The study examines how language‑model agents compact context for a current question and whether that compacted context remains useful for subsequent questions. Using four corpora that annotate evidence for multiple questions per text, the authors compare various compaction methods against random deletion at the same token budget. While sentence‑level compaction improves accuracy on the current question (e.g., gpt‑6‑luna boosts MultiRC accuracy by 11.4 points at 25% tokens), it does not outperform random deletion on later questions; word‑level deletion (LLMLingua‑2) shows better performance on later MultiRC questions and retains more evidence content words, though its advantage varies across datasets and evaluation settings.
By Junwei Yao, Yongxi Zhou, Zike Yang, Zhijun Wang, Jianan Liu, Yanxian Chen, Wenbo Ye
World2HOSI is a multi‑expert harness that generates long‑horizon human‑object‑scene interaction (HOSI) data by orchestrating existing human‑object interaction (HOI) and human‑scene interaction (HSI) models. It builds a persistent scene state from decoupled background and object assets, plans action stages, invokes experts via their native interfaces, and updates the scene after each action, ensuring physical plausibility through bounded residual optimization and contact‑driven physics simulation. Experiments and a perceptual user study demonstrate that World2HOSI achieves superior task completion and interaction quality while producing more diverse data than prior approaches.
By Jiushu Li, Haoran Wang, Yumeng Liu, Zhiyang Dou, Yuan Liu, Yuexin Ma, Cheng Lin
ClawMMU is a protection layer for large language model agents that tags every memory fragment with its owner, scope, origins, and authority. It enforces that information can only increase in authority but never self‑elevate, checking these labels whenever content enters the model’s context, is stored, retrieved, or used to drive tool calls. By basing checks on provenance rather than content, ClawMMU turns memory attacks into classical protection faults that the agent harness can block even if the model is misled.
By Jinhao Hu, Bardia Mohammadi, Mofasshara Binte Rafique, Lars Henning Klein, Akhil Arora, Ashvin Goel, Laurent Bindschaedler
arXiv:2604. 07667v2 Announce Type: replace Abstract: Multi-agent debate improves LLM reasoning, yet agreement is not evidence of correctness.
By Mengdie Flora Wang, Haochen Xie, Guanghui Wang, Aijing Gao, Guang Yang, Ziyuan Li, Qucy Wei Qiu, Fangwei Han, Hengzhi Qiu, Yajing Huang, Bing Zhu, Jae Oh Woo
The paper introduces Hybrid Belief Reinforcement Learning (HBRL), a two‑phase framework for coordinating multiple autonomous agents in spatial exploration tasks. In phase one, agents build spatial beliefs via a Log‑Gaussian Cox Process and plan information‑driven trajectories using a Pathwise Mutual Information planner; in phase two, a Soft Actor‑Critic policy is warm‑started through dual‑channel knowledge transfer that incorporates belief states and replay‑buffer demonstrations. Experiments on UAV wireless service provisioning and NYC taxi demand sensing show significant performance gains, with the dual‑channel transfer outperforming single‑channel approaches.
By Danish Rizvi, David Boyle
The paper investigates how data should be scaled in agentic reinforcement learning, where data is now a structured design space of tasks, environments, and trajectories. By fitting per‑axis scaling laws across various large language models and RL algorithms, the authors find that environment diversity has the strongest impact, followed by task and trajectory diversity. Using these laws, they price synthesis methods, identify critical synthetic‑to‑real ratios, and formulate a compute‑optimal budget allocation that can improve efficiency by up to 1.7× compared to the common practice of scaling trajectories alone.
By Bowei He, Yankai Chen, Xiaokun Zhang, Changjiang Han, Ye Yuan, Chong Li, Xue Liu
The paper introduces LAMB, a recurrent framework that stores latent multi‑hypothesis beliefs for long‑horizon language agents. LAMB separates a shared Common state containing interaction evidence from weighted Delta components representing competing hypotheses, and updates these hypotheses using relative evidence from action‑outcome predictions without generating textual memory. Experiments on ALFWorld, BabyAI, ScienceWorld, and Crafter show that LAMB matches or surpasses prior methods, notably improving success rates and reducing redundant actions.
