Review
-
Large Language Models, Machine-Assisted Cultural Analytics, and the Agentic Turn
Abstract
Technologies like large language models make it easier to convert cultural material into analyzable variables, while LLM-driven agentic systems extend machine assistance to other parts of the research workflow. Rather than a full how-to guide or an exhaustive review of recent LLM use cases, this chapter looks at some crucial aspects of this new paradigm as well as the bigger picture. I argue that these technologies expand the scale, methodological range, and scope for delegation, while the basic logic by which those measurements support cultural claims remains mostly unchanged.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12792
PDF
-
SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
2026-10-08 18:01:00Wei Minn, Zhensu Sun, Lwin Khin Shar, David LoAbstract
Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository's existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12569
PDF
-
Hybrid Cinematography: Previsualizing and Managing Hallucination Risk in Generative Video Reshooting
2026-10-08 17:59:11Nhan (Nathan), Tran, Neal Wadhwa, Abe Davis, Stefan StojanovAbstract
On a film set, the camera move is committed during a take. Generative video reshooting lets filmmakers change it afterward, but may require hallucinating unrecorded content, a gap sometimes discovered only after leaving the set. We present Hybrid Cinematography, a workflow that bridges physical capture and generative reshooting to manage hallucination risk while filmmakers can still act on it. Using an editable 3D shot plan and a proxy of the take, our previsualization evaluates hallucination risk in real time. Seeing where the take lacks support, filmmakers can iteratively adjust the plan, explore moves that balance capture and generation, shoot guided pickups, or knowingly accept hallucination. We demonstrate the workflow through a mobile augmented reality application for on-set planning, capture, and review, and an offline pipeline for existing video. A study with experienced filmmakers reveals how previsualizing risk informs camera decisions and exposes tensions between creative intent and generative hallucination.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12455
PDF
-
RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments
2026-10-08 17:55:12Zimo Wen, Yijin Chen, Yuxuan Cao, Wendi Chen, Yanwen Zou, Wenye Yu, Fuhang Kuang, Han Xue, Jun Lv, Chuan Wen, Cewu LuAbstract
A generalist robot should not only perform diverse tasks but also improve through experience, turning what it learns during execution into capabilities that later tasks can reuse. Robot agents that act through code can already repair programs from execution feedback, yet it remains a central challenge to organize this experience around the task structure that gives it meaning, so that each repair is attributed to the responsible capability, supported by execution evidence, and validated before it is reused. We introduce RoboRSI, a robot self-improvement system built on Top-Down Skill Refinement (TSR). TSR decomposes tasks into compound, atomic, and base skills with scoped responsibilities and explicit input--output contracts, attributes each execution outcome to the responsible branch, and confines revision to that branch. Building upon this structure, a Manager, Planner, Engineer, and Reviewer coordinate planning, execution, diagnosis, and the validated release of new skills, while people steer the process through objectives and corrections; stable skill sequences are further consolidated into reusable compound skills. On a mobile manipulator, RoboRSI develops multi-object household cleanup over 104 rounds. In simulation, it achieves the highest success rate on LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin, exceeding the strongest baseline by 2.7 to 11.0 percentage points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12424
PDF
-
Machine Learning Meets High-Energy Nuclear Physics: From Pattern Recognition to Physics-Integrated Discovery
2026-10-08 16:47:33Xun Chen, Weiyao Ke, Yu-Gang Ma, Long-Gang Pang, Kai ZhouAbstract
