The paper proposes treating each atom as an agent that predicts the behavior of its neighbors through collective predictive coding (CPC), a framework borrowed from language emergence. By applying CPC to classical spin systems—such as two-dimensional Ising models, a stripe‑ordered spin model, and the Sherrington–Kirkpatrick spin‑glass—macroscopic order parameters, including replica overlap, arise automatically from observed spin configurations. For the Ising cases, the equilibrium distribution of the learned collective predictive order reproduces a Landau‑type potential, demonstrating that macroscopic order can emerge without reference to microscopic Hamiltonians.
By Koji Hashimoto, Tadahiro Taniguchi
The paper introduces EvidenceConnect, a data‑synthesis framework that generates training examples by zooming into related image regions to create questions requiring comparison, aggregation, or chaining of visual evidence. By retaining only reference answers agreed upon by independent vision‑language models and discarding questions that survive region ablation, the authors train base VLMs with RLVR to produce Zooming without Zooming v2 (ZwZ2). Although the synthesized data are not benchmark‑specific, ZwZ2 shows improvements across fine‑grained and spatial perception, embodied spatial reasoning, multimodal long‑context understanding, multimodal agent tasks, and general visual reasoning, demonstrating that strengthening the shared capability of connecting evidence yields more generalizable multimodal perception.
By Lai Wei, Yatai Li, Liangbo He, Jun Lan, Renjie Gu, Linhan Cao, Junkai Lin, Huijia Zhu, Weiqiang Wang, Linghe Kong, Yue Wang, Zhuosheng Zhang, Weiran Huang
The paper investigates how language‑model agents use external procedural skills and discovers a bias called Procedural Exposure Bias (PEB), where repeated exposure to a skill’s pattern makes the model favor its associated action even when the context is unchanged. PEB is much weaker when the skill is presented visually rather than textually, mainly due to visual encoding rather than layout. Experiments across multiple environments show that visual skills lead to higher task success than textual ones, highlighting the importance of skill representation for agent decision‑making.
By Qing Yang, Yuqi Xiong, Pengcheng Huang, Zhipeng Xu, Zhenghao Liu, Zulong Chen, Yu Gu, Ge Yu
arXiv:2610. 37276v1 Announce Type: cross Abstract: Large language model (LLM)-driven agent scaffold evolution improves agent workflows through iterative candidate generation and evaluation.
By Sichun Luo, Yi Huang, Mukai Li, Haochen Luo, Guanzhi Deng, Haibo Wang, Jian Xu, Zefa Hu, Junlan Feng, Qi Liu
The paper investigates how different natural language plan representations affect the performance of large language model (LLM) web agents. Using the PlanAhead framework, the authors automatically classify WebArena tasks into three difficulty levels and evaluate four plan types—sequential subgoals, narrative, pseudocode, and checklist—on hard tasks across multiple multimodal LLM agents. They introduce a decomposition of Success Rate into Achievement Rate and Solved-Task Consistency to better understand stochastic variability, finding that both plan formulation and the LLM used to generate the plan significantly influence agent robustness and success.
By Alejandra Zambrano, Sara Vera Marjanovic, Imene Kerboua, Xing Han L\`u, Leila Kosseim
The paper defines an autonomous agent as a system that maintains task goals and viability within explicit identity and authority boundaries, continuously acting with world and self-models, and adjusting control mechanisms through feedback. It traces the evolution from homeostasis to biological autonomy and identifies six operationalization lines, culminating in a one-boundary, four-loop framework that organizes task execution, self-maintenance, metacontrol, and governed learning. Five simulation experiments test permission constraints, road-closure response, health maintenance, capability re-estimation, and cross-task learning, offering preliminary evidence for the feasibility of these mechanisms while noting the need for broader testing.
By Hongda Zhang, Ziqing Zhou, Yuning Chen, Linqiang Hu, Yi Liu, Ding Zhou, Zhao Huang, Haining Xu
AT-AT (Automatic Tool Optimization from Agent Traces) is an agentic tool optimizer that improves shared toolsets offline using historical agent executions while keeping agents unchanged. It employs a lightweight coding agent to read agent trajectories and evaluation results, selectively editing tool descriptions and implementations when allowed. In benchmarks, AT-AT achieves the highest task completion rates in Airline and Telecom domains at lower cost than strong baselines, and it also improves performance on TerminalBench-2 and TBLite, while reducing cost or error rates on production MCP servers.
