Chat
-
LLM Agents for Automated Post-Training via Structured Recipe Optimization
2026-10-13 15:39:34Ruohong Liu, Bohan Tang, Lekang Jiang, Jack Umenberger, Yiwen GuoAbstract
Automated post-training transforms pre-trained large language models (LLMs) into models capable of handling diverse domains by automatically constructing and optimizing a recipe of coupled decisions over data sources, data processing, cross-domain composition, and training configuration under a limited experimental budget. Existing automation either searches predefined recipe spaces that enable systematic optimization but remain fixed rather than adapting their available choices to the target tasks, or uses open-ended agents that discover task-specific choices but leave them implicit in trial-level edits, limiting their systematic reuse and recombination across trials. To fill this gap, we introduce optimization over a task-adapted recipe space, a novel paradigm that can adapt the recipe space itself to the target tasks while representing its choices explicitly for systematic reuse, recombination, and joint optimization. We instantiate this paradigm with Structured Recipe Optimization (SRO), comprising an LLM-based Recipe-Space Designer that constructs the task-adapted recipe space and a Multi-Domain Recipe Optimizer that uses structured trial evidence to guide LLM candidate construction and multi-objective Bayesian trial selection. Extensive experiments show that SRO achieves the highest average score in all evaluated target-model and agent-LLM combinations. In joint post-training across three capability domains and six benchmarks, SRO reaches 41.1 versus 15.4 for the strongest baseline on Qwen3-1.7B with GPT-5.4, and 53.8 versus 40.8 on Qwen3-4B with DeepSeek-V4-Flash. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18572
PDF
-
Aligning Faithfulness for Theory-Scale Autoformalization
2026-10-13 14:53:04Yuming Feng, Tadd Mao, Dhruva Arekar, Tianjun Zhong, Jiani Huang, Xujie Si, Ziyang LiAbstract
Autoformalization (AF) requires more than compilation: formal statements must preserve the meaning of their informal sources, a property known as faithfulness. An emerging frontier is the automated evaluation of this property, yet most existing methods operate at the statement level and produce whole-item verdicts, failing to reveal how individual expressions correspond or which contextual interpretations support the judgment. This omission is consequential because AF is inherently theory-scale: target statements derive their meaning from surrounding mathematical material and unstated conventions. We formulate faithfulness evaluation as a context-aware align-then-label problem: identifying phrase-level correspondences between informal and formal statements and assigning a faithfulness label to each under explicitly specified contextual interpretations and assumptions. We also introduce FaithAlign, the first theory-scale faithfulness benchmark comprising 460 items and 4,982 annotated phrase-level links, alongside FaithAlign-Gap, a companion set of 146 items with informal context and assumptions withheld. Each FaithAlign item includes the target statement pair, relevant context, formal dependencies, and, where needed, assumptions specifying one given interpretation. Across 8 models, GPT-6 Astra achieves the highest defect F1 at 47.8% and identifies all defects in only 4.5% of items with two or more defects. On FaithAlign-Gap, Astra identifies only 28.2% of undecidable items as such. These results show the difficulty of recovering context-dependent correspondences, establishing FaithAlign as a testbed for fine-grained, context-aware faithfulness evaluation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18498
PDF
-
LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations
2026-10-13 14:38:55Mohan Li, Chengyu Yu, Francesco Sovrano, Marc Langheinrich, Martin GjoreskiAbstract
Large Language Model (LLM) alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of alignment generalization. Moreover, existing evidence is largely grounded in natural language variation already represented during pretraining, leaving unresolved whether alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of alignment generalization and identify an empirical pattern we term Alignment--utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32717
PDF
-
How Well Do LLMs Perform in Peer Review? Evidence from Finance and Accounting
2026-10-13 14:11:35Hai Lu, Yichen Luo, Luping YuAbstract
