Chat
-
Adaptive Consistency Graph for Long-Horizon Agents
2026-09-29 06:28:56Jiecong Wang, Hao Peng, Zhanyi WangAbstract
Large language model agents can often make reasonable local decisions on short tasks, yet their performance degrades when success requires long sequences of dependent actions and tool calls. During execution, task requirements, historical evidence, and the current execution state may gradually become disconnected, so later decisions can drift from the original objective. We study this problem by introducing the Adaptive Consistency Graph (ACG) for long-horizon execution. ACG incrementally organizes execution evidence and its provenance in a persistent graph, then constructs a temporary requirement-centered view for each decision under a bounded context budget. Rather than replacing the base agent's planner or tool executor, ACG provides a structured and traceable context view for each decision. In the matched evaluation, ACG improves GPT-5.6-luna's average success from 44.5\% with ReAct to 50.2\%, with the largest gain on BrowseComp-Plus (73.5\% versus 62.4\%). We further analyze trajectory structure and inference cost to characterize this improvement. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.32754
PDF
-
Spotter: Let the Embodied Model Lead, and the VLM Reflect for It
2026-09-29 06:21:50Long Li, Qichao Zhao, Yue Yang, Fan Xu, Zhe Wang, Alan Wee-Chung Liew, Chao Qu, Heng Tao Shen, Shirui PanAbstract
Current embodied models do not respond to their own failures, although what just went wrong could inform a small adjustment on the next attempt, the kind of reflection behind the gains of thinking in language models. We test whether they can repair a known error, which requires producing a correction and judging whether it is right. Stopped at a failure and allowed to retry, they seldom repair it through their own randomness or from a language description of the error, and best-of-N selection cannot pick the successful candidate after a failure. We attribute this to training only on successful demonstrations and to inputs too narrow to show what went wrong, and conclude that reflection must come from a vision-language model (VLM), which takes in far more information, such as the episode history and text, and is more general. Prior VLM-led work has the VLM plan every step and invoke the embodied model as a tool, placing the VLM on the critical path. We propose Spotter, which reverses the roles: the embodied model leads and executes continuously, while the VLM runs in parallel, monitors through a lightweight local screener, intervenes only when an error is detected, reflects on and corrects it, and returns control. We run Spotter with Qwen and with GPT as the VLM, and both improve the embodied models; with GPT, Spotter improves Cosmos Policy and $\pi_{0.5}$ by 5.6 and 7.5 percentage points on RoboCasa, and raises $\pi_{0.5}$ from 47.2% to 57.0% on the Hard setting of RoboTwin 2.0 and from 53% to 83% on a real robot. Because the VLM steps in only when an error is confirmed, a successful episode with Qwen takes only 13 to 16 s longer than with the embodied model alone and about 70% less time than with a VLM-led baseline using the same model. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36808
PDF
-
Code4Scene: Benchmarking Coding Agents for Constructing and Editing 3D Scenes
2026-09-29 05:49:37Xiaokang Ye, Siddhant Hitesh Mantri, Zimeng Chen, Edward Zhang, Zhaoxu Zheng, Yuanheng Li, Yizhao Chen, Tianyang Huang, Lianhui QinAbstract
Frontier coding agents can now write and execute code that authors 3D environments, but whether they reliably understand 3D structure and precisely control scene state remains unclear. The generated 3D scene is a persistent, executable artifact: a convincing render can hide incorrect spatial relations, intersecting objects, or unintended modifications. We introduce Code4Scene, a benchmark of 190 Unreal Engine cases built from human-assembled scenes that evaluates coding agents on two complementary settings under a shared execution interface. Construction tests scene-level spatial reasoning from open-ended language specifications, where many realizations are valid; editing tests precise control of scene state, where the agent must recover the target scene from reference images while preserving everything else. Rather than scoring code or rendered views, Code4Scene evaluates the generated engine-native scene for task fulfillment, artifact integrity, and static physical validity, with edits additionally compared against withheld ground truth. Across 14 coding-agent configurations on the 95-case public set, construction and editing performance are strongly correlated but not interchangeable (Spearman $\rho = 0.78$): Claude Fable 5.1 leads construction, Gemini 3.8 Flash leads editing, and GPT-6 Astra narrowly leads overall. Spatial Composition is the weakest construction category for every agent, while editing remains imprecise: the best Repair F1 is only 0.527, and 35.8% of edits that fully recover the target still introduce unintended changes elsewhere in the scene. These results expose a gap between plausible 3D generation and reliable spatial reasoning and state control.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36777
PDF
-
EASE: Behavior-Adaptive Skill Curation for Self-Evolving Agents
2026-09-29 05:18:24Zhen Xiong, Qiaoyu TanAbstract
