Chat
-
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
-
Beyond Successful Play: Diagnosing Opponent Modeling in Large Language Models
2026-10-12 19:08:43Hanbo Xie, Yue Zhao, Robert C. WilsonAbstract
Language-model agents are expected to adapt to others whose behavior changes through interaction. Yet adaptation may rely on learned behavioral patterns rather than theory of mind (ToM): the ability to infer others' mental states, including beliefs and intentions. We introduce a benchmark and diagnostic framework for repeated hide-and-seek against algorithmic opponents with explicitly defined recursive belief updates. Their ground-truth beliefs and action probabilities let us distinguish adaptive play, action prediction, and belief tracking through live interaction, fixed-history evaluation, and controlled information interventions. With structured elicitation and behavioral supervised fine-tuning (SFT), we find improved play in several tested settings, but diagnostics reveal limitations in opponent prediction and belief tracking. We test whether supervising opponent action probabilities can address this predictive deficit in Qwen2.5-7B-Instruct. Predictions improve when the model is told the opponent's type and parameter settings, but become less accurate when it must infer the opponent from history without this information. Training on examples that also omit these details reduces the gap, although improvements vary across opponents. Is this limitation specific to smaller open-weight models, or does it also appear in a frontier model? GPT-5.4 also benefits from correct opponent beliefs: supplying them improves its predictions and choices, whereas deeper explicit belief reports and self-generated memory do not consistently improve forecasting from interaction history. This contrast suggests that using useful belief information is easier than inferring it autonomously, even for a frontier model. Together, our findings distinguish improved play and action prediction from reliable belief tracking, providing a framework for evaluating mentalizing and testing targeted improvements.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17310
PDF
-
Reliable Text-to-SQL in Resource-Constrained Settings with Small LLMs via Iterative Retry Workflows
2026-10-12 18:59:25Bilel Abderrahmane Benziane, Xavier ScheplerAbstract
Large Language Models (LLMs) have recently improved Text-to-SQL generation, but their computational cost can hinder deployment in resource-constrained settings. This paper evaluates the Text-to-SQL capabilities of small and medium open-source LLMs (1B-8B parameters) under a controlled inference budget, focusing on how model specialization and prompt formulation affect performance. Using the Spider benchmark, we compare multiple prompt representations and assess the accuracy of generated queries. Our results show that SQL-specialized models are consistently more reliable than general-purpose and chat-oriented models at small scales, and that performance can vary substantially depending on the prompt used. To reduce this variability without resorting to larger models, we introduce three lightweight inference workflows that automatically retry with alternative prompts and/or models when failures occur, including an optional syntax-correction step. These workflows significantly improve end-to-end success rates, with a trade-off in additional inference calls. These results demonstrate a practical path to stronger Text-to-SQL performance with small LLMs, by trading modest additional inference cost for higher reliability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17305
PDF
-
The Agent Knew: Internal States Forecast Unauthorized Actions and Stop Them, and the Lies That Follow
2026-10-12 18:53:33Arya Fayyazi, Mehdi Kamal, Massoud PedramAbstract