Machine learning (ML) in high-energy nuclear physics (HENP) is entering a new stage in which physical knowledge is incorporated more directly into data analysis, simulation, and physics inference. This mini-review focuses on developments that have matured in the past several years. Whereas earlier applications emphasized event classification, pattern recognition, and surrogate models for selected observables, recent work has moved toward physics-integrated workflows: calibrated Bayesian extraction of QCD matter properties, dense-matter equation-of-state inference from heavy-ion and neutron-star data, generative event modeling, neural unfolding of weak physical signals, differentiable inverse solvers, gauge-equivariant and diffusion-based lattice-field samplers, and neural reconstruction of model functions in holographic QCD. We survey recent applications of ML in heavy-ion collisions, neutron-star physics, lattice QFT, and holographic or continuum QCD. The emphasis is not on ML architectures alone, but on how they enter concrete physics workflows, how physical constraints such as symmetries, conservation laws, causality, thermodynamic stability, and topology are imposed, and how uncertainty quantification and validation determine whether an AI-assisted result can support a reliable physics conclusion.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12293
PDF
-
One Word Opens the Gate: The Option-Channel Attack on Typed Decision Models as Agent Guardrails
2026-10-08 16:46:43Seyedarmin Azizi, Erfan Baghaei Potraghloo, Massoud PedramAbstract
A typed decision model reads a piece of text and returns a probability over caller-defined options, each with a short written definition, generating no text. Recent work places these models in agent systems as guardrails: the component that reads a proposed tool call or incoming message and decides whether to allow it. We evaluate seven open-weight models in that role and report the two error directions separately: a fail-open error allows a prohibited action and is a vulnerability; a fail-closed error blocks a permitted one and is only a cost. On prompt-injection, jailbreak and toxic-content screening, accuracy at the allow-or-block decision ranges from 36% to 72% against a chance level of 50%. A low error rate in one direction only reflects which answer a model defaults to: one allows nearly everything, another blocks nearly everything. On a synthetic suite of agent tool calls, six lines of server log text that say nothing about the policy raise a gate's fail-open rate from 0% to 63% on a policy it otherwise decides correctly. Giving the permissive option a misleading name, with its definition and the judged text untouched, raises that rate to between 93% and 100% on the four models that place the label in their input. Every defense we tested is defeated, either by an attacker who targets its mechanism or by attacker-controlled text. Escalating the least confident decisions does not help either: a decision an attack has reversed is no less confident than the one it replaced. Parsing each policy field into a typed value does eliminate one attack, but it also makes the model unnecessary: a deterministic rule over those values reaches 100% accuracy on all six policies. These models can reduce how many cases reach a reviewer, but on this evidence they should not be the component that decides. Code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12292
PDF
-
OA-MAP: Evidence-Grounded Multi-Agent Multimodal Framework for Interpretable Knee Osteoarthritis Progression
2026-10-08 15:23:02Sixu Chen, Mingrui Yang, Qiang Guan, Xiaojuan LiAbstract
Knee osteoarthritis (KOA) progression prediction can support patient monitoring, requiring the integration of multimodal data and multidomain expertise. Moreover, isolated risk estimates provide limited insight underlying a prediction. To automate the progression assessment workflow and reduce manual effort while providing interpretable findings and supporting evidence, we present OA-MAP, an autonomous multi-agent framework for evidence-grounded assessment of structural and pain progression in KOA. The system incorporates modality-specific agents including MRI, X-ray, and clinical agents, together with a coordinator agent. This framework can autonomously recruit specialist agents, select tools for prediction and analysis, and retrieve literature as external evidence based on user request and available patient information. An uncertainty-informed human-in-the-loop mechanism enables clinicians to review and correct intermediate findings, triggering recomputation of affected results. We evaluate the prediction models using 600 participants from the FNIH Osteoarthritis Biomarkers Consortium cohort. On the test set of 100 participants, the fusion models achieve AUROCs of 0.80 for structural progression and 0.68 for pain progression. A case study illustrates how OA-MAP combines risk estimates with intermediate findings, cross-modal conflicts, literature support, and uncertainty indicators to support interactive review.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12134
PDF
-
A persistent accuracy ceiling in automated verbal deception detection
2026-10-08 15:15:32Riccardo Loconte, Jonas Festor, Zane Fatjanova, Mariam Bolkvadze, Bennett KleinbergAbstract