By Peien Ding, Hui Li, Yuhan Liu, Jinsong Su
The paper introduces TriMem, a memory system for lifelong LLM agents that stores raw dialogue segments, extracted atomic facts, and synthesized profiles to preserve detail, enable efficient retrieval, and support deep reasoning. It uses TextGrad-based prompt optimization to iteratively refine extraction and profiling prompts without updating model parameters. Experiments on LoCoMo and PerLTQA show TriMem outperforms strong memory baselines across multiple LLM backbones.
By Jingwei Sun, Jianing Zhu, Jiangchao Yao, Tongliang Liu, Bo Han
SpatialScout is a benchmark that tests vision‑language model agents on active, object‑level 3D mapping. It includes 80 indoor scenes with 3,140 target objects, where agents must control a camera, gather RGB observations, and estimate object centers in a shared coordinate frame. Performance is measured by localization accuracy and target coverage using object‑scale overlap scores, revealing that current VLMs, including Claude Opus 5.5, struggle to build precise metric maps.
By Zhaoxi Zhang, Hao Liang, Linxuan Wu, Yiming Wu, Yifu Yuan, Kun Liang, Chenming Tang, Meiyi Qiang, Wentao Zhang, Weikang Li
The paper introduces SMART, a Self‑evolving Multi‑Agent Refinement Framework for conversational text‑to‑motion generation. It enables users to iteratively refine specific spatio‑temporal regions of a motion sequence based on natural language feedback, preserving the rest of the motion. The framework incorporates a semantic‑guided masked editing mechanism and a memory‑based self‑evolving strategy to reuse past editing knowledge and reduce redundant exploration.
By Jianan Li, Jiawei Zhao, Yangtao Zhou, Zhifu Zhao, Shaoxin Wu, Yutao Wang, Dezhen Wang, Hua Chu, Fu Li
HERMES++ is a unified driving world model that simultaneously performs 3D scene understanding and future geometry prediction. It uses a BEV representation to merge multi‑view spatial data for LLM compatibility, introduces LLM‑enhanced world queries for knowledge transfer, and employs a Current‑to‑Future Link to condition future geometry on semantic context. A Joint Geometric Optimization strategy enforces structural integrity by combining explicit geometric constraints with implicit latent regularization, and extensive benchmark tests show that HERMES++ outperforms specialist methods in both point cloud prediction and scene understanding.
By Xin Zhou, Dingkang Liang, Xiwu Chen, Feiyang Tan, Dingyuan Zhang, Hengshuang Zhao, Xiang Bai
DriveVLA-M0 introduces a failure-aware latent memory system for Vision‑Language‑Action models in autonomous driving. It stores past failure cases with scene representations and expert trajectories, and uses a retrieval model to decouple static road structure from dynamic agent interactions. During inference, retrieved cases are injected via a lightweight LoRA‑based test‑time training mechanism, yielding significant performance gains on NAVSIM benchmarks with minimal latency overhead.
By Zebin Xing, Yupeng Zheng, Qiang Chen, Linbo Wang, Yichen Zhang, Pengxuan Yang, Junli Wang, Deheng Qian, Xiaoqing Ye, Junyu Han, Yifeng Pan, Qichao Zhang, Dongbin Zhao
InteractAgent is a multimodal large language model system designed for long‑horizon human‑scene interaction. It plans, executes, observes, and reflects in a closed loop, using a single MLLM for planning and evaluation and a frozen MotionLLM for execution. The system incorporates online observation for step‑wise recovery decisions and memory‑augmented reflection for multi‑round plan refinement, achieving improved task success and reduced penetration on the UniHSI benchmark compared to training‑free baselines.