Agent skills provide a lightweight mechanism for self-evolving agents to accumulate reusable procedural knowledge without updating model parameters. However, existing learned skill curators typically optimize curation without explicitly modeling downstream executor behavior. We show that this can cause systematic cross-executor degradation: curators trained with different executors perform best when paired with their own training executor, indicating that effective skill curation is executor-dependent. We formulate behavior-adaptive skill curation and introduce EASE, a framework that learns a single curator that adapts its decisions to different executor behaviors. EASE maintains an online behavioral profile of recent execution patterns and conditions the curator on this profile, the current trajectory, and retrieved skills to add, modify, or remove skills from an evolving repository. We train the shared curator jointly across multiple frozen executors with reinforcement learning, using retrieval-aware and behavior-aware temporal attribution to focus optimization on curation actions with observable downstream influence. Across ALFWorld, ScienceWorld, and WebShop, with executors ranging from Qwen3-8B/32B and GPT-OSS-120B to unseen Kimi K2.6, DeepSeek V4 Flash, and Gemini 3.5 Flash, EASE outperforms strong skill- and memory-based baselines without per-executor finetuning. EASE also maintains 34.5--41.0% fewer skills, improves skill retrieval by 36.3--38.7% and measured edit utility by 51.8--60.0%, and reduces deployment-time inference tokens by 9.1--14.5%. These results establish behavior-adaptive skill curation as an effective principle for building self-evolving agents.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36746
PDF
-
Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models
2026-09-29 05:07:47Ayan Sengupta, Vaibhav Seth, Tanmoy ChakrabortyAbstract
Knowledge Distillation (KD) trains a smaller-capacity student model to imitate a larger-capacity teacher model by matching output distributions, implicitly assuming the teacher to be a reliable oracle. In large language models (LLMs), this assumption often fails: teacher predictions can exhibit high entropy and hallucinations, causing standard KD to degrade well-calibrated student priors. We propose CaRE-KD, a confidence-gated distillation framework that replaces static objectives with uncertainty-adaptive optimization. CaRE-KD has two components: a token-level loss (CaRE-Divergence) that adaptively switches between Forward and Reverse KL divergence based on teacher--student confidence, and a batch-level epistemic rejection mechanism (Revival) that suppresses updates when the teacher is more uncertain than the student. We provide a gradient-level analysis showing how this dual-granularity design induces a conditional calibration mechanism that prior static divergences cannot reproduce. Empirically, across eight teacher--student pairs and eleven benchmarks spanning instruction following, chat alignment, code generation, and mathematical reasoning, CaRE-KD delivers consistent gains over strong baselines (Skewed-KL, $\alpha$--$\beta$ divergence). Highlights include up to $+3.2$ average ROUGE-L on instruction-following tasks, $+2.1$ pass@1 on MBPP, $+1.7$ accuracy on GSM8k, and $+1.8$ accuracy on CollegeMath over the strongest baseline, with consistent gains in LLM-as-a-judge factuality (up to $+2.5$ per task over Skewed-RKL). Revival further acts as a principled, loss-agnostic plug-in that systematically strengthens existing distillation objectives by filtering epistemically unreliable teacher supervision.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36734
PDF
-
CoSec: Benchmarking Agent Security in Communities
2026-09-29 04:44:54Hao Chen, Wenhui Dong, Ye Chen, Jiezhi Yao, Chenbo Xia, Yuwen Qu, Renxiang Wang, Fudong Yuan, Camil Hamami, Chenglong Pan, Xinquan Yue, Ziyu Wang, Fengyu Ye, Chenyang Si, Caifeng ShanAbstract
LLM agents operate in persistent collaborative environments involving multiple users, communities, memories, files, and tools. Community boundaries may remain fixed or evolve with changes in membership, roles, composition, and relationships. Agents must complete legitimate tasks and prevent unauthorized disclosure of protected information. Existing evaluations do not fully examine these risks in agent systems. We introduce \textbf{CoSec}, an executable benchmark for evaluating privacy and authorization enforcement in LLM agent systems operating within and across communities. CoSec contains 208 canonical scenarios spanning fixed and evolving boundaries, protected information belonging to the agent owner or other participants, and attacks through dialogue, environmental content, persistent memory, and composed workflows. CoSec executes complete agent systems with persistent sessions, memory, files and tools. It verifies information flows against the active authorization state using execution traces and artifacts. Across harness and model configurations, agents frequently complete benign tasks but violate privacy and authorization boundaries. Privacy behavior varies across harnesses, attack surfaces, and community states, revealing how memory, files, tools, and workflows can carry protected information beyond its authorized scope. These findings show that task utility does not imply privacy or authorization compliance and that authorization in community settings remains an unresolved security challenge for persistent LLM agents.