Language-model agents sometimes act beyond their authorization and then misreport what they did. On CAPE-Bench, matched agent tasks in which one fact decides whether an action is authorized, frontier open-weight agents do both, and one behavior marks the lie: retroactive permission laundering, acting and then asking permission as if the action were pending, precedes a false report in 70% of 1081 fresh informed violations, against 8% otherwise (preregistered odds ratio 7.0); frontier agents on tau^2-bench instead fabricate the authorization before violating. A linear readout of the agent's residual stream forecasts the violation before the action (AUROC 0.94) and the false report before the report, and beats chain-of-thought monitors when the agent is pressured to hide its reasoning. Gating every proposed action on this forecast cuts violations over whole episodes in four model families, and sending repeated flags to an approver keeps authorized work (Gemma-4: violations 0.70 -> 0.09, completion 1.00; a model approver that escalates unclear cases keeps this gain with a person seeing 25% of asks), at about 10-8 of the agent's compute and 109 times less than a chain-of-thought monitor. On AgentDojo, an external prompt-injection benchmark, the same readout calibrated on 29 tasks, with the steering unchanged, cuts attack success on unseen tasks in two families (Gemma-4 11.7% -> 5.7%, gpt-oss 28.5% -> 17.5%) while keeping benign utility (Gemma-4 0.637 vs 0.637; gpt-oss 0.667 vs 0.681). Preventing the violation is also the most reliable way to prevent the lie (misreported violations 0.10 -> 0.01 pooled over three families, preregistered), whereas lie-specific fixes after a violation are weak. All tests are preregistered and hash-locked; code, data and the full record, including tests that did not pass, are at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17301
PDF
-
Does a Governed Confidence Score Know When a Text-to-SQL Answer Is Right? An Exploratory Study on BIRD and a Pre-registered Confirmation on Spider
Abstract
Conversational analytics lets people ask questions of records spread across many systems and get answers in plain language; it is useful only if each answer can be trusted and explained. A production system of this kind attaches a confidence score to every answer it writes over a database and uses it to show the answer, show it with a warning, or withhold it. The score's weights, caps and thresholds were set by design. This paper audits that design. On the 1,534 questions of BIRD dev, answered by GPT-5.1 with a benchmark-adapted production prompt and five samples per question, the designed score separates right from wrong answers weakly (AUROC 0.624, 95% CI 0.595-0.652) and withholds 35.7% of correct answers. Nearly all of its ranking signal comes from agreement among the samples, and most of it is lost where agreement is combined by a minimum with a checklist that drops to zero on any unsupported claim, including claims in commentary. We froze a rule in which agreement with the shown answer is the score and deterministic checks act as vetoes and caps, and pre-registered a test on Spider dev before any Spider answer existed. The primary hypothesis held: AUROC rose from 0.566 to 0.725 (paired difference +0.159, 95% CI +0.097 to +0.221), and correct answers withheld fell from 49.0% to 8.8% with PASS accuracy 95.7%. Under one matching policy the choice of agreement statistic matters little; the gain comes from how agreement enters the score. Whether added checks help depends on how correctness is defined, while they always cost unqualified coverage. A pre-registered empty-result check did not meet its effect-size criterion on Spider. Correctness here means the core answer; labels are model-judged, and a blinded two-person audit, not yet complete, will be reported in a revised version.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17256
PDF
-
D2: High-Capability Full-Duplex Spoken Language Models from Half-Duplex Backbones
2026-10-12 17:15:29Cheng-Kuang Chang, Alexander H. Liu, Brian Kingsbury, Hilde Kuehne, Rogerio Feris, Samuel Thomas, James Glass, Kai-Wei ChangAbstract
Full-duplex spoken language models (FD-SLMs) support simultaneous listening and speaking, but often lag behind half-duplex spoken language models (HD-SLMs) in general capabilities. We propose Dialogue-to-Duplex (D2), a framework that converts pretrained HD-SLMs into high-capability FD-SLMs at fine temporal granularity. D2 adapts the original audio encoder to streaming through decoder-free representation distillation and learns full-duplex interaction through supervised fine-tuning. We systematically study the effects of chunk size on full-duplex interaction, general capabilities, and inference computation. We apply D2 to two distinct HD-SLM architectures, Qwen3-Omni and LLaMA-Omni 2. At a 12.5Hz update rate, Qwen3-Omni-D2 achieves 99.2% turn-taking success with 521-ms latency, 99.5% interruption success, and a state-of-the-art VoiceBench score of 70.1.
