M-SQE lifts the lowest-resource languages most (+12.9pp on Hindi and +5.6pp on Swahili) and achieves strong performance across all six culture regions, thereby moving agentic skill use toward linguistic and cultural equality.

Abstract: Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings, safety failures may emerge only after multiple turns, yet existing evaluations often reduce agent behavior to task or attack success, obscuring whether an agent acts, refuses, or remains appropriately calibrated as the interaction evolves. We introduce Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using agents. Blindspot evaluates complete user-agent-environment trajectories through adaptive adversarial interaction, stateful tool execution, and execution-grounded adjudication. Its current instantiation contains 22 attack families and 35 scenarios across seven domains, yielding more than 2,500 long-horizon trajectories with an average interaction length of 14.7 turns. Each trajectory is assigned one of five outcomes: Safe Completion, Correct Refusal, Unsafe Completion, Over-Refusal, or Indeterminate. Unlike fixed attack datasets, Blindspot is an extensible live-simulation framework in which attacks, scenarios, tools, policies, domains, and agent configurations can be added without redesigning the evaluation pipeline. We evaluate 13 proprietary and open-weight LLMs using eight metrics covering unsafe completion, appropriate refusal, benign utility, over-refusal, repeated-run robustness, and post-refusal failure. Preliminary results reveal substantial differences in safety-utility calibration across models and show that failures can emerge only after several initially safe interaction steps. These findings motivate treating agent safety as a trajectory-level property rather than a single-turn or binary success criterion.

Abstract: Artificial intelligence threatens to outrun human understanding and control. New mathematics is needed to design AI that is legible, steerable, and cooperative with humanity. I organize this invitation by mathematical field, so you can turn straight to your own: logic and game theory for cooperation; probability for agency and world-models; algebra and representation theory for learned features; analysis and geometry for generalization and training dynamics. Each section ends with an open problem that is accessible to a working mathematician with no prior experience in AI safety.

Abstract: Large language model agents are increasingly deployed for long-horizon task execution. However, current evaluation paradigms face three major limitations: terminal-only assessment ignores intermediate processes and struggles to localize errors efficiently and accurately, single-reference matching penalizes valid alternative solution paths, and post-hoc trajectory judging incurs high costs without the ability to halt failed runs early. To address these issues, we propose DynSTEER, a dynamic stage-wise trajectory evaluation framework for agents. DynSTEER segments rollouts into stages anchored by key completed actions, focusing evaluation on essential milestones with adequate context while enabling targeted strategy adjustments. It compiles a path-tolerant milestone graph from public task views to respect diverse legitimate strategies without leaking ground truth. Furthermore, it adaptively routes evaluation queries across multi-tier judges and halts unrecoverable executions online to curb resource waste. Experiments demonstrate that DynSTEER improves evaluation discriminability across LLM agents by 85.2\% over native evaluation, separates all model pairs with statistical significance, and saves 34.51\% of execution steps on failed rollouts.

Abstract: Mixture-of-Experts (MoE) has become a widely adopted architecture for Large Language Models (LLMs), as it improves model capacity while limiting computational overhead through sparse expert activation. This property makes MoE-based LLMs particularly attractive for resource-constrained distributed environments. However, federated fine-tuning of MoE-based LLMs remains challenging under heterogeneous client data. Since clients often correspond to different task preferences, directly aggregating their local updates may weaken expert specialization and introduce conflicting update directions on shared experts. To address these challenges, we propose FedTAR, a task-aware federated fine-tuning method for MoE-based LLMs. FedTAR establishes the association between local updates and task preference via routing outputs. Specifically, we apply Singular Value Decomposition (SVD) to both routing features and local updates to extract low-dimensional task coordinates and update directions. Based on the task coordinates, FedTAR performs intra-cluster aggregation among clients with similar task preferences and inter-cluster aggregation across different task groups. The aggregated update is then reconstructed through the learned task-to-update mapping, ensuring that the final update remains aligned with task-specific optimization directions. In this way, FedTAR preserves expert specialization and mitigates destructive interference among heterogeneous clients. We evaluate FedTAR on four benchmark tasks under different non-IID settings. Experimental results demonstrate that FedTAR consistently outperforms strong federated fine-tuning baselines and achieves state-of-the-art performance.