Abstract (translated)
URL
https://arxiv.org/abs/2609.34790
PDF
-
Normalize-Then-Precondition: A Hierarchical Approach to Marginal Scale and Interaction Geometry for LLM Training
2026-09-29 04:32:28Zixuan Gong, Zeyu Gan, Jiaye Teng, Yong LiuAbstract
Matrix optimizers have emerged as a promising direction, with Muon standing out as a prominent design. Revisiting Muon through its full-Gram representation, we observe that it jointly processes marginal-scale and interaction information. This opens an alternative way to organize geometric information hierarchically, motivating the Normalize-Then-Precondition framework. Specifically, it first uses diagonal-Gram information to construct a marginally normalized update, then applies spectral preconditioning to its directional interaction geometry. Building on this framework, we develop NormPre with NormPre-G and NormPre-L adopting global and localized spectral preconditioning, grounded in spectral-norm steepest descent and a regularized formulation followed by leading mode selection, respectively. To enable large-scale training, NormPre-G uses Newton-Schulz iterations and NormPre-L employs randomized sketching to approximate the leading interaction eigenspace. Theoretically, we establish $\mathcal{O}(T^{-1/2})$ convergence guarantees for simplified versions of NormPre. Across extensive pretraining experiments on GPT-2 Small, LLaMA and Qwen3, both variants consistently outperform AdamW, Muon and MANO under matched training budgets. Further efficiency and spectral analyses reveal the complementary strengths of two variants and characterize their performance-efficiency trade-off. We open-source our code through a GitHub repository at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36692
PDF
-
Quantization Error Is Spectrally Flat: A Single Random Probe Is a Calibrated, Data-Free Sensitivity Estimator, with Application to Budget-Targeted Mixed-Precision Quantization
Abstract
A single random Gaussian probe gives an unbiased estimate of the squared Frobenius norm of a layer's quantization error. The estimator is well-behaved because round-to-nearest error is spectrally flat. Across 1,683 tensors from a 35B MoE and a 9B dense model, effective dimensionality is 0.93 to 0.96 times the i.i.d. noise value of the same shape, and on the MoE the median is unchanged from 2-bit to 8-bit. The probe coefficient of variation is predictable from tensor shape. One probe measures per-tensor sensitivity to within 4 to 7%; twenty probes reach 1.3 to 1.4%.RAM applies the propagated form of this estimator to budget-targeted mixed-precision quantization with no calibration data. Gaussian probes carrying the network's own input statistics score every tensor at six bit-widths. A knapsack solver allocates bits under an exact byte budget, with guardrails against catastrophic 2-bit assignments. One probe pass serves any budget. Isolated and propagated scores rank tensors independently on Qwen3.5-35B-A3B (Spearman -0.01), yet the propagated probe rank-correlates 0.81 to 0.83 with the GPTQ layer objective from real activations, while the isolated estimator is uncorrelated with it. That objective is the wrong allocation target: at matched bytes on Qwen3.8-27B, a block-output probe beats a vendor IQ3_M mix and an oracle that allocates from the real-activation objective. On Qwen3-8B the propagated probe ties HAWQ-V2 at matched bytes. Across seven architectures from 8B to 122B, with probe timing up to a 400B model in nine minutes on one workstation, RAM reaches 3.5 to 13.6% lower median WikiText-2 perplexity than size-comparable uniform 4-bit builds on the tested MoE models. (Black Sheep Ai this http URL)
Abstract (translated)
URL
https://arxiv.org/abs/2609.33923
PDF
-
MCP Error Messages Written for Developers Hurt the Most Capable Agents Most
Abstract
Many Model Context Protocol (MCP) servers wrap web APIs built for human developers, and their error messages tell the reader to run a command, edit a configuration, open a web page or wait. Many agents that read them can only call the server's tools. In 150 widely used MCP servers, 949 of 3,001 error messages tell the caller what to do next, and half of these steps depend on something the server cannot see about the caller. On credential errors, 62 of 67 steps ask for a terminal command, a configuration change or a web page; on rate limits, 20 of 30 say to wait and retry without naming the call to repeat. We tested five OpenAI models that act only through the tools of Berkeley Function Calling Leaderboard tasks, and the agents did what the step said. On expired credentials, a terminal command in the step left 45% of tasks recovered, and the loss it caused grew from 18 points for GPT-5.5 to 69 for GPT-6 Astra. On a rate limit, GitHub's "Wait before retrying." left 6%. We tested two remedies. For MCP developers, naming a server tool in the step raised recovery on expired credentials to 84%, with the login tool in place of the command, and on a rate limit to 88%, with the call to repeat in place of the bare wait. For agent developers, deleting the step with a one-sentence prompt before the model reads it raised recovery on expired credentials to 82%.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35381
