The paper investigates a phenomenon called inertia in computer‑use agents, where agents repeat ineffective actions despite recognizing their futility. By analyzing high‑dimensional activation states, the authors find that inertia corresponds to an absorbing region in activation space where values become stale. They propose a method called R$^3$—Reset, Reroute, Restore—to temporarily reset the agent’s context trajectory, escape the absorbing region, and then restore historical context, achieving a 17‑55% reduction in measured inertia across models.
By Giulio Segalini, Zhi Wen Soi, J\'er\'emie Decouchant, Lydia Chen
The paper introduces Proof‑R1, a reinforcement‑learning framework that trains large language models to generate verifiable proofs for natural‑language logical reasoning tasks. Proof‑R1 only accepts a generated conclusion into the proof state when it satisfies formal verification constraints, ensuring each reasoning step is machine‑checkable. The method also reconstructs the dependency closure that supports the final answer, aligning credit with valid proof steps, and shows improved answer accuracy and verifiability across multiple benchmarks and models.
By Qili Zhang, Qianren Mao, Hanze Cai, Kaiming Zhao, Yuening He, Xihan Lei, Yashuo Luo, Hanwen Hao, Yutong Gu, Likang Xiao, Zhijun Chen, Weifeng Jiang, Haoyi Zhou, Jianxin Li
The paper introduces a framework for designing proactive large‑language‑model agents, centered on three joint principles—Task Capability, Temporal Allocation, and Trust—alongside a five‑dimensional design space. It proposes PROACTIVITY‑GYM, a simulation testbed for evaluating proactive assistance across multi‑day scenarios, and presents empirical findings that highlight performance gaps and the importance of aligning interventions with user trust. Human studies show that misaligned interventions can sharply reduce trust, even when outcomes are correct, underscoring the need for careful joint optimization of the 3T principles.
By Jio Oh, Seunghyun Do, Young-Jun Lee, Steven Euijong Whang, Dongyeop Kang
SeekVLN is a new framework for Vision‑Language Navigation that addresses the problem of agents acting on insufficient evidence, termed Progress Myopia. It combines semantic progress reasoning with active evidence seeking, trained first with Future‑guided Reverse Generation to augment expert trajectories, and then refined via Counterfactual Contrastive Policy Optimization to reward beneficial seeking actions. Experiments on simulated benchmarks show significant gains, improving success rates by 12.7% on R2R‑CE and 7.5% on RxR‑CE, and real‑world tests demonstrate human‑like evidence‑seeking behavior.
By Zhimin Wang, Meiyuan Zhu, Duo Wu, Linjia Kang, Yajun Wang, Yuan Ni, Xiaohang Wang, Tianlu Pan, Jingyan Jiang, Yaowei Wang, Zhi Wang
SkillGym is an automatic pipeline that generates verifiable environments for training skill-use agents. It crawls internet skills, filters for reproducible workflows, and uses a builder‑reviewer process to create difficulty‑controlled tasks with reference solutions and verifiers. The system builds 6.8k environments, collects 19k successful trajectories, and fine‑tunes LLMs from 2B to 122B parameters, improving performance and skill invocation rates.
By Renxi Wang, Mingshan Hee, Fajri Koto, Timothy Baldwin, Haonan Li
The paper introduces FOCUS, a training‑free framework that compresses the interaction history of large language model agents by preserving only the past interactions that causally influence future decisions. Unlike prior methods that learn compression policies offline, FOCUS operates entirely at test time, requiring no additional data collection or fine‑tuning and can be applied to any closed‑API model. Experiments on a variety of agentic benchmarks show that FOCUS reduces peak context length by up to 48% and dependency by 73%, while improving task success by up to 8.9 percentage points.
By Shantanu Dixit, Anson Bastos, Xuchao Zhang, Chetan Bansal, Saravan Rajmohan
The paper introduces Context Language Models (CLMs), which treat context as a mutable file that the model can update freely, enabling the model to learn what information to retain. CLMs built zero‑shot from existing models outperform state‑of‑the‑art context‑management methods on several benchmarks, achieving higher accuracy with fewer FLOPs. The authors also demonstrate that CLMs can be steered via natural‑language instructions and online reinforcement learning, and they propose a suffix‑cache reuse strategy that further reduces server‑side compute.