By Maya Ravichandran, Pragya Tripathi, Dharamendra Kumar, Adith Swaminathan, Nathan Kallus
BellowsKV introduces a training‑free approach that maps the hidden states of multiple adjacent layers in a large language model to a single shared latent representation per token, allowing each layer to retrieve its own keys and values via a layer‑specific projection. The method retains the latent at full dimension and offers two compression strategies: BellowsKV‑Quant, which stores the latent at low precision, and BellowsKV‑Evict, which makes a shared eviction decision per block. Experiments on Llama‑3.1‑8B, Mistral‑7B, and Qwen2.5‑7B show that BellowsKV‑Quant preserves about 99% of full‑cache accuracy at 3.6× compression and 97‑98% at 13‑14×, outperforming existing cross‑layer methods that drop to at most 84% accuracy at 16× compression.
By Jiamu Zhang, Hanjie Chen
The paper introduces Hamiltonian Temporal Contrastive Curiosity (HTC²), a method that combines multi‑horizon temporal representation learning with Hamiltonian phase‑space structure to guide exploration in physical control environments. HTC² uses a horizon‑conditioned contrastive critic to learn relationships between future state pairs within an episode, and the resulting representation distance serves as an intrinsic reward. The reward is further modulated by bounded multi‑step Hamiltonian variation, and the framework is integrated with Soft Actor‑Critic and Proximal Policy Optimization, evaluated on six DeepMind Control Suite tasks against several baseline curiosity methods.
By Shuwei Wang, Qiang Guan, Brian Qiu, Haoxiang Huang, Haoxi Wang, Chenyu Li, Bo Xu, Sen Cui
The paper investigates how fine‑grained credit assignment affects learning in multimodal agentic reinforcement learning, using video editing as a testbed. Four movie‑trailer editing tasks—Content Selection, Narrative Sequencing, Rough Cut, and Fine Cut—are introduced to represent real‑world editing workflows with varying complexity and horizon lengths. The study finds that decaying process rewards significantly boost performance on the long‑horizon Fine Cut task, while offering modest gains on shorter tasks; however, fixed process‑reward weighting can lead to reward‑hacking and degraded outcomes, and on‑policy self‑distillation provides limited benefit.
By Lei Zhao, Zongheng Cao, Xinyu Hu
FOCUS is a reinforcement learning method designed to improve long‑horizon language‑model agents by using selective branching to train a critic network more accurately. The rollout controller actively samples continuations from states that are expected to reduce the critic’s estimation error, and the refined value estimates are used to score segments between consecutive states with a TD(λ) advantage. Experiments on legal reasoning, deep search, and code localization tasks show that FOCUS outperforms baselines such as GRPO and VinePPO, achieving significant accuracy gains on benchmarks like BrowseComp‑Plus and DeonticBench.
By Zhengping Jiang, William Jurayj, Liaoyaqi Wang, Chunsheng Zuo, Anqi Liu, Benjamin Van Durme
AgentBaton explores how capable and efficient language models can collaborate on long‑horizon agentic tasks by handoffs, where one model takes over from another. The study finds that handoffs from a capable to an efficient model at appropriate stages preserve or improve accuracy while significantly cutting cost, whereas the reverse handoff is less effective. AgentBaton, a lightweight detector, learns the optimal handoff point and achieves an average API cost reduction of 65.9% across 11 model pairs while largely maintaining task accuracy, even gaining up to 8.8 percentage points over capable‑only execution.
By Junlong Li, Jindong Jiang, Yonggan Fu, Jiarui Yao, Hanrong Ye, Wonmin Byeon, Peter Belcak, Hao Zhang, Pavlo Molchanov
The study shows that in memory‑agent benchmarks, simply retrieving the correct fact does not guarantee accurate answers. On the FactConsolidation task, increasing the evidence budget from 64 to 512 tokens caused accuracy to drop, even though the newest statement was present in every retrieved set. Experiments with Qwen models revealed that replacing an unrelated statement with an obsolete competing statement drastically reduced accuracy, indicating that outdated information can be more harmful than irrelevant evidence.