Referees increasingly use large language models (LLMs), while finance and accounting journals disagree about whether to allow it. We ask which part of a referee's job the models can do. GPT-4o, whose training data end before the 2024 conference season, referees the papers presented at the 2024 AFA and NBER finance meetings, and we compare its verdicts with publication in the top-three finance journals over the next two years. Papers the model scores one standard deviation higher are 44\% more likely to be published, net of author reputation. The result is not driven by look-ahead bias, as the model does not recall the outcomes. The information comes from its judgements of correctness and exposition. Its judgements of originality and significance carry none, and a referee guide that stresses importance does not help. The split between correctness and contribution holds for accounting papers and for a newer model, GPT-5. LLMs can check whether research is right, but not whether it matters. The technical check can go to the model, while the judgement of contribution stays with people.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18435
PDF
-
Agaram 2.5B: The First Tamil-First Language Model Trained from Scratch with Tool Calling. A Playbook for Under-Represented Languages
2026-10-13 14:00:05Ramchand Kumaresan, Madhusudhanan G, Kaushik KezendiraneAbstract
We trained Agaram 2.5B, a 2.5-billion-parameter Tamil-English language model, from scratch on about 50 billion tokens (58% Tamil, the rest English, math, code and parallel text) on one node of eight H100 GPUs in about six days, with a new tokenizer that, unlike our first model's, keeps line breaks. Supervised fine-tuning on 134,428 conversations, whose replies an open-weight teacher model wrote, and three short targeted rounds turned it into a chat model that answers in Tamil script, romanized Tamil (Tanglish) and English and can call tools: on 25 tool prompts whose schemas appear in training it picks the right tool 23 times and gives exactly the expected arguments 13 times. The released model translates better than Sarvam-1, a 2B Indic base model, in all six FLORES and IN22 directions (FLORES chrF 54.3 into Tamil and 55.2 into English, against 49.2 and 44.2), and nearly doubles its base model on native-Tamil IndicIFEval (19.3% from 10.0%). It is weak on knowledge: reading-comprehension and knowledge benchmarks sit near chance, and on 50 held-out everyday questions written by one native speaker, one LLM judge rated 46% of its replies usable (22-24 of 50 over three passes; 95% interval for 23 of 50, 33-60%), with nearly every failure a wrong fact. Our two further SFT experiments changed answer format but not the knowledge measures. We also describe our first model, Agaram 1.5B, whose failures shaped the choices here; how the judge's scores varied between passes on identical replies; and, item by item, what the targeted rounds trained against and where frontier-model text entered the data. The project ran on compute credits and our own funds, about USD 13,300 across both models; the weights are released (licence to be announced after legal review), with GGUF files for this http URL, at this https URL code at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.18420
PDF
-
Lightweight but Linguistically Fair? Quantisation Robustness across Bavarian, German, and English
2026-10-13 13:15:59Jophin John, Michael Hoffmann, Piroska Lendvai, Florent DufourAbstract
Post-training quantisation reduces the memory cost of large language models, but it remains unclear whether compression preserves language skills of regional varieties as reliably as their standard-language counterparts. We study this question for Bavarian, Standard German, and English using BAVGROUND, a multiple-choice benchmark, where 206 source questions are aligned across all three languages. Across 14 open-weight 7B--10B instruction models, we compare BF16 with 8-bit GPTQ and two 4-bit settings, GPTQ and AWQ. Eight-bit GPTQ remains close to BF16 and largely preserves the relative behaviour of the three languages, while 4-bit quantisation was found to be much less stable. The two 4-bit methods reduce average accuracy by about 1.4--1.5 percentage points and change 11.9--16.0% of BF16 decisions, with changes concentrated on examples where the BF16 model is already uncertain. This instability does not translate into a consistent Bavarian disadvantage. Although Bavarian, a low-resource dialect, has lower baseline accuracy than German and English, the language most affected at 4 bits varies across models and quantisation methods. Calibration further shapes these cross-language effects: the relative Bavarian--German effect changes with calibration language, while changing the calibration source has little average effect on GPTQ-4 but can affect AWQ-4 much more strongly. Free-form evaluation on five models further shows that 4-bit compression can degrade generated responses even in settings where multiple-choice accuracy changes only modestly, while 8-bit GPTQ remains comparatively stable. Overall, our results show that quantisation robustness for regional language varieties depends on the model, quantisation method, calibration data, and evaluation setting, and cannot be measured by aggregate multilingual accuracy alone.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18343
PDF
-
JevGPT: Thinking Fast and Slow in Autonomous Research