Abstract: Low reported scores on leading physics benchmarks, including those featured in the Artificial Analysis Intelligence Index (2026), suggest that frontier language models still struggle with advanced physics, a demanding test of their scientific reasoning and quantitative problem-solving abilities. Yet this impression does not always align with domain experts' experiences using these models in their work. We revisit these reported findings by evaluating frontier models on six widely used physics benchmarks and auditing them with experts, focusing on text-only problems with verifiable final answers. For each subfield of physics, faculty and graduate researchers with relevant expertise carefully review problem statements, reference solutions, and model responses to distinguish genuine model errors from grader errors, incorrect reference solutions, and ambiguous or underspecified questions. Most audited cases initially evaluated as incorrect reflect these benchmarking issues rather than errors in the models' physics reasoning. We then ask experts to address these benchmarking issues by correcting erroneous reference solutions and repairing or excluding flawed questions. We find that GPT-5.6-Sol's measured mean@4 rises from 47.3% to 78.7% on HLE-Physics and from 61.0% to 87.2% on CMT-Benchmark, while its corrected pass@4 reaches 94.4% on the 54 retained CritPt challenges. Corrected scores are computed on the retained evaluation subsets following expert review. Scores on the audited subsets of UGPhysics, PRISM-Physics, and PHYBench also rise substantially after correction. These findings suggest that current benchmarks substantially understate frontier models' ability to solve well-posed physics problems. Near-saturation on these closed-ended tasks highlights the need for more demanding, expert-validated evaluations.

Abstract: On-Policy Self-Distillation (OPSD) has emerged as a popular paradigm for large language model (LLM) self-improvement, allowing models to act as their own teachers by leveraging privileged information such as ground-truth solutions. However, recent findings indicate that OPSD can severely degrade the performance of LLMs on complex reasoning tasks: By forcing the student to imitate an artificially confident reasoning trace conditioned on privileged information, OPSD inadvertently suppresses expressions of uncertainty and penalizes the exploratory, self-corrective behaviors required to solve challenging problems. To address this, we introduce Negative Self-Distillation (NSD), a new framework that optimizes LLMs by diverging from flawed reasoning rather than imitating privileged solutions. Instead of relying on ground-truth answers or external supervision, NSD uses the model itself to generate a question-specific negative condition (eg, acting as a ``careless reasoner'') and pushes the student's distribution away from this self-generated negative teacher. Naively applying unlearning objectives to achieve this divergence is problematic, as flawed reasoning tokens are confounded with basic linguistic tokens; indiscriminately penalizing both risks catastrophically degrading the model's foundational language capabilities. We resolve this by designing a dynamic gating mechanism that automatically identifies and isolates reasoning-critical tokens, ensuring gradient updates target only behavioral flaws while preserving the model's linguistic priors. Empirically, NSD consistently outperforms OPSD and other label-free, self-bootstrapping reinforcement learning (RL) baselines.

Abstract: Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing skill optimization methods often rely on costly execution-based evaluation and substantial task data. We introduce \textbf{COBRA-Skills}, an efficient framework that formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space. COBRA-Skills couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising or informative candidates while continually refining the skill population from execution feedback. Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance among compared methods, while reducing optimization cost by 55--58\% relative to SkillOpt and using only 50 unique optimization examples per benchmark. Further analyses show that COBRA-Skills remains robust to changes in the agent harness and performs effectively when the target model itself is used for skill generation and refinement.