Abstract (translated)
URL
https://arxiv.org/abs/2610.17124
PDF
-
WeGR: Unified Generative Retrieval for Industrial Search Systems in WeChat
2026-10-12 15:38:47Jiajun Han, Guoxin Ma, Felix, Liangsheng Xu, Ruiqi Yu, Yingdi Guo, Zongbing Wang, Xinhua Ye, Chenyan Hua, Xin Xie, Pengcheng Yang, Haijun WuAbstract
Generative retrieval (GR) offers a promising paradigm for industrial search by unifying candidate generation within an autoregressive framework. However, its effectiveness is limited by two key challenges. First, existing Semantic ID (SID) tokenizers struggle to incorporate collaborative signals while maintaining effective codebook utilization, limiting their ability to distinguish items and achieve high-quality retrieval. Second, existing reinforcement learning (RL) methods suffer from severe reward sparsity under exact SID matching, leaving many sampled trajectories without effective learning signals and hindering preference optimization. To address these challenges, we propose WeGR, a unified generative retrieval framework consisting of two complementary components. Specifically, the Collaborative and Collapse-free Tokenizer (CoCo-VAE) incorporates query-to-item and item-to-item collaborative signals while mitigating codebook collapse through codebook refresh and complementary regularization, enabling more informative and balanced SID representations. Building upon these SIDs, Dynamic Horizon Policy Optimization (DHPO) progressively expands both reward evaluation and policy optimization horizons, reducing learning difficulty while leveraging behavioral reward signals at different levels of the SID hierarchy, without relying on auxiliary reward models. Extensive offline experiments demonstrate the effectiveness of WeGR in improving retrieval and ranking performance. Moreover, a two-month online A/B test in WeChat E-commerce Search demonstrates relative improvements of 2.539% in orders, 1.508% in click-through rate, and 1.381% in conversion rate, while reducing online CPU and GPU usage by 25.5% and 26.1%, respectively. The system has been successfully deployed across multiple search scenarios in WeChat, serving billions of users, generating tens of millions of PVs daily.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16979
PDF
-
SarcasTraps: Effect of Sarcasm on Turkish Hate Speech Detection
2026-10-12 15:25:27Niyazi Ahmet Metin, Dilara Kek\"ull\"uo\u{g}luAbstract
Standard hate speech detection datasets often emphasize explicit harmful language, but directed harm can also be expressed indirectly, without clearly hateful wording. Sarcasm is one such form of indirect expression, as it can convey harmful intent through mock praise, exaggeration, humor, or apparent politeness. In this paper, we introduce SarcasTraps, a targeted external test set for analyzing the effect of sarcasm on Turkish hate speech detection. SarcasTraps contains 479 entries annotated along two dimensions: hate speech and sarcasm. We train classifiers using BERTurk and XLM-RoBERTa on two existing Turkish hate speech datasets and evaluate them on SarcasTraps, using it strictly as an external test set. We also evaluate GPT-4.1 and Llama-3.3-70B in a zero-shot setting on SarcasTraps. Results show that sarcasm substantially affects hate speech detection. Fine-tuned encoder models lose recall on sarcastic hate, while zero-shot LLMs maintain higher recall but suffer substantial precision drops on sarcastic examples. These findings show that both fine-tuned hate speech detectors and zero-shot LLMs struggle when sarcasm is involved, and SarcasTraps provides quantitative evidence of this challenge through a targeted Turkish test set. Due to the presence of sensitive and hateful language, SarcasTraps will be shared upon request for academic research purposes. Warning: This paper has examples from the dataset which contains offensive and hateful content towards groups of people.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16957
PDF
-
Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate
2026-10-12 15:20:01Harsh Kumar, George Eilender, Ashton AndersonAbstract
Descriptions of how people use AI are becoming training data for the next generation of models. We ask whether what a model reads about its users changes how it treats people. Using contrastive synthetic-document finetuning, we train Qwen3-32B on four matched corpora built from the same 1,625 real-world usage scenarios. The corpora differ only in the predominant motive they attribute to users of AI chatbots: helping others (prosocial), exploiting others (antisocial), legitimate self-advancement, or task completion. Finetuning shifts the models' stated beliefs about users, how true those beliefs look along an internal truth direction, and their broader judgments of how safe and good the world is. These social priors carry into behavior. In trust games, antisocial models expect less reciprocity and send less money than prosocial models. Their distrust is specific and targeted at people, and users of AI chatbots in particular. After an exploitative record, they send almost nothing to a person but most of their endowment to a random device with the same record. In a common-pool resource simulation, a single antisocial agent leads every group of otherwise sustainable agents to exhaust the resource, after privately planning to mislead them. Antisocial training also raises dark-triad scores and lowers alignment ratings. Belief shifts appear when antisocial documents are 1% of training tokens, persist when descriptions of assistant behavior are suppressed, and are not reproduced by training on discourse about misaligned AI. Capping Assistant-Axis drift and steering emotion-related directions leave the gaps largely intact. What models learn about the people they serve shapes how they subsequently treat them and others.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16946