2026-10-13 12:29:13Yasheng Sun, Yanning Dai, Yimeng Chen, Nanbo Li, Firas Laakom, Yifan Xu, Piotr Pi\k{e}kos, Mateusz Ostaszewski, Wenyi Wang, Cheng Luo, Yuhui Wang, J\"urgen SchmidhuberAbstract
Autonomous research agents must acquire and reason over evidence within a limited budget. Yet relevant sources can remain uninspected or unused, leaving a gap between retrieval coverage and the evidence that informs research. The challenge is to align fast local judgments with changing evidence needs and feed their findings back into research decisions. We introduce JevGPT, a dual-system architecture for autonomous research that couples fast evidence processing with deliberative research planning. System Two sets evidence objectives and integrates findings, while System One locates support, assesses relevance, and guides further inspection. Independent source judgments run in parallel; selected original passages inform the next research step. Experiments on fixed-corpus question answering and document ranking show that JevGPT achieves leading task performance, with faster completion and lower cost than comparably effective research agents. Evidence analysis shows broader source inspection and greater coverage of supporting documents in the research loop, highlighting efficient evidence processing as a route to more capable autonomous research.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18272
PDF
-
TextualRL: Textual Reinforcement Learning for Context Management
2026-10-13 11:47:32Tianyu Huai, Jie Zhou, Bo Mao, Qin Chen, Liang He, Xipeng QiuAbstract
LLM agents can adapt to task feedback by revising the textual context that guides their behavior, without changing their model parameters. However, turning execution rewards into reusable context updates is challenging: individual trajectories offer limited comparative evidence about what to change, while task-specific fixes may overgeneralize when incorporated into a shared context. We introduce TextualRL, a gradient-free reinforcement learning framework that treats a group of rollouts from the same task under the same context, rather than an isolated trajectory, as the unit of policy critique. Groupwise Policy Critique contrasts successful and failed rollouts within a task; when outcomes are uniform, it draws on cross-task evidence to identify behaviors worth preserving or formulate repair hypotheses. Cross-Group Policy Update reviews the resulting proposals against evidence pooled across groups, including observations that yield no proposed edit, to refine their content and applicability. Reviewed proposals form a bounded textual patch, with adoption of the candidate context determined by held-out validation rewards. Together, these components establish a reward-driven policy optimization loop in which multiple attempts jointly inform reusable context updates. We evaluate TextualRL with five target agents on six benchmarks. Averaged over the six benchmarks, TextualRL improves test accuracy over the base model by 24.4 and 15.3 percentage points for GPT-5.5 and Qwen3.8-27B, respectively.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18216
PDF
-
CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions
2026-10-13 10:13:07Xinyang Liao, Lingyu Li, Kaicheng Shen, Yang Yao, Na Lu, Yan Teng, Xia HuAbstract
AI companions provide users with emotional support and ongoing companionship, but their sustained relational interactions also introduce distinct safety risks. Prior work largely targets general harms or specific companion platforms and risk categories, while relying on manually collected conversations, limiting scalable and user-grounded evaluation and mitigation. To address this, we introduce CompanionSafe, a unified evaluation-to-mitigation framework for AI companion safety. Its evaluation component, CompanionEval, provides scalable multi-turn evaluation by constructing 1,300 relationally grounded settings from diverse personas, matched companion characters, and dialogue scenarios across four turn tiers. Its mitigation component, CompanionGuard, converts benchmark-exposed failures into turn-localized supervision for companion-specific risk detection. Our results show that even strong LLMs and existing guard models remain vulnerable to companion-specific relational harms, while both closed- and open-source models exhibit increasing safety risks as interactions become longer. Together, CompanionEval and CompanionGuard form a practical evaluation-to-mitigation loop for AI companion safety, turning relational safety failures exposed during multi-turn interaction into supervision for companion-specific guardrails.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18095
PDF
-
LifeSearch: Benchmarking Deep Research Agents for Real-World Life and Productivity
2026-10-13 10:12:26Yibo Wang, Senjie Jin, Yansong Ning, Xinbing Liang, Haotian Luo, Jiazheng Zhang, Yuxuan He, Will Wu, ClarenceaiAbstract