PDF
-
Replay the Curvature: Accurate and Scalable NVFP4 Quantization for Large Language Model Inference
2026-09-29 03:55:29Ruiyi Ding, Jie Li, Kang He, Ziyan Liu, Chengru Song, Yuedong Xu, Yuan ChengAbstract
Large language models make weight storage and memory traffic major inference costs, motivating low-precision formats that represent each weight with only a few bits. Such formats use a scale to map floating-point values into a small codebook; NVFP4 improves local range utilization by letting every 16 E2M1 weights share an E4M3 block scale. Choosing that scale is difficult in GPTQ because quantizing one column updates those that follow, so evaluating a block independently can misestimate its final reconstruction error. Large models pose a second challenge: full-precision weights, calibration activations, and second-order state cannot all remain on one accelerator, while assigning complete layers to devices leaves each time-consuming layer solve serial. We introduce \emph{Schur Replay}, a scale-selection algorithm that reproduces the GPTQ updates caused by each block scale and scores the resulting block error after accounting for compensation from unquantized columns. Separately, our execution infrastructure keeps only the active layer resident, tiers activations across device, host, and disk, retires full-precision layers after export, and distributes independent output rows across tensor-parallel ranks. Together, the algorithm and infrastructure attain $99.35\%$ and $100.84\%$ question-weighted recovery from BF16 across seven benchmarks on Qwen3.5-397B-A17B and Llama-3.3-70B-Instruct. On the 397B model, the infrastructure reduces measured per-layer time by $15.17\times$ over ModelOpt and $23.14\times$ over LLM Compressor, with lower memory used per GPU.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36654
PDF
-
Inducing Process Supervision from Outcome-Only Reinforcement Learning
2026-09-29 03:46:45Shengda Fan, Xin Cong, Zhong Zhang, Haotian Chen, Yankai LinAbstract
Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning. However, training strong PRMs remains costly: human step annotation is difficult to scale, while Monte Carlo estimation is computationally expensive and can drift from the intrinsic correctness of steps. To get effective PRMs at low cost, we introduce TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning (RL) framework for training generative PRMs. In TIPS, the model generates a chain-of-thought (CoT) followed by step-level labels and an outcome label. The reward depends solely on whether the predicted outcome matches the ground truth, and the resulting group-relative advantage is used to optimize the entire generated response. Intuitively, when checking intermediate steps helps determine the outcome, more accurate checks can lead to better outcome judgments and higher rewards. Outcome-only RL can therefore reinforce step-level verification without explicit process supervision. We validate the effectiveness of TIPS across math and agent benchmarks and four backbone families. Notably, TIPS-Qwen3-4B-Thinking-2507 reaches 85.2 F1 on ProcessBench with only 3.2K outcome-labeled trajectories, surpassing all evaluated trained PRMs and strong prompt-only judges such as GPT-5.4-Instruct and Claude-4.7-Opus, while still trailing o1-mini. Code and data are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36641
PDF
-
Generating Edit-Inducing Questions for AI Research Manuscripts
2026-09-29 03:29:57Sebastian Joseph, Zichao Wang, Jennifer Healey, Alexa Siu, Junyi Jessy Li, Ani NenkovaAbstract
We study the ability of LLMs to generate edit-inducing questions whose answer will improve a paper draft. On a dataset of paired submission and camera-ready papers from ICLR and NeurIPS, we compare the helpfulness of questions from GPT models with or without full paper context to that of human reviewers. GPT produces more edit-inducing questions and its questions are associated with more extensive edits and cover a broader range of edited content compared to questions from reviewers. However, a much smaller percentage of the GPT questions are edit-inducing. Our analyses confirm that automated questions can be beneficial to authors and highlight an example task where proper attending to long context deteriorates reasoning model ability to produce helpful output.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36617
