The paper titled "Cheap, open agents make LLM pollution harder to mitigate" reports that open-weight language models combined with open-source agentic frameworks can produce synthetic survey responses that are competitive with commercial agents and harder to detect. The authors compared nine agent configurations, finding that fully open agents run locally without usage fees and that no single detection check reliably identifies all agents. Open-text responses were the most effective at distinguishing agents from humans, highlighting the need for multilayered detection strategies.
By Raluca Rilla, Anne-Marie Nussberger, Rui Mata, Dirk U. Wulff
Scepsy is a serving system designed to efficiently schedule arbitrary multi‑LLM agentic workflows on GPU clusters. It leverages the observation that each LLM’s share of execution time remains relatively stable across requests, profiling LLMs under various parallelism levels to build an Aggregate LLM Pipeline that predicts throughput and latency. Using this predictor, Scepsy searches for optimal GPU allocations—balancing fractional GPU shares, tensor parallelism, and replica counts—and then heuristically places them on the cluster to reduce fragmentation and honor network topology, achieving up to 2.5× higher throughput and 1.0–3.3× lower latency compared to baseline approaches.
By Otto White, Marcel Wagenl\"ander, Britannio Jarrett, Xijin Zhao, Yanda Tao, Pedro Silvestre, Guo Li, Huanzhou Zhu, Llu\'is Vilanova, Peter Pietzuch
The paper discusses the "authorship hazard" in agentic dataspaces, where AI agents can draft and publish governance artifacts without human approval. It argues that agents should be treated as subjects of governance, not authors, and proposes mechanisms to enforce this, such as closing the authorization channel and treating drafting as an enforcement problem. Experiments show that unapproved agent drafts can reverse many authorization decisions, highlighting the need for robust review processes.
By Seungho Lee, Changbin Lee
The paper introduces the concept of LLM Parkinsonism, describing how large language models can persist in low‑value actions after completing their objectives. It proposes a Global Executive Control (GEC) architecture that separates action generation from project‑level oversight, achieving comparable success to candidate‑set control while significantly reducing token usage and complexity. Experimental results on a 24,000‑episode benchmark show GEC cuts mean token use by 36.4% and limits token consumption at the 40,000‑token ceiling by 18.7%, eliminating pre‑completion drift.
By Dongsheng Xiao, Zeyuan Wang, Xuzhe Xia, Bo Zhao, Yankai Cao
SlideLab is a training‑free, multi‑agent framework that generates scientific presentations directly from research papers. It first plans a coherent narrative, then iteratively builds and refines a shared slide deck using agents for content planning, visual generation, layout refinement, and grounding verification. In a blind human preference study, SlideLab outperformed both open‑source and commercial systems on 77% of papers while using about four times fewer inference tokens than the strongest open‑source baseline. The authors also introduce ConfArena, an audience‑oriented evaluation framework that simulates a conference room and assesses presentations slide by slide, matching human system rankings and detecting issues such as falsified numbers, degraded figures, dropped slides, and shuffled slide order.
By Vidushee Vats, Karun Sharma, Yuxia Wang
ToolSearcher is a reinforcement learning framework designed to improve large‑scale tool selection for large language models. It introduces category‑constrained discrimination, event‑level search modeling, and trajectory‑aligned credit allocation to better distinguish similar tools, optimize multi‑turn search, and provide fine‑grained rewards. Experiments on large‑scale benchmarks show that ToolSearcher outperforms strong baselines in iterative search and complex tool composition scenarios.
By Zhenlong Dai, Xujie Song, Zitong Wang, Tong Niu, Jian liu, Weiqiang Wang, Xiu Tang, Sai Wu, Chang Yao, Jingyuan Chen
The paper introduces the concept of Evolutionary Safety for recursive self-improving AI, focusing on how safety properties evolve as an AI system and its successors change. It identifies key risks such as intent drift, error accumulation, and safety-property erosion, and presents a taxonomy covering agent state, model state, evaluation, environment, and update mechanisms. The authors propose methods for discovering and evaluating evolutionary risks, and outline governance principles for modification, selection, authorization, provenance, and recovery, while highlighting open problems for maintaining safety in persistent, adaptive, and recursively self-improving systems.
By Chang Gong, Jingping Bi, Di Yao, Xinjian Liang, Chao Xiang, Ruijie Guo
The paper introduces skill cascading attacks, where a malicious goal is spread across multiple seemingly benign skills, causing harmful outcomes when combined. It presents SkillCascade, an automated red‑teaming framework, and releases SkillCascade‑Bench, a benchmark of 213 validated cascading test cases across various agent systems and domains. Experiments show that these cascaded interactions reliably induce harmful behaviors while evading existing per‑skill scanners and runtime monitors, revealing a gap between component‑level integrity and system‑level safety.