Deep research (DR) agents address open-ended information needs through multi-step reasoning, long-horizon web search, and evidence synthesis. However, existing benchmarks primarily focus on academic research or synthetic tasks, leaving DR agents' capabilities in real-world life and professional productivity settings insufficiently evaluated. To bridge this gap, we introduce LifeSearch, a challenging benchmark comprising 203 Chinese and English tasks derived from real-user query seeds across 7 concrete life and productivity domains. With more than 4,500 hours of human annotation, we construct over 4,000 expert-written, fine-grained rubrics, covering both report-content criteria for coverage and accuracy and concrete criteria specifying values, units, dates, and URLs. We conduct extensive experiments on 12 advanced DR agents under a unified evaluation harness, systematically analyzing their performance patterns and failure modes. We find that even the best-performing frontier agent, GPT-5.6 Sol, achieves a score of only 63.4%, leaving substantial room for improvement on realistic life and professional DR tasks. We select the grading configuration by comparing candidate LLM judges against human rubric labels, and a human ranking study shows strong alignment between aggregate benchmark scores and human judgments (Spearman's $\rho = 0.89$, Kendall's $\tau = 0.81$). LifeSearch provides a rigorous evaluation framework and foundation for advancing DR agents toward real-world life and productivity.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18092
PDF
-
Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents
2026-10-13 09:41:12Bofan Xie, Chaokun Wang, Guanchen Ge, Songyao Wang, Fang Niu, Hang Zhang, Wei WangAbstract
Long-term memory has become a core component of large language model (LLM) agents. As agents accumulate personal information over long-term use, users increasingly ask them to delete sensitive or outdated facts. A memory system should honor such requests. It should remove exactly what was requested and keep the rest of its memory intact. However, existing memory benchmarks evaluate only memory writing, retrieval, and organization, so whether agents can forget at the user's explicit request remains under-evaluated. To address this gap, we propose Bemula, the first benchmark dedicated to memory unlearning. It comprises 50 multi-session dialogue instances, 135 forget requests of six types, and 1,742 test questions. Each dialogue has a control version without the requests, so that only systems that have stored the target can earn credit for forgetting. Across the paired versions, five test suites measure storage, removal of the requested information, preservation of unrelated memories, relearning of similar facts, and resistance to extraction attacks. On Bemula, we evaluate six mainstream memory systems under three backbone LLMs, together with the Full Context LLM on two of them. The results are negative. Across all evaluated systems and backbones, no system passes more than 1.5% of forget requests end to end, and most pass none at all. These systems either leave the requested information extractable under simple attacks or delete far more than requested. Memory unlearning therefore remains an open challenge for agent memory design. We release our data and code at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.18052
PDF
-
Evaluating the Causal Faithfulness of SHAP as a Candidate Attribution Method for SAE Features
2026-10-13 07:54:39Samuel Xie, Noura Al-MoubayedAbstract
Sparse autoencoders decompose language model representations into interpretable features, but feature interpretation in this literature remains predominantly correlational. Attribution methods offer more direct estimates of feature influence, yet their faithfulness to actual causal effects has not been systematically evaluated in the SAE setting. We apply SHAP-based attribution to SAE latent features and evaluate it against direct causal intervention, using sentiment classification on SST-2 with GPT-2 Small and a 32,768-feature layer-7 SAE as a testbed. We compare SHAP to Integrated Gradients, Gradient$\times$Input, and probe-weight baselines across linear and MLP probe arms, scoring each against feature ablation and activation steering measured through the language model's own logit difference. Most faithfulness associations we measure do not survive multiple-comparison correction, and attribution methods that agree strongly with one another ($\rho = 0.42$--$0.80$) agree poorly with measured causal effects, indicating that inter-method agreement is not evidence of causal faithfulness. Two effects hold under control. First, local per-example attribution retains a directional advantage over global averaged attribution, while its apparent importance advantage is explained by feature activity rather than superior causal ranking. Second, under known-ground-truth minimal pairs, attribution recovers inserted sentiment signal well above chance (AUROC $0.792$) but fails to localize deletions (AUROC $0.520$), suggesting that these methods detect the presence of a feature's influence more reliably than its absence. We argue that importance faithfulness and directional faithfulness are separable properties requiring separate evaluation, and that attribution agreement should not substitute for causal validation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17898