Automated methods have been proposed to overcome the limitations of human verbal deception detection, but evidence remains fragmented across disciplines. We systematically reviewed 25 years of research (289 reports, 6,136 classification models) and meta-analyzed 3,653 models nested within 97 datasets. Pooled accuracy was 74.4% (95% CI: 71.2%-77.4%) with substantial heterogeneity. Accuracy was driven by methodological quality (ground truth, data source, class balance, evaluation procedure) more than by model complexity: the adoption of embeddings and large language models has not translated into improved predictive performance. Only 12.46% of reports used data with verifiable ground-truth, and only 23.96% of models were evaluated on independent data. The pooled accuracy aligns with meta-analyses of manual approaches, suggesting a ceiling of 70-75%, unlikely to be lifted by current research conventions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12118
PDF
-
LIVIN: Benchmarking Spatial and Embodied Intelligence in Digital Twins of Lived-In Homes
2026-10-08 14:47:30Peijun Xu, Chuansen Nie, Yiyang He, Yinuo Bai, Jingyang Liu, Kuixiang Shao, Yuyang Jiao, Kuanhao Xia, Jiayi Zhu, Zitian Yang, Yanqi Zhang, Tianye Tan, Shuwei Di, Junyi Xu, Jingyi Yu, Jiayuan GuAbstract
Realistic household simulation must capture not only diverse environments but also the lived-in object arrangements and spatial constraints that shape robot motion and interaction. Existing resources often trade off scale, real-world correspondence, and interaction readiness, leaving a gap in faithful, interactive replicas of how real homes are actually arranged. To this end, we introduce LIVIN, a benchmark for spatial and embodied intelligence built on digital twins of 30 diverse lived-in homes. These replicas preserve observed room layouts, furniture configurations, and everyday belongings. To construct them, we design a human-in-the-loop workflow comprising instance recognition, architectural reconstruction, and object generation and placement, with intermediate results reviewed and corrected by humans against the source observations at each stage. We evaluate four tasks in LIVIN: 3D detection, 3D reconstruction, navigation, and loco-manipulation. Our evaluations show that current methods remain challenged by the dense object arrangements, occlusions, limited free space, and constrained interaction regions found in realistic lived-in homes. We hope LIVIN will help advance embodied AI in real-world homes, from spatial understanding to robotic interaction, and ultimately bring embodied intelligence into everyday home environments.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12069
PDF
-
Natural Language to First-Order Logic LLM-based Autoformalization
2026-10-08 14:24:12Andrea Brunello, Cristian Curaba, Luca Geatti, Michele Mignani, Angelo Montanari, Nicola SaccomannoAbstract
Large Language Models (LLMs) have renewed interest in autoformalization. Yet, when First-Order Logic (FOL) is considered as the target formalism, the field still lacks a unified task formulation and a systematic survey. This paper addresses this gap: we first provide a principled definition for the FOL-autoformalization task by distinguishing Ontology Extraction from Logical Translation, showing how their conflation obscures (cross-study) evaluation; we review existing datasets, evaluation metrics, and LLM-based methods, including fine-tuning, prompting, and verification-based refinement; we identify open challenges in benchmarking, semantic evaluation, ontology-aware methods, and end-to-end applications.
Abstract (translated)
URL
https://arxiv.org/abs/2610.12030
PDF
-
How Is Automated Research Evaluated? A Survey of Benchmarks and Evaluation Practices
2026-10-08 12:50:23Liulei Zhang, Dejing Zhou, Chuyue Huang, Guanhua Chen, Yutong Yao, Lidia S. Chao, Chi Man Vong, Derek F. WongAbstract
Automated research systems support literature synthesis, ideation, experiments, writing, and peer review, but their evaluation is dispersed across tasks, benchmarks, and studies that are difficult to compare directly. We review this literature from the perspective of evaluation design and evidence, covering six targets: literature synthesis, research ideation, executable workflows, scholarly writing and communication, automatic peer review, and end-to-end research. We compare task construction, evidence sources, evaluators, and scoring procedures to explain the capabilities assessed by different designs. Our synthesis highlights three recurring lessons: output checks, process checks, and human studies provide complementary information; evaluator calibration is specific to the property being assessed; and resource budgets and attempt selection are integral to interpreting performance comparisons. We identify diagnostic evaluation designs and documented gaps in supporting evidence, and translate these comparisons into reporting and audit recommendations for specific evaluation settings. The survey helps readers navigate existing evaluations, select appropriate benchmarks, and design subsequent studies.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11877