PDF
-
RoboChrono: A Real Robot Benchmark for Streaming Task Understanding
2026-09-29 03:21:32Yuzhou Wu, Longteng Fan, Zimeng Li, Yu Wanchan, Ting Zhang, Yiyang Ma, Shihao Li, Wei Ying, Jianbin Qin, Jiajian Jing, Fangwen Chen, Yifan Wu, Zichen Zhang, Ruiqi Yang, Weibin Kong, Yihang Xu, Haoran Liu, Zonghang He, Xuyang Liu, YiFan Xiong, Siteng Huang, Tao Xu, Zhuo Xu, Long Chen, Ruoxiang LiAbstract
Understanding ongoing robot manipulation requires models to interpret visual observations in relation to interaction history and task progress. We introduce RoboChrono, a benchmark for streaming task understanding comprising 39 scenarios and 34,713 evaluation instances, constructed from real robot executions and complementary bare-hand human recordings. The benchmark evaluates seven tasks grouped into recognition, alignment, and temporal grounding, covering action understanding and anticipation, visual correspondence, temporal ordering, and action localization. Zero-shot evaluation of 18 vision-language models reveals substantial differences across tasks. GPT-6-Astra achieves 98.3% accuracy on Frame Matching but 68.3% on Frame Ordering, while RynnBrain1.1-122B-A10B exhibits a larger gap, reaching 95.4% and 32.9%, respectively. Input ablations on matched questions with five open-weight models further reveal distinct dependencies on visual evidence: removing visual observations reduces Current Action Recognition accuracy by 22.1 percentage points, whereas Next Action Prediction decreases by only 0.7 points. These findings show that strong visual matching does not consistently coincide with strong temporal ordering, and suggest that next-action prediction can be supported by task and action priors even when visual evidence is unavailable. RoboChrono provides a diagnostic setting for examining these differences, highlighting the need for capability-specific evaluation beyond aggregate scores when assessing task understanding in robot manipulation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36605
PDF
-
Grounded Revision vs. Prior Injection: Probing Retrieval-Augmented Patent Claim Amendment
2026-09-29 02:37:33Josepha Michiko Leo, Hyun-seok Min, Yehoon Jang, Irvan Zidny, Jin-Woo Chung, Sungchul ChoiAbstract
Retrieval-augmented generation is widely used in professional writing, yet whether retrieval grounds revision or merely injects templates is rarely tested where "correct" has a definable meaning. Patent claim amendment supplies that signal: the examiner names the attacked limitation and cites prior art, providing per-case ground truth. We release three artifacts: (i) a corpus of 7,385 USPTO prosecution cases with XML-aligned pre/post claims, rejection, and cited prior art; (ii) a seven-probe battery comparing random and structural-match retrieval as two policies under a fixed prompt scaffold; (iii) a deterministic five-channel metric (C1-C3 and C5 in main, C4 supplementary) requiring no LLM evaluation. Across 9,600 pre-registered calls on four frontier LLMs (Claude Sonnet 4, Claude Haiku 4.5, GPT-5.4, GPT-4o-mini), no tested model exhibits detectable classical prior-injection behavior; retrieval effects are small and direction-inconsistent between random and structural retrieval, and the null is unchanged under a dense (semantic) retriever, across retrieval depths k in {1,3,5,10}, and under a paraphrase-sensitive grounding metric. Revision locality reveals a model-specific difference that the template channel misses. The four-cell taxonomy, which we treat as exploratory, leaves the prior-injector cell unoccupied.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36550
PDF
-
Towards Breaking the Learning System Wall Using Multimodal Tutoring Transcriptions
2026-09-29 01:56:08Danielle R. Thomas, Marie Cynthia Abijuru Kamikazi, Ashish Gurung, Ishan Miglani, Shivang Gupta, Zachary Levonian, Conrad Borchers, Kenneth R. KoedingerAbstract
Past research using log data has faced the "learning system wall," whereby few methods exist for generalizing models of student learning across platforms. Increasingly, online learning is captured by richer forms of data, including dialog and video, with new affordances. An example of this is remote tutoring programs, where human tutors support students who use learning systems while video conferencing. Toward better platform-general modeling of learning, we introduce an AI-driven multimodal transcription system that processes screen-recording videos into unified screenplay-style transcripts containing audio dialogue and annotated learning log actions. We describe a planned method for temporally aligning AI-generated multimodal transcripts with MATHia learning logs and for identifying and classifying student learning processes to align with MATHia logs. Lastly, we highlight challenges and potential solutions in capturing learning processes in one system, offering initial steps towards generalizing log data across diverse systems.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36502