PDF
-
Same Attack, Different Judge: A First Controlled Adversarial Comparison of Jev, LLM Judges, and Purpose-Built Safety Classifiers
2026-10-13 06:43:29Xiang Li, Aichun Huang, Mingming Zhang, Zheli LiuAbstract
TypeSafe AI's Jev is a new kind of safety judge. It generates no text and no label: it answers questions the developer defines in advance and returns a probability for each answer. The vendor admits that adversarial input can change these answers but gives no numbers. In this paper, we present the first controlled adversarial evaluation of Jev. We build a test harness that sends every system the same payload under the same policy, and run it on four public-data tasks, from rule-based ones (input moderation, injection detection, routing) to a judgment-based one (phishing detection), with fifty attack variants from thirteen families. The targets are Jev, five GPT and GLM judges, and two local classifiers (Llama Guard 3, Prompt Guard 2). We find that Jev is the most robust system on input moderation: eight forgeries of its structured input bypass none of its decisions, and a sentence asking only why someone might make a harmful request bypasses 9% of its blocks, against 56% for GPT-6 Astra. On phishing detection. However, the same forgeries bypass 19-29% of the emails it flags, and even without attack Jev detects only 62% of phishing emails, against 77-81% for the GLM judges. Jev is also the fastest judge, but it flags the most benign requests (27%). The local classifiers fail differently: a training-example statement makes Llama Guard 3 miss 62% of the phishing it had flagged. These results suggest that Jev resists attacks on tasks with a clear rule to check but is more exposed on tasks that call for an overall judgment.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17820
PDF
-
GIFT-Eval-CTX: A Benchmark For Contextual Forecasting Model Evaluation
2026-10-13 06:21:50Xu Liu, Hanshu Yan, Taha Aksu, Quang Pham, Zhongzheng Qiao, Jun Hao Liew, Silvio Savarese, Doyen Sahoo, Junnan LiAbstract
Contextual forecasting seeks to improve time series forecasting by incorporating forward-looking information, such as planned events, external conditions, and expert expectations. However, reliable evaluation remains difficult: existing benchmarks often have limited domain coverage, lack standardized history-future splits, or provide text that is not guaranteed to inform the future. We introduce GIFT-Eval-CTX, a benchmark that isolates whether forecasting methods can translate verified future information into accurate numerical trajectories. Built upon GIFT-Eval, GIFT-Eval-CTX contains 10,000 instances across 20 datasets. Each instance provides structured effects derived from the realized future: Macro effects describe horizon-wide behavior, while Micro effects specify localized changes. We construct these annotations with an agentic pipeline combining LLM-based discovery and deterministic code validation. We benchmark TSFMs, frontier LLMs, and tool-using agentic systems. The results show that structured effects reduce aggregate relative MASE by 40.7% for GPT-5.6 Sol and 29.6% for Claude Opus 4.8 relative to history-only prompting. The evaluated MoiraiAgent system achieves the best overall result. We further analyze the contributions of structured effects and agentic systems to performance, and identify opportunities for future research.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17794
PDF
-
Instructions Are Evidence, Not Objectives: Measuring Outcome Alignment of Agents
2026-10-13 05:39:34Kaiming Liu, Yuxuan Liu, Jinrui Ju, Minyuan Ruan, Xuanyu Lei, Peng Li, Yang LiuAbstract
Large language model agents can follow complex human instructions to solve real-world tasks. In practice, however, initial instructions are often incomplete. In our analysis of 100 user-agent trajectories from SWE-chat, 56\% of initial instructions omit requirements that later emerge as necessary, and 52\% of the useful follow-up requirements could have been inferred from the project. A user states requirements from their own understanding of the task, and these \textbf{stated requirements} often differ from the \textbf{true requirements} that completing the task imposes. An agent should therefore treat the instruction as evidence rather than as the objective, inferring the true requirements from the instruction and the environment and satisfying them even when the user is not yet aware of all of them. We call this ability \textbf{outcome alignment}. In this work, we construct \textbf{TiGanEval}, a benchmark