PDF
-
What Output-Only Review Cannot Verify: Study Contracts for Research Agents
2026-10-08 11:42:23Eitan Waks, Ben GlockerAbstract
Some defects in an AI-generated study can be identified from its artifacts; others require knowledge of what was approved before execution. We propose study contracts that bind declared experimental choices, run obligations and claim scope to recorded execution evidence, and distinguish this contract-relative verification from scientific truth. A diagnostic using eight self-authored clean/mutated pairs illustrates the information boundary. A deterministic checker applying a registered, fault-specific rule to approved and executed objects detected all eight registered mutations. Across eighteen recorded judge aliases given individual metadata-filtered packages without pair context or the registry-selected fault label, 104 of 144 mutated evaluation cases received defect flags; the remaining cases comprised 32 abstentions and eight terminal failures, with no explicit clean decisions on mutated cases. Some packages retained approval and execution fields, including digests. The prompt instructed judges to abstain when evidence was insufficient. These results characterize a deliberately information-asymmetric development setting; they do not isolate the effect of authoritative information from differences in task specification and rule selection, and they are not comparative verifier quality or agent reward hacking. We identify full-information comparisons, legitimate-adaptation controls and closed-loop agent evaluations as necessary tests of whether contract checks improve useful compliant completion under optimization.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11754
PDF
-
Opera: A Verbal Critic Framework for Long-horizon Coding Agents
2026-10-08 11:11:53Kai Mei, Zhiyuan Hu, Yutong Dai, Juntao Tan, Yifan Zhang, Dingjie Song, Dimitris N. Metaxas, Silvio Savarese, Ran Xu, Zeyuan ChenAbstract
Long-horizon coding agents need timely corrections, yet feedback can be ineffective or even harmful when it misjudges ongoing work or fails to address the underlying problem. Existing critics focus on evaluating trajectories and generating feedback, but rarely track what happens after feedback is delivered. We present Opera, a verbal critic framework that treats each correction as a persistent note, followed until the diagnosed problem is resolved. Opera decides when to review through periodic and event-driven triggers, diagnoses issues with typed operators, audits feedback against visible evidence before delivery, and tracks the agent's subsequent actions to distinguish mere compliance from actual resolution. As a test-time critic, Opera improves the resolve rate of non-critic agents by up to 12.4, 15.0, and 8.9 percentage points on Terminal-Bench 2.1, a SWE-Bench Pro subset, and DeepSWE v1.1, respectively, across four policy models, and achieves the highest mean resolve rate among competitive critic baselines on all three benchmarks, and also improves policy models when the policy critiques itself. Beyond inference, Opera-guided rollouts provide approximately on-policy training data: fine-tuning Qwen3.5-9B on them improves its resolve rate on held-out SWE-Bench Pro repositories by 10.2 percentage points without a critic at inference time, matching fine-tuning on rollouts from a stronger model, while preserving its performance when switching harness, i.e., from Openhands to Terminus-2, which the latter substantially degrades. Our code is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.33987
PDF
-
VESSI - VLM-Enhanced Support for Surveillance and Investigations
2026-10-08 10:50:17Saverio Cavasin, Pietro Tedeschi, Mattia Tamiazzo, Alessandro Brighente, Simone Milani, Mauro ContiAbstract
Automated video surveillance analysis has become a critical component of intelligence infrastructures and Law Enforcement agencies. Traditional systems lack the semantic module for comprehensive situational awareness and forensic tasks, limiting their ability to interpret events meaningfully or support post-incident investigations. This slows operational insight and increases the burden on human analysts. Recent advances in Vision-Language Models (VLMs) offer promising pathways to bridge this gap. To address this, we propose VLM-Enhanced Support for Surveillance and Investigations (VESSI), a VLM-based framework designed to enhance automated video surveillance analysis through prompt-driven interrogation of video sequences where salient visual features are