PDF
-
UA-Bench: Benchmarking Large Audio-Language Models on Underwater Acoustic Recognition
2026-10-12 15:18:11Zhe Sun, Liang An, Jiayu Yao, Ronglai Qian, Jianing Cai, Ziwei Huang, Qing Fan, Zhiyuan Yang, Yujun Cai, Hongli Cao, Yiwei Wang, Xiaoyan WangAbstract
Large audio-language models (LALMs) are fast becoming a general interface to sound, yet how far their broad pretraining carries into specialized domains remains largely open. Vessel monitoring, maritime safety, and ocean sensing all depend on recognizing underwater signals, but to our knowledge LALMs have not been systematically evaluated on them. Annotated recordings are costly to collect, and a faithful synthetic alternative has to bring together many interacting factors: the structure of detection pulses, communication waveforms, and ship-radiated noise, and the sound-speed profile, geometry, and multipath propagation through which each reaches the receiver. We address this by proposing \textbf{UA-Bench}, to our knowledge the first large-scale benchmark for evaluating LALMs on underwater acoustic signal recognition. It comprises 375,051 audio records that couple parameterized sources with modeled propagation, so that every record can be traced to the conditions that produced it. The records span these three signal families, and each carries a three-level label hierarchy (13 leaf classes), a templated multi-turn dialogue, and source and channel metadata. We evaluate on a balanced 2,600-record subset with a gated coarse-to-fine protocol that locates each error at source routing, fine-grained classification, or output format, and add a paired silent-audio control to test whether the waveform is actually used. Nine open LALMs, evaluated without task-specific adaptation, reach only 2.0\%--13.1\% leaf accuracy, against a random reference of 7.7\%. The waveform helps source routing far more than leaf recognition (for one model, +29.5 versus +2.3 points over silence), and just four models gain reliably at the leaf level, by at most 4.7 points. UA-Bench thus gives a reference point for these models, and its released training partition supports follow-up adaptation work.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16942
PDF
-
WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search
2026-10-12 15:10:01Yingtao Ren, Aolin Li, Cheng Cheng, Huiyun Hu, Huasheng Liang, Baoxing Xie, Jun Yuan, Haijun Wu, Chenxing WangAbstract
Short-video search requires compact query rewriters that infer personalized intent from brief, ambiguous queries under strict latency constraints. The challenge is to decide when personalization is warranted and generate a reformulation that preserves the user's immediate information need. A rewriter optimized for fluency or engagement may still introduce unsupported details and redirect the search toward a different intent. Existing methods often underutilize users' actual reformulations and behavioral traces, relying instead on synthetic reasoning and indirect behavioral rewards that may produce more engaging candidates but cause intent drift. To build an effective rewriter that remains faithful to user intent, we propose WeFair, Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in WeChat short-video search. WeFair assigns real user behavioral signals to a progression from demonstration to preference to verification. We first mine actual human reformulations from search sessions, then apply multidimensional LLM screening and contextual verification to filter unsupported transitions without extensive manual annotation. During training, we begin with a supervised curriculum that progressively builds rewriting capabilities from large-scale human personalized query demonstrations, and then use direct preference optimization to align with users' subsequent consumption. In the final on-policy optimization stage, we construct a strict verification reward based on actual human reformulations to further boost performance with group relative policy optimization. Further analysis shows that auxiliary rewards can activate more rollout groups in GRPO but reduce intent fidelity. Online A/B testing in WeChat short-video search shows a 1.576% relative increase in effective clicked-query rate (VV>10s) and a 1.731% relative reduction in query reformulation rate.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16927
PDF
-
Observe, Share, or Predict? Disentangling Partner Information in Embodied Multi-Agent Collaboration
2026-10-12 14:49:31Adri\'an Szlatincs\'an, Filippo Ziliotto, Bruno Lepri, Luciano Serafini, Massimiliano Luca, Tommaso CampariAbstract
LLMs are used as high-level controllers for embodied agents, but it remains unclear whether they support collaborative reasoning, such as anticipating a collaborator's actions in real-time teamwork without dialogue. We ask if observing the partner's state, exchanging object knowledge, and predicting the partner's next action, shape task outcomes and coordination when agents can share percepts but never intentions, plans, or task-level messages, and whether the two agents occupy symmetric roles. Using the Habitat-based PARTNR benchmark, we introduce a Decoupled Partial Observability regime that enforces strictly individual perception, and we augment decentralized zero-shot ReAct planners with Other Agent's State (OAS) observation, object Knowledge Exchange (KE) between co-located agents, and Other Agent's Next Action (OANA) prediction, evaluated on the full 2 x 2 