PDF
-
DuplexSpeechBench-Document Grounding: Benchmarking Document Grounding and Hallucinations in Voice Agents
2026-09-29 01:43:55Puneet Mathur, Nedim Lipka, Zeyu Jin, Dinesh ManochaAbstract
Voice agents enable low-latency, natural interaction, yet their ability to faithfully ground responses in external documents remains underexplored. We introduce DuplexSpeechBench-Document Grounding (DSB-DG), a benchmark for evaluating document grounding in voice agents across five professional domains. DSB-DG targets three failure modes: Context Saturation, which measures grounding under increasing document length; Grounding Decay, which measures retention of document facts across multi-turn dialogue; and Proactive Grounding, which evaluates whether context re-injection mitigates conversational drift. The benchmark contains 1,636 adversarially verified QA pairs from 50 documents covering five professional domains, and supports fully automatic evaluation of grounding accuracy, hallucination, and response latency. Across systems spanning cascaded, proprietary full-duplex and real-time, and open-weight speech2speech architectures, we find substantial differences in effective grounding capacity. While cascaded pipeline (ASR-LLM-TTS) achieves the highest grounding accuracy, Gemini-Live and GPT-Realtime closely trail behind. Open-weight systems exhibit distinct failure modes, most notably an abrupt context-capacity collapse and multi-turn grounding decay. More broadly, grounding fidelity degrades with context and conversational load, and failures frequently manifest as unsupported generations rather than abstention. We show that contextual grounding as a key unresolved challenge for reliable full-duplex voice agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00316
PDF
-
Reward-rate Policy Gradient for Efficient Machine Learning Engineering Agents
2026-09-28 23:46:08Muhang Tian, Sherry YangAbstract
Traditional reinforcement learning (RL) techniques focus on maximizing expected cumulative reward, where each action assumes to take a constant unit of time. However, this assumption does not hold for agentic RL tasks such as machine learning engineering (MLE) agents, where actions involve data loading, feature engineering, and model training that take variable durations. Efficiency matters in modern agentic RL where actions are costly. To address this limitation, we adapt from continuous-time RL and Semi-Markov Decision Process (SMDP) formulation and propose Reward-rate Policy Gradient (RPG), where we focus on optimizing the reward rate -- the long-term reward per unit of time. RPG estimates the reward rate from off-policy samples, then charges each action for the time it consumes at that rate. We first conduct theoretical analysis in the bandit setting to establish that RPG approximates the optimal reward rate and empirically demonstrate it outperforms baselines while avoiding enumeration over the policy space, a known issue for an existing method. We then further apply RPG on a small language model (Qwen3.5-4B) with self-improvement loops and empirically show it obtains higher rewards within a fixed time budget than vanilla RL on MLE-Bench and NanoGPT, with a 19.2% and 85.7% margin, respectively. Our method provides a practical solution for optimizing performance under wait time considerations in modern agentic RL tasks, where actions interact with external environments and cost time.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36393
PDF
-
LEGO-Anything: Coding Agents for 3D Scene Reconstruction
2026-09-28 23:17:39Xirui Li, Peng Shi, Mingwen Dong, Sheng Zhang, Zhuoyan Xu, Dongkyu Lee, Shuaichen Chang, Yi Xiang, Lin Pan, Jiarong JiangAbstract