of software engineering and research tasks. In every task the user genuinely lacks part of the requirements, each requirement stays discoverable from the code and runtime behaviour of the environment, and requirements the agent infers are separated from those the user adds after observing delivered work. These properties together make outcome alignment measurable. Evaluating eight frontier models including GPT-5.6-Sol, we find outcome alignment broadly missing: the models lose 53 to 67 points of success rate when moving from the full specification to the partial instruction, and consequence feedback across four submissions does not close the gap. Furthermore, a plan-mode harness cannot compensate for the missing ability of the model itself, which places outcome alignment among the abilities the next generation of models should be developed for.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17757
PDF
-
Auditing Failure Attribution: Label Predictability and Judge Behavior
2026-10-13 05:26:33Tianyang Xu, Jing GaoAbstract
Failure-attribution benchmarks ask an LLM judge which step or agent caused a failure. A high score should mean the judge checked the faulty step against the run (a cross-step check). However, spurious signals (surface cues) that predict the annotation can produce a high score. In this paper, we propose four tests to separate checking from cue use. Input restriction asks whether partial input predicts the annotation. Cue following checks whether moved cues change answers. Residual accuracy scores judges where cue-only predictors fail. Aggregation asks whether the failure mix or metric changes rankings. Real observations motivate these tests: on three benchmarks, simple predictors match or beat GPT-4.1. In two Who&When Pro systems, a classifier reading one step's reasoning finds 95-100% of faulty steps, motivating input restriction. In one of them, error injection leaves no reasoning after the faulty step; judges fine-tuned on its labels beat GPT-4.1 by 41 points. Adding task-related reasoning after that step moves this cue: fine-tuned judges lose 31-99 points, but zero-shot judges lose at most 9, motivating cue following. On Aegis, blaming every agent matches GPT-4.1 on micro-F1 but trails by 19 points in exact-set accuracy, motivating aggregation. We recommend reporting these tests when developing failure-attribution benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17747
PDF
-
Evidence-First Agentic Verification
2026-10-13 04:58:04Reza Sanayei, Razvan-Gabriel Dumitru, Sunjie Hou, MohammadHossein Rezaei, Haris Riaz, Robert Vacareanu, Minglai Yang, Ying Liu, Mihai Surdeanu, Daniel Yue ZhangAbstract
Long-horizon, semi-verifiable agentic tasks often rely on rubric-based evaluation, but the evidence needed to assess these tasks can be spread across the environment state and the execution trace. A judge may overlook this evidence or accept unsupported claims, including an agent's own reports of success. When rubric verdicts serve as reinforcement-learning rewards, such errors can reinforce undesirable behavior. We introduce Sauron, an evidence-first agentic verifier that investigates both state and trace and relies on a claim only after independently validating it against its source. Bounded queries and joint investigation of related criteria reduce context use and redundant work, and each verdict carries an auditable receipt of its supporting evidence. Across 2,586 human-labeled criteria from 100 personal-assistant episodes, Sauron achieves the highest F1 of five verifiers on all eight backbones tested while costing less than Gandalf, the strongest baseline. With GPT-5.6 Sol, it reaches 0.96 F1 versus Gandalf's 0.92 at 23% lower cost, with a wider margin on process criteria. On a paired fault-injection suite covering 13 families, several of which simulate reward hacking, Sauron detects 97% of injected failures while flagging only 2% of clean controls, outperforming baselines even when they are given the full trace. On TRAIL, a benchmark for detecting and localizing real agent errors, Sauron leads both Gandalf and TRAIL's own judge on all three quality metrics.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17724
PDF
-
Cowork-Gym: Synthesizing Interactive Work Environments at Scale
2026-10-13 03:56:58Anni Zou, Cheng Chen, Jiaji Deng, Weikang Zhou, Xinji Mai, Yue Wu, Jianing Lin, Lipeng Xie, Lixiong Qin, Qingyue Zhao, Que Shen, Shuchang Tao, Xuchen Pan, Yaliang Li, Yunpeng Zhai, Yushuo Chen, Zhihai Wang, Zhuo Zhang, Zhaoyang Liu, Bolin Ding, Jingren ZhouAbstract