Abstract: Large Language Models (LLMs) can memorize and reproduce sensitive, copyrighted, or otherwise undesirable training content, creating privacy, safety, and regulatory concerns. Machine unlearning offers a practical alternative to full retraining, but many existing methods apply broad or fixed parameter updates that can degrade utility and remain brittle under deployment changes such as post-training quantization, where forgotten knowledge may partially re-emerge. We propose Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer-level unlearning framework that selects transformer layers using a forget-to-retain significance score. This score identifies layers with high influence on the forget set and low sensitivity to the retain set, allowing FOM-UL to concentrate updates where they are most effective while leaving most of the model unchanged. This targeted update strategy improves the forgetting-utility trade-off and provides an empirical path toward quantization-resilient unlearning by reducing the chance that small, diffuse updates are erased by low-bit rounding. Across TOFU, KnowUnDo, and MUSE-style evaluations, FOM-UL reduces residual memorization compared with strong GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines while preserving retain-set utility close to the vanilla model. Under 8-bit and 4-bit post-training quantization, FOM-UL maintains stronger memorization suppression and utility preservation than competing methods, and adversarial prompt evaluations show lower recovery of forgotten content. Overall, FOM-UL provides an efficient unlearning strategy that improves targeted forgetting, utility preservation, and deployment robustness without claiming formal guarantees of erasure.

Abstract: Code generation benchmarks are widely used to evaluate Large Language Models (LLMs), but benchmark data leakage into training sets can inflate performance and undermine evaluation validity. DetectLeak, a method specifically designed for code generation benchmark leakage detection, relies on perplexity scores to identify likely leaked samples. However, perplexity mainly reflects general familiarity with code patterns and may perform poorly on complex or rare samples. It also overlooks other useful signals, such as code similarity, functional correctness, and semantic representations. To address these limitations, we propose CGMIA (Code-Generation-specific Membership Inference Attack), a method for detecting leakage in code generation benchmarks. CGMIA fine-tunes a shadow model on a subset of benchmark samples to construct labeled member and non-member data. For each sample, it collects the input prompt, generated code, and reference solution, and extracts expert features, including CodeBLEU, edit distance, test pass rate, and perplexity, together with semantic features from CodeBERT embeddings. An integrated learning module combines these features to capture both surface-level memorization signals and deeper behavioral patterns, enabling a classifier to predict whether a sample was included in the target model's training set. Experiments on eight code generation benchmarks show that CGMIA outperforms eight existing membership inference methods in most cases. It also effectively detects known leaked APPS samples in StarCoder-7B's training data.

Abstract: Autonomous agents powered by large language models (LLMs) continuously accumulate experience through interaction, creating an opportunity to improve future behavior through self-evolution. A fundamental challenge is how to transform abundant, task-specific interaction experience into reusable model competence without sacrificing the ability to adapt rapidly to newly observed evidence. Explicit textual states, such as skills and agent harnesses, provide fast, human-readable and editable adaptation, but incur persistent dependence on external context; parametric policies provide compact and reusable competence, but are substantially slower to update. We present \textit{Experience Funnel}, a self-evolving framework that couples fast state adaptation with slow policy consolidation in an alternating loop. Interaction trajectories are first distilled into an explicit textual state, where newly acquired experience can be rapidly incorporated and validated. The framework then selectively identifies state-enabled behavior that remains useful across state revisions and consolidates it into the policy through transition-aware distillation. The updated state--policy pair subsequently generates new rollouts, providing fresh evidence for the next round of state adaptation and policy consolidation. Experiments across diverse agent benchmarks show that \textit{Experience Funnel} consistently improves agent capability over state-only evolution and policy-internalization approaches, while progressively converting useful explicit experience into autonomous policy competence.

Abstract: Multi-agent large language models solve complex tasks by coordinating several policies in a shared environment. However, existing reinforcement learning methods usually optimize each response or trajectory separately, even when several outputs jointly cause one state transition. Consequently, the update unit differs from the action executed by the system. To address this problem, we propose SRPO (Setwise Relative Policy Optimization), which treats the active set the minimal set of outputs consumed by one transition, as one multi-agent action. Specifically, SRPO combines member log-ratios into one cardinality-normalized set ratio, assigns one relative advantage, and clips the set once. This formulation unifies division of labor and joint co-evolution as actions with different set sizes. Experiments on mathematical reasoning and multi-turn search demonstrate one training interface for fixed, mixed, and dynamically routed workflows across four model scales, with the strongest macro-average results among the reported comparisons. Optimization diagnostics further characterize its stability under different event reductions and set sizes.