converted into textual descriptions. We test our framework with four state-of-the-art models. Since most datasets for this task are unlabeled, we also propose the Composite Model Utility Score (CMUS) to assess VLM performance. Experimental results show that our solution substantially improves the analysis capabilities of human operators and enhances the flexibility of automated surveillance systems. In our evaluation, the most reliable model flagged potentially relevant activity in more than 66% of the videos while reducing review time by more than 85%, offering a practical balance between selectivity and efficiency. The model ordering produced by the reference-free CMUS evaluation was reproduced by the normal-video CMUS evaluation and matched the false-positive-rate ordering obtained from 5,909 manually referenced frames. This agreement supports the operational use of the score within the evaluated setting.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11674
PDF
-
Runnable Commit Untangling for Coding Agents
2026-10-08 09:42:27Jinfeng Jiang, Dongsun Kim, Dayi Lin, Zhou YangAbstract
Coding agents produce large, tangled patches that mix multiple development purposes, making the code hard to review and maintain. Commit untangling offers the promise of organizing such large patches into untangled, manageable commits. This paper emphasizes two important limitations in existing commit untangling studies. First, they do not consider that untangled commits are ordered and should leave the code runnable. In practice, maintainers are unlikely to accept commits that prevent the code from running. Second, existing studies claim that commit untangling helps software maintenance. However, they conduct syntactic comparisons between the untangled commits and developers' original commits without directly showing the claimed maintenance benefits. To address these gaps, this paper makes two novel contributions: (1) RucTangle, the first agentic method that untangles commits while keeping the code runnable after each commit; and (2) TangleEval, the first evaluation framework that quantifies how untangled, manageable commit histories help coding agents repair bugs. We compare RucTangle against four untangling methods on 131 agent-generated patches. All histories produced by RucTangle are runnable, while baselines produce 20.6%-37.4% unrunnable commit histories. We further collect 453 agent-generated patches that introduce regressions (i.e., causing previously passing tests to fail) and ask two other coding agents to repair regressions. Augmenting agent context with RucTangle-produced histories yields 5.2% absolute improvement in pass@1. We also analyze agent trajectories to learn how they use untangled commits to navigate and fix bugs. Our findings demonstrate the value of adopting established software engineering practices in the era of coding agents, which broaden the future research agenda: how can agents actively use software history to make better development decisions?
Abstract (translated)
URL
https://arxiv.org/abs/2610.11593
PDF
-
AgentTime: Can Agents Estimate and Control Their Own Runtime?
2026-10-08 09:06:36Michael Ofengenden, Maksym AndriushchenkoAbstract
An essential control of AI agents is their ability to manage runtime. This ability requires a sense of time-awareness, to predict and estimate wall-clock time and to control their own actions. Prior work has focused on time-awareness, but duration-following and control in native agent harnesses remain unexplored. We present AgentTime, a benchmark for testing whether agents can work for a requested duration, predict their runtime, and estimate elapsed time afterward. It comprises 222 tasks from 18 sources spanning coding, computer use, agentic work, and automated research. Duration-following experiments append a single instruction specifying how long to work, with requests ranging from about a minute to multiple days. Accuracy on these instructions varies substantially: Fable 5.1 in Claude Code deviates from requested runtimes by a typical factor of 2.9$\times$, compared with only 1.2$\times$ for GPT-6 Astra in Codex. However, matching the requested runtime does not, by itself, establish continued work on the task. Among 158 reviewed Astra runs with classifiable transcripts, 14 explicitly slept after appearing to finish. In forecasting experiments, predictions tend to overestimate natural runtimes. In retrospective experiments, removing temporal information more than doubles deviation for Sol and Astra and nearly doubles it for Fable. An agent's ability to complete a task does not guarantee that it can control its own time or work for the whole requested duration. For agents to run reliably, safely, and autonomously over long horizons, we require the evaluation of both.