A 3D scene reconstructed from a single image is most useful when represented not as a rendering or a fixed 3D output, but as an explicit scene program whose execution yields a scene that can be inspected, edited, and queried. We present LEGO-Anything, an Image-to-Code framework in which a coding agent iteratively writes and executes Blender code, inspects scenes and renderings, and revises the program. To evaluate end-to-end scene recovery, we introduce LEGO-Bench, a simulator-grounded benchmark with 208 images from 104 diverse indoor and outdoor scenes. LEGO-Bench separately scores artifact validity, visible-surface geometry, and rendered appearance. Its simulator-grounded design enables extensibility and precise automatic evaluation. Among evaluated agents, GPT-6-astra achieves the strongest overall results, with 53.4% indoor and 39.6% outdoor scores, yet substantial gaps remain between delivering valid scene artifacts and faithfully recovering scene geometry and appearance. Analysis of agent construction trajectories reveals three recurring issues: weak scene initialization, regressive edits during iteration, and unreliable self-evaluation. These findings motivate LEGO-Plugin, a training-free harness plugin for more controlled iterative scene construction, which improves all six evaluated models, with relative gains of up to 62.7% in overall score. Finally, we test whether reconstructed scenes can represent natural images and support vision tasks. In LEGO-World, we derive object detections, instance masks, and relative depth as deterministic queries on scenes reconstructed by GPT-6-astra. These readouts show non-trivial performance across all three tasks but fall well short of specialized vision models, suggesting that program-constructed scenes from current coding agents are a promising but not yet sufficiently precise representation of natural images.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36380
PDF
-
ThuRunel: Dynamic Decoupling for Structured Advisory Dialogue
2026-09-28 22:23:43Yuyan ChenAbstract
High-stakes advisory domains such as medical aesthetics, legal consultation, and educational planning exhibit a two-phase structure. The early phase requires empathetic elicitation and emotional support, and the late phase requires authoritative specialist judgment. Neither fully automated agents nor human junior consultants adequately address this structure at scale. We formalize the core design challenge as dynamic decoupling, asking how an AI advisory agent should decide what to ask, when to stop, what to resolve autonomously, and what to forward to the specialist. We present ThuRunel, an advisory agent combining a finite-state belief management framework, a chain-of-thought teacher synthesis protocol, and learned generation adapters. Against eleven baselines, ThuRunel achieves consistent improvements in elicitation completeness and specialist brief quality. ThuRunel is publicly deployed as a bilingual web application in which the same decoupling decisions operate from the client's side, grounded in a curated knowledge base that cites its sources in every answer.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36340
PDF
-
Illusory Truth or Mere Exposure? Model-Dependent Repetition Effects in LLM-Based Social Media Simulations
2026-09-28 21:09:47Azza Bouleimen, Nicol\`o Pagan, Anik\'o Hann\'akAbstract
Generative agent-based models (GABMs) are increasingly used to simulate social media dynamics, including misinformation spread. For such social simulations to be valid proxies of human behavior, LLM agents should replicate established human cognitive biases, among them the Illusory Truth Effect (ITE), where repeated exposure to a claim increases its perceived truth value. We investigate whether and how the ITE manifests across four LLMs (Gemma-3-4b-it, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and GPT-5-nano) in a social media simulation context. We propose a two-phase within-context experimental design that embeds the repetition manipulation inside a realistic news feed interaction. Using this design, we collect 336,000 truth, importance, sentiment, and interest ratings across 100 statements, 10 feed variants, and 3 replications. The key comparison is between ratings assigned to repeated statements, seen throughout a simulation phase, and completely unseen ones, rated within the same experimental context window. We distinguish genuine ITE (truth-specific repetition boost) from mere exposure effects. We run an OLS regression followed by a Linear Mixed-Effect Model to account for differences across models and ratings. Our results reveal four qualitatively distinct patterns: Gemma-3 exhibits a genuine ITE; Qwen2.5 shows a mere exposure effect; GPT-5-nano displays no repetition effect on truth and mild skepticism toward repeated content; Llama-3.1 shows a small truth boost alongside decreases in evaluative dimensions. Crucially, temperature has no effect on these findings, and a variance decomposition highlights the high context-sensitivity of LLM rating behavior. Our findings caution against assuming uniform ITE replication across LLMs in social simulations, while suggesting that Gemma-3-4b-it may offer the most behaviorally realistic approximation for misinformation-related simulations.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36278
PDF