x 2 grid with four planners on 1,000 validation tasks. OAS observation improves task quality, reducing extraneous effort for every planner, whereas KE primarily improves execution efficiency, cutting simulation steps by up to 25% with little change in success. The two are largely redundant for strong planners but complementary for the weakest, for which KE without the OAS is the worst configuration. OANA prediction lowers the success rate in 15 of 16 model x context pairs while raising explicit partner mentions to 55-100%, and its accuracy gains lie mainly in the coarse action type, partly because agents recognize the action the partner is already performing. The robot retrieves objects sooner and completes its tasks earlier than the simulated human. Shared context reduces total task time by shortening the remaining duration that the human works alone. Also, coordination benefits arise from responding to shared context, especially the OAS, rather than explicit anticipation, yielding shallow, planner-dependent partner modeling.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16895
PDF
-
Before Attention Sinks Form: Hidden-State Geometry Reveals and Modulates Their Formation
Abstract
Attention sinks are usually studied after they have formed. We ask whether a measurable hidden-state change appears beforehand and whether changing that geometry alters what happens later. To remove network depth as a shortcut, we compare prompts in the same model and at the same layer. When sink timing varies across prompts, simple spectral summaries of the hidden states add information beyond current attention, with the clearest prompt-specific signal in a short window about two to three layers before sink formation. The predictive strength varies across cohorts, so we treat the spectrum as a readout of a pre-sink state change rather than a universal warning statistic. The signal is broader than one measured extreme activation. We then modify only the singular values of the post-layer hidden states after current attention has already been computed. These interventions change later sink behavior in Qwen2.5, GLM-Z1, and GPT-OSS relative to norm-matched random spectral perturbations. In Qwen2.5, the immediate suppressive effect is accompanied by a reduction in next-layer query--key preference for position~0, and the learned spectral directions remain stable when estimated from substantially less training data. Together, the results show that attention sinks are preceded by a short spectral reorganization of hidden states, and directly perturbing this geometry can causally alter subsequent sink formation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16745
PDF
-
VisExMEM: Visual Evidence-Based Explainable Multimodal Evaluation Metric
2026-10-12 09:07:37Subhadeep Roy, Christoph Leiter, Steffen EgerAbstract
Current image-text evaluation often relies on a global score, which gives limited insight into which semantic details are supported by the image. Explainable metrics provide rationales, but the visual evidence supporting those rationales can still be difficult to verify. We introduce VisExMEM, a reference-free evaluator that decomposes captions into atomic claims, localizes visual evidence, verifies each claim, and averages claim-level support into a global score. Explanations use the same evidence and decisions. We train Qwen3.5-based 4B and 9B models on 100K samples with multi-task SFT, counterfactual refinement, and evidence-aware GDPO. Across eight benchmarks, VisExMEM-9B achieves the highest descriptive composite among open-source evaluators in our comparison, exceeding VQAScore by 11.7 points while remaining competitive with GPT-5.6 Sol. We also introduce AGITA-Bench, containing 600 human-annotated image-text pairs and 8.7K atomic claims for global and claim-level evaluation. For untuned Qwen3.5 and GPT-5.6 Sol controls, decomposed prompting raises the same composite by 3.6-5.1 points relative to holistic prompting. Human evaluation shows that VisExMEM explanations outperform prior explainable metrics and approach GPT-5.6 Sol.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16441
PDF
-
Correct Choices, Failed Sequences: Choice2Seq Benchmarks Correctness Retention Across Output Interfaces
2026-10-12 05:11:37Tianyou Lai, Wentao Yue, Zhiying Deng, Qilei LiAbstract
A model may select the correct answer yet fail to construct it through a sequence of local decisions. We introduce Choice2Seq, a diagnostic benchmark that measures correctness retention across complete-answer selection and sequential construction in independent runs on the same problems and candidate sets. Sequence Composition Rate (SCR) measures the fraction of correctly selected items that are also answered correctly through local construction. Jev achieves 97.1% and 99.5% selection accuracy on arithmetic and lexical retrieval, yet retains only 33.0% and 2.1% of those correct choices under serial construction, respectively. Substantial gaps also emerge for StartLux and Qwen, including when Qwen emits characters directly. Restricting local options to candidate-supported characters improves construction, while explicit position indexing improves copying fidelity, even for incorrect selections. These interventions distinguish task correctness, plan fidelity, and output compliance. Matched dialogue-state evaluations reveal that local field construction largely preserves Jev's correct states but reduces Qwen's mean-turn accuracy, highlighting task- and interface-dependent differences. Sequence-level confidence and request-cost analyses further characterize completed outputs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.16191
PDF