By Oleeviya Babu Poikarayil, Tony Basil Babu, Abdulrahman Nahhas, Het Darshan Mehta, Jawid Ahmad Baktash, Ernesto William De Luca, Klaus Turowski
The paper introduces Policy-Compatible Reinforcement Learning (PICO), a method that addresses the learning cliff in reinforcement learning with verifiable rewards (RLVR) by performing distribution-constrained joint decoding. PICO guides token-level decoding with expert data while preserving the model’s token probabilities, and uses a dynamic top‑k mechanism to keep sampling within a finite divergence budget. The resulting trajectories are verified and policy‑compatible, and are jointly optimized with on‑policy rollouts using a mixed‑behavior objective that mitigates distribution mismatch via importance sampling, leading to consistent performance gains over standard RLVR and prior off‑policy baselines on diverse reasoning datasets and agentic benchmarks.
By Zehao Wang, Hanzhu Chen, Runquan Gui, Zhe Wang, Lin Yang
CASEL (Consequence-Aware Selective Execution Layer) is a runtime allocation policy for tool‑using language agents that chooses among direct execution, verification, sandboxing, clarification, approval, and abstention for each proposed tool call. It combines calibrated failure probability, state‑dependent consequence, reversibility, ambiguity, and evidence freshness to minimize a per‑mode cost. In controlled simulations, CASEL cuts harmful‑side‑effect trajectories from 14.75 % to 7.02 % and reduces mean normalized harm from 0.184 to 0.123 compared to an uncertainty‑ranked gate.
By Yanyan Zhang, Fei Shen, Ting Yan, Hao Zhang
The paper introduces Re4, a scientific computing agent that employs a four-step logical chain—rewriting, resolution, review, and revision—using three large language models: a Consultant, a Programmer, and a Reviewer. The Consultant rewrites problem descriptions by linking them to domain knowledge, the Programmer generates and runs code, and the Reviewer iteratively refines the code through self-debugging. Experiments on PDEs, fluid dynamics, ill‑conditioned linear systems, and data‑driven physics show that this collaborative framework markedly improves bug‑free code generation and execution success rates across several LLMs.
By Ao Cheng, Lei Zhang, Guowei He
PIVOT (Plan-Inspect-eVOlve Trajectories) is a self‑supervised framework that refines large‑language‑model agent plans by treating trajectories as optimizable objects. It operates in four stages—PLAN, INSPECT, EVOLVE, and VERIFY—using structured losses and textual gradients to iteratively improve trajectories and ensure compliance with task constraints. Experiments on DeepPlanning and GAIA show that PIVOT achieves up to 94% relative improvement in constraint satisfaction with human‑in‑the‑loop feedback and still delivers significant gains autonomously, while using 3x to 5x fewer tokens than competing methods.
By Tuo Zhang, Alin-Ionut Popa, Yan Xu, Rui Song, Dimitrios Dimitriadis
The paper "Cross-User Poisoning: User-Task Boundary Failures in Multi-User Collaborative Language Agents" examines how shared language agents, which handle interleaved messages from multiple users, can be exploited through cross-user poisoning (CUP). CUP occurs when an adversary injects a message into the shared context that later influences the agent’s actions for a benign user, leading to unauthorized responses or actions. The authors validate CUP on two deployed agents, Continua and ElizaOS, and introduce the MURMUR framework to evaluate such attacks. Their experiments across Slack, Workspace, and Airline domains show CUP’s high success rate and persistence, outperforming prompt-injection attacks. They also test defenses, finding that context summarization alone is insufficient, while task clustering and provenance prompting can significantly reduce non-adaptive propagation. The study concludes that robust multi-user agents must implement explicit user/task scoping rather than relying on generic input filtering or context compression.
By Atharv Singh Patlan, Peiyao Sheng, S. Ashwin Hebbar, Prateek Mittal, Pramod Viswanath
The paper introduces Agentic Distributional Regression (ADR), a method that uses a language‑model agent to generate a short program estimating conditional densities from iid data. The agent constructs the estimator from a library of probability distributions, producing a parametric, mixture, or multivariate factorization whose parameters are explicit functions of covariates. The authors provide finite‑sample error bounds, an oracle inequality for model selection, and show asymptotic normality of fitted constants, with simulation results indicating ADR outperforms other methods on twelve real datasets and a new post‑training dataset.
By Lorenzo Tedesco (University of Bergamo)
The paper introduces the Theory of Strategic Evolution, outlining three core principles—situated choice, reproductive transmission, and institution‑mediated coupling—that govern strategic replicators. It formalizes how agents’ choices influence continuation or reproduction, how successors inherit organizational structures, and how institutions shape these dynamics. The work includes formal models, proofs, and examples, such as a market‑rent‑entry model and a deployment scenario, to illustrate the theory’s application to strategic replication.
By Kevin Vallier