By Rulin Shao, Shannon Zejiang Shen, Junjie Oscar Yin, Yuetai Li, Minheng Wang, Hamish Ivison, Radha Poovendran, Nathan Lambert, Teng Xiao, Mike Lewis, Wen-tau Yih, Luke Zettlemoyer, Pang Wei Koh
The paper introduces TOCOMAS, a Topology‑Coherent Multi‑Agent System that enforces topological coherence—consistent responsibility, handoff, and memory boundaries—within self‑evolving multi‑agent systems. TOCOMAS grounds task graphs in tool interfaces, groups compatible task nodes into reusable responsibility domains, and derives collaboration and memory visibility rules that respect task dependencies. In experiments on BBEH, WorkBench, SWE‑Bench‑Verified, and CoMemBench, TOCOMAS outperforms baseline methods in task success, verified progress, handoffs, and memory isolation.
By Sen Zhao, Ruiqi Kong, Zuyu Zhang, Lifeng Shen, Xinyu He, Xu Zhang, Qinghua Zhang
The paper introduces agentic meta‑reasoning, a structured inference‑time framework that explicitly manages control decisions—such as selecting partial work, restarting, or stopping—during long‑horizon agentic tasks. By delegating task execution to workers and consolidating decisions through a lightweight controller that references persistent memory, the method reduces the need to replay full histories. Experiments on ProgramBench and other benchmarks show that meta‑reasoning improves performance over direct control baselines, especially as computation budgets increase, and reveals greater reuse of earlier work and higher solution coverage.
By Paras Dahal, Anton Bakhtin, Taco Cohen, Zhengxing Chen, Carole-Jean Wu, Rob Fergus, Scott Yih, Gabriel Synnaeve, Ruslan Salakhutdinov, Sanjeev Arora, Jason Weston, Anirudh Goyal
The paper introduces a structured approach to extracting skill and responsibility level pairs from free text using the Skills Framework for the Information Age (SFIA). It evaluates five methods—including lexical baselines, retrieval‑augmented generation, and multi‑agent crews—against expert‑mapped European ICT role profiles, finding that generative strategies are more precise and that only explicit level‑prediction strategies reliably assign responsibility levels. The study also releases an automated SFIA‑9 corpus and establishes the first reproducible baseline for level‑aware skill extraction.
By Ranuga Disansa, U. S. Samarasinghe, Lasith Gunawardena
The paper investigates whether compact surrogate models can reduce the cost of simulating large language model (LLM) societies while still reproducing their collective behavior. Using 9,455 published trajectories and new opinion‑dynamics experiments, it finds that incorporating neighbor information improves individual predictions across all 16 public‑data settings and enhances pooled collective forecasts on held‑out questions, though the collective gains vary with transfer conditions. Additional tests on 24 new statements do not confirm earlier contrasting history effects, and Qwen shows benefit from history only after three observed rounds, underscoring the need for direct collective validation, explicit limits on available observations, and comparisons with simple baselines.
By Igor Itkin
The paper introduces the Agentic Commerce Bench (ACB), a benchmark for measuring fraud in AI agents that autonomously spend money. It presents a taxonomy of agentic commerce fraud, a dataset of twenty fraud classes derived from real production data, and an open‑source detector stack called gordonguard for auditing and replaying hostile counterparties. The study shows that current reasoning layers and security scanners perform poorly on many classes, highlighting the need for better detection mechanisms.
By Ankit Srivastava, Debjyoti Paul
The paper introduces StepLearn, a nonparametric framework for prequential test‑time learning in large language model agents. StepLearn separates immediate use of informative transitions from persistent trust, turning each transition into a hypothesis that guides the next step and only reusing it after prospective validation across episodes. Experiments on WebArena‑Lite and ALFWorld show StepLearn improves success rates by 2.2–12.7 percentage points over the strongest baseline, with benefits evident from the first task attempts.