Abstract (translated)
URL
https://arxiv.org/abs/2610.09944
PDF
-
Why machines will still not rule the world
Abstract
In our book Why machines will never rule the world [13, 14] we argue that arti- ficial general intelligence is mathematically impossible. This is because the human beings and the processes which exhibit intelligence are complex systems whose be- haviour cannot be captured by the kinds of models that we can generate with or without computers. Proponents of contemporary machine intelligence respond with two lines of argument: a theoretical one, grounded in the universal approximation theorems for neural networks and the Church-Turing-Deutsch principle; and an em- pirical one, grounded in rapidly rising scores on standardized benchmarks. In this communication we examine and reject both responses. First, we show serious issues in the physicalist counter-argument based on the Church-Turing-Deutsch principle. Second, we review recent evidence to the effect that prominent benchmarks are compromised by training-data contamination, flawed test construction, and strate- gic optimization. Our central argument remains: That models required to perform cognitive behaviour in open-ended, thermodynamically complex and non-ergodic environments are not and will not become achievable.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11424
PDF
-
Writing for the Reviewer: Defensive Writing in GPT Models
Abstract
Researchers increasingly use ChatGPT to revise their papers, and recent GPT versions often narrow or even retract the authors' claims. We call such changes defensive writing when the given material does not support them, and we test two explanations: the model corrects the authors' overclaiming, or it writes for an anticipated reviewer. We ask GPT versions and models from other developers to rewrite paragraphs from papers written before ChatGPT, or to write from an evidence sheet that lists a paper's method and results. Defensive writing grows with GPT version. GPT-6-astra retracts the authors' claims outright, and when it writes from the evidence sheet, it still adds the most ungrounded qualifications. The results favor the anticipated-review explanation, and correcting overclaiming explains only a small part. When the models are only asked to polish, defense stays near the level of the originals; mentioning review raises it, and one round of self-review raises it further. At the same time, fewer than one in ten of the claims GPT-6-astra retracts are overstated. AI reviewers score defensive rewrites higher, while human readers find them harder to read and the authors less certain. Combining AI writing with AI review may amplify this style.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11355
PDF
-
Deception by Omission: Language Models Knowingly Hide Their Mistakes
2026-10-08 06:45:02Lucas Florin, Amelie Knecht, Ulysse Schaller, Thilo HagendorffAbstract
Large language models (LLMs) increasingly act as agents with little human oversight, so potential mistakes they make can go unnoticed. Users then depend on the model to report what went wrong. An honest model discloses its mistakes, while a deceptive one conceals them. However, it is unclear how current LLMs behave in such situations. In this study, we prefill LLM trajectories with synthetic mistakes. The trajectories resemble real deployments in chat and agentic settings. Models fail to disclose their mistake in 36.4% of chat and 67.1% of agentic rollouts. In 2.4% and 5.3% of rollouts, respectively, they are aware of the mistake in their chain of thought but still deceptively conceal it. Rates vary by model: for instance, Gemini 3.5 Flash knowingly conceals mistakes in up to 19.9% of agentic rollouts. In 11.9% of chat and 51.8% of agentic rollouts, models show no awareness of mistakes, even though they reliably spot them when reviewing the same transcript as an outside observer. Our results show that, as agents take on more tasks with less oversight, users cannot rely on them to self-report possible mistakes. Developers should instead use independent monitors that review agent trajectories, or specifically train models to check their past actions and disclose what they find.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11351
PDF
-
How to post-train on a surrogate: Envelope sampling mitigates reward hacking
2026-10-08 05:48:31Sanjit Dandapanthula, Shuvom Sadhuka, Samir Khan, Michael Oberst, Aaditya Ramdas, Alexandra ChouldechovaAbstract
Large language models (LLMs) are commonly post-trained against LLM judges and other cheap surrogates because the true reward, such as human preference, is too expensive to query at scale. This practice often leads to reward hacking, where reinforcement learning against a miscalibrated surrogate leads to undesirable side effects. In this work, we study a setting in which a small number $n$ of model outputs are annotated with ground-truth labels (e.g., from expert review) and used to recalibrate the LLM judge before optimizing against it. Prior approaches to judge recalibration are costly or heuristic, and it is known that on-policy sampling fails when the surrogate is miscalibrated on a rare set of outputs. In this work, we propose envelope sampling, a theoretically-grounded method for judge recalibration that seeks to minimize an upper bound on the regret of the post-trained model under the assumption that the human reward and re-calibrated reward lie in an $L^2$ ball around the judge. We give practical algorithms to sample from the envelope by rejection or by fine-tuning against a modified reward, and experiments on clinical note generation and on a controlled sycophancy task show that recalibrating on envelope samples mitigates reward hacking where recalibrating on base-model samples does not.
Abstract (translated)
URL
https://arxiv.org/abs/2610.11281
PDF