By Zihao Zhu, Siwei Lyu, Adel Bibi, Baoyuan Wu
The paper investigates cost-inefficient behaviors in coding agents, analyzing 1,200 trajectories from Claude Code and Mini-SWE-Agent on SWE-bench Verified. It identifies three main inefficiencies—subsumed retrieval, similar script generation, and test re-execution—that affect 79–98% of tasks and contribute up to 22.75% of costs. The study evaluates mitigation strategies, finding that developer-designed skills reduce costs by up to 41.73%, outperforming other approaches.
By Yiran Hu, Nan Jiang, Shanchao Liang, Anik Dey, Yi Wu, Lin Tan
Self-evolving agents can continually improve their behavior, while tools define the executable action space through which they interact with the environment. However, exposing the full tool library to...
LLM-based agents have shown strong capabilities in automated data analysis and are increasingly moving toward long-horizon, multi-stage analytical workflows. However, as the analytical process evolves...
As LLM agents increasingly rely on persistent memory for long-horizon and personalized behavior, they can retain and reuse information across interactions, but this also creates a lasting channel thro...
Simon Willison delivered a closing keynote at the WeAreDevelopers World Congress North America, outlining the key developments in large language models for 2026. He highlighted the release of Claude Opus 4.5 and GPT‑5.1 in November 2025, noting that their coding agents became reliable enough for daily use. Willison also discussed early benchmarks, such as generating an SVG of a pelican riding a bicycle, and mentioned a new GitHub repository called "Warelay" that emerged during the year.
The article argues that well‑designed architecture can inadvertently eliminate signals that tooling relies on, turning a structural issue into a search problem. It highlights how drawing boundaries in systems can strip away essential cues needed by agents. The piece emphasizes the importance of considering signal preservation when designing architecture.
By Yonatan Sason
The article discusses Meta’s new AI system, Muse, which offers each user a persistent Linux VM in the cloud and is marketed as an easy-to-use, consumer‑friendly agentic AI. It highlights the system’s technical innovation and the company’s packaging as a cute mascot, while noting that users may not fully grasp the power and potential danger of such an advanced tool, especially when running it locally on a Mac.
The article explains that Retrieval-Augmented Generation (RAG) is a retrieval system, while agents are responsible for action. The author built a distinct layer that explicitly connects retrieval to action, and tested this setup across nine tasks alongside standalone RAG and agent systems.
By Emmimal P Alexander
MeshHeal is a fully decentralized self‑healing framework for decentralized LLM‑based multi‑agent systems that addresses gray failures—situations where an agent remains responsive but its task‑solving quality degrades. It operates on two timescales: a fast adaptive hierarchy that escalates uncertain or low‑scoring outputs to committee review and correction, and a slow peer‑relative detector that aggregates scores to distinguish persistent degradation from normal variation, triggering mandatory review and eventual exclusion of degraded agents while allowing recovered agents to rejoin. MeshHeal’s evaluation, using Model‑Backed MAS Evaluation, shows it achieves higher degraded‑phase accuracy (0.839) on BBH, MATH, and MMLU‑Pro with fewer tokens per task compared to the baseline Symphony.
By Keru Chen, Sen Lin, Yingbin Liang, Nathaniel D. Bastian, Shaofeng Zou
The study evaluates how large language model agents maintain consistency over extended interactions by simulating a 20‑step delayed‑gratification task. Researchers ran 84,540 trajectories across eight model families, using survival analysis to track when agents first claim a reward and discrete‑time hazard regression to assess how factors like social visibility, persona stressors, and deliberation policy affect failure risk. They also developed a seven‑category taxonomy from 13,780 deliberation traces, revealing that early failures are impulse‑driven, later ones are fatigue‑ or cost‑benefit‑framed, and public settings elicit norm‑oriented justifications; longer deliberation correlates with higher intra‑rationale contradictions, challenging assumptions about reasoning depth and consistency.
By Igor Bogdanov, Olga Manakina, Chung-Horng Lung
GRASP is a multi-stage planning framework that improves the reliability of large language models on complex tasks. It separates planning into three specialized modules—GenPlan for global macro-guidelines, RevPlan for exploring localized strategies, and VerPlan for multi-criteria evaluation—allowing context isolation and strict macro-regularization. Experiments show GRASP outperforms direct LLM planners by significant margins on datasets such as Natural Plan Calendar Scheduling, ZebraLogic, and SciBench Math, and it mitigates performance collapse in multi-task and dual-task settings.
By Arunabh Srivastava (Amir), Mohammad A. (Amir), Khojastepour, Srimat Chakradhar, Sennur Ulukus
The paper investigates how well large