By Tong Zhao, Reed Li, Yuyang Hu, Yutao Zhu, Haijin Liang, Haibo Shi, Yu Lu, Zhicheng Dou
The paper introduces RedHerring, a defense mechanism that inserts safe decoy vulnerabilities into code repositories to divert autonomous LLM agents’ verification efforts away from real security flaws. By embedding CVE-derived vulnerability chains with false bridges and providing a private certificate for quick verification, RedHerring forces agents to spend a significant portion of their limited resources on decoys. Experiments on 33 OSS‑Fuzz projects show a 38.7‑60.4% reduction in discovered real vulnerabilities, even when agents are aware of decoys.
By Kaikai Zhang, Zihan Zhang, Yuchong Xie, Zesen Liu, Shuangjie Yao, Zhixiang Zhang, Dongdong She
Mnemon is a memory agent that stores conversations as raw, dated records and uses a fast System 1 decision model (Jev) to quickly judge the relevance of records, while a slow System 2 LLM plans searches and composes answers. The agent consolidates records into topic timelines and value histories in the background, enabling efficient retrieval without rewriting conversations into structured formats. Experiments show Mnemon achieving high scores on LoCoMo and LongMemEval‑S with low context length and cost, and Jev outperforming LLMs in evidence separation and speed.
By Guangren Wang
PADM'E is a method for synthesizing preference‑aligned data to meta‑evaluate language‑model (LM) evaluators of agentic behaviors. It reframes meta‑evaluation as a preference judgment problem, generating criterion‑based data with small LMs and no human involvement. In a prototype, PADM'E produced 1,000 samples across four domains and three criteria, and human validation showed agreement with human judgment rising from 73% to 85% compared to a naive baseline.
By Cheng Chang, Yining Mao, Peng Qi
The paper introduces ReviveBench, a benchmark designed to evaluate coding agents’ ability to revive non‑running software and reconstruct industrial engines from open specifications. It comprises two families of tasks—revival (ten tasks addressing dependency issues, missing modules, legacy builds, and GPU models) and reconstruction (thirteen tasks covering numerical, geometric, hardware, and transactional systems). The benchmark uses hidden verifiers calibrated against native environments, engineering tools, or reference implementations, and the authors report that the strongest evaluated model passes all revival tasks and most reconstruction tasks, while also uncovering verifier defects that highlight measurement error in executable verification.
By Tianyu Liu, Dingyuan Dai, Yufan Du, Zhen Yang
The paper introduces Verbalization Training (VT), a technique that encourages large language models (LLMs) to openly express their evaluation awareness (EA) without directly supervising their internal beliefs. VT works by truncating model rollouts just before spontaneous verbalizations, creating training prefixes that signal awareness, and then applying a reinforcement learning objective to increase calibrated verbalization. Experiments on models such as Qwen3.6-35B-A3B, Kimi K2.6, and Inkling show that VT boosts verbalized EA by 2.4–2.9× while keeping latent EA and overall behavior largely unchanged, and a causal study confirms that VT-induced verbalizations reflect newly acquired meta‑knowledge.
By Usman Anwar, Sahar Abdelnabi, David Krueger
The paper introduces Audience‑Bound Persistent Memory, a system that tracks the audience of each memory item and enforces authorization throughout the memory lifecycle. Each item carries the audience present at recording, and derived items are partitioned or suppressed based on the intersection of source audiences, expanding only through explicit grants. The authors implement the approach in two reference architectures—a flat store and a relationship graph—and evaluate it on 10,000 multi‑party histories, showing that no forbidden items entered any context while unscoped retrieval exposed forbidden items in 82% of cases, and that entitled recall matched policy‑equivalent baselines and outperformed unscoped retrieval by 0.30 Recall@5.
By Sibo Liu
arXiv:2609.36570v1 Announce Type: cross
Abstract: Indirect prompt injection makes an LLM agent treat untrusted retrieved text as instructions. We present CounterSteer, an inference-time defense that...
By Mark Russinovich