By Senhe Zhang, Tao Zhang, Yujia Zhang, Zeyu Zhang, Dong Gong
arXiv:2610.24264v2 Announce Type: replace
Abstract: As LLM-based agents are deployed with increasing autonomy in real-world environments, a critical failure mode emerges: agentic misalignment, where...
By Eric Zhang, Terry Jingchen Zhang, Chijioke Ugwuanyi, Jerick Shi, Bernhard Sch\"olkopf, Zhijing Jin
S$^2$-MAS is a society‑scale multi‑agent simulation framework that models 500 million synthetic household personas based on real‑survey data. Its Genese engine separates decision making from state updates, while FinL‑mlx, a compact decision model trained on 2.93 million examples, achieves high accuracy in consumer financial‑choice scenarios. The framework evaluates eight financial task families and simulates 12 million distinct agents, revealing how purchase goals, credit limits, and demographic factors influence consumption, cash flow, debt repayment, and arrears over a 24‑month panel.
By Yang Li, Di Wan, Yike Wang, Chujun He, Zongsi Xu, Xiyu Chen, Sheng Zhong, Deng Binjie, Zhenwei Guo, Hou Jun, Kunhan Guo, Weihao Lin, Tuo Zhou, Hai Liu, Ye Luo
The paper examines inference‑time observation attacks on Safe Multi‑Agent Reinforcement Learning (SMARL), where an adversary perturbs observations of some agents to alter their actions. It introduces the MinMax Attack framework, which jointly selects adversarial target actions for compromised agents while considering the cooperative resistance of benign agents. The authors derive a bound on the induced safety cost and show through experiments that MinMax Attack causes higher safety costs than existing baselines while maintaining task reward, thus improving attack effectiveness and stealthiness.
By Taozheng Zhu, Shixiong Jiang, Fanxin Kong
The paper introduces TEAL, an unsupervised framework that allows coding agents to automatically build and refine reusable tools for data‑lake analytics without requiring labeled questions. TEAL employs an Explorer that proposes tools from the data lake and a Critic that evaluates proposals based on the agent’s own execution traces, retaining only changes that the agent adopts without extra steps. Experiments on KramaBench and LakeQA show that TEAL consistently improves performance across multiple LLM workers, especially under tight step budgets, and that synthetic question generation (TEAL‑Q) can match real‑question evolution.
By Yuqing Yang, Yurong Liu, Deqing Fu, Manav Garg, Sravan Babu Bodapati, Fatma Ozcan, Alon Y. Halevy, Jiani Zhang
The paper introduces the Research Progress Graph (RPG), a framework that captures the full, evidence‑driven evolution of research ideas through a closed‑loop process of experimentation, insight, challenge identification, literature exploration, and revision. Unlike current large‑language‑model agents that only record the final idea, the RPG records each research state, its temporal order, and the reasoning that links successive states, providing a complete, authentic trajectory. The authors compare thirty‑five existing agents to the RPG, showing that none fully preserves the evolving record, and argue that future agents should predict next research states, evaluate ideas as trajectories, and use authentic trajectories as training data.
By Ruochen Li, Zhiyu Chen
The paper introduces Automated Multi-Agent Organization Design (AMOD), a framework that treats the organization of agents as a design object and uses an execution-driven loop to iteratively refine responsibilities and coordination rules. AMOD assigns functional responsibilities to agent teams, evaluates outcomes and interaction traces, and revises the organization to address missing, duplicated, or ineffective roles. The approach also proposes testing responsibilities by removal or replacement to validate their necessity.
By Shao-Jun Xia, Yuner Zhang, Huixin Zhang, Di Wang, Yifan Jiang, Liangxi Liu
arXiv:2610.27714v1 Announce Type: new
Abstract: When does experience memory help a closed-loop LLM driving agent? Published gains cannot say, because the two factors we find to matter most, which war...
By Yingxin Jia, Xuanyi Li