Cowork agents navigate dependencies across tools and files while adapting to changing user requirements and edits to shared workspaces. Scaling such environments requires maintaining consistency among task objectives, executable resources, user behavior, and success criteria. We introduce Cowork-Gym, a framework that synthesizes Cowork environments through a build-execute-revise loop that uses execution evidence and review feedback to validate and repair generated environments. To capture realistic Cowork dynamics, a stateful user simulator responds to the agent's actions through dialogue and direct edits to the shared workspace. We synthesize 64.9K environments across 14 application scenarios and use them to post-train 9B, 27B, and 35B-A3B Qwen models with supervised fine-tuning, reinforcement learning, and on-policy distillation. Averaged across six agentic benchmarks, these models outperform their corresponding base checkpoints by 8.7, 6.4, and 4.1 points, respectively. Ours-27B ranks second overall among the evaluated models and achieves the highest average scores on PawBench and ClawEval, outperforming models with over 100B total parameters on both benchmarks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17670
PDF
-
RAN-JEPA: An Action-Conditioned World Model for Energy-efficient Mobile Network Control
2026-10-13 02:05:04Binghan Wu, Kai Zeng, Kai Yang, Ya-Qin Zhang, Joseph Sifakis, Ye OuyangAbstract
Operational mobile networks are rich in observations but poor in interventions: they continuously generate large volumes of passive telemetry, while controlled action--response trajectories remain comparatively scarce and costly to collect. This observation--intervention asymmetry is a central obstacle to building practical world models for RAN control. We address this problem with RAN-JEPA, an action-conditioned world model that separates representation learning from intervention learning. The model first learns a predictive latent representation from observation-only network data, adapts that representation to the target deployment domain, and then learns action-conditioned latent dynamics from controlled sleep/wake trajectories while keeping the encoder fixed. We evaluate RAN-JEPA using public operator performance data for pretraining and a real-equipment four-cell 5G testbed for domain adaptation, post-training, and control. Across four future 15-minute intervals, RAN-JEPA achieves lower action-conditioned prediction error than a GPT-2 world-model baseline and a mechanistic rule, while remaining close to a strong persistence reference. What-if evaluation further shows that the model produces distinct and coherent future trajectories under matched SLEEP and KEEP plans. In closed-loop control, RAN-JEPA consistently supports higher sleep ratios than the two baselines at comparable congestion cost. These results show that observation-first, action-conditioned learning can turn abundant passive telemetry and limited intervention data into a practical predictive model for energy-aware RAN control.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17584
PDF
-
Reinforcement Learning In Partially Verifiable Environments for Colon Cancer Treatment Planning
2026-10-12 21:39:21Li-Ching Chen, Andr\'es Franco Valiente, Yuta Ishiyama, Zachary R. Chalmers, William Chen, Kelsey Natsuhara, Taema Brinjikji, Madhumita Sushil, Travis Zack, Julian C. HongAbstract
Clinical decision-making is a sequential process of information collection, clinical inference, and forecasting under partial observability, often without a uniquely correct treatment decision. Large language models (LLMs) have demonstrated strong medical question-answering capabilities but extending these advances to clinical reasoning remains challenging because many clinical tasks are only partially verifiable. Recent clinical reinforcement learning (RL) work has therefore largely focused on diagnostic endpoints, where correctness can be more directly verified. Oncology treatment decision making is a particularly challenging instance of this problem that extends beyond medical question-answering in which multiple decision trajectories can be clinically defensible for the same partially observed patient state. We formulate oncology treatment planning as an interactive RL problem under partial observability and construct patient-specific, process-aware rewards from decision processes observed in real world clinical records. The agent must learn to sequentially acquire information from a structured patient representation before committing to a first-line treatment recommendation, and is evaluated on information gathering, treatment selection, and reasoning alignment. Using GRPO and GDPO to optimize Llama-3.1-8B-Instruct within this environment, RL substantially improves the base policy beyond frontier model capabilities. The GDPO-trained policy increases Mean@3 scalar clinical reward by $79.60\%$ over base Llama and exceeds GPT-5.2 by $18.82\%$ under the same interactive environment, while reducing query count during training. Overall, our results show that patient-specific, partially verifiable rewards can provide an effective learning signal for interactive clinical RL when both the patient state and the underlying clinical decision process are only partially observed.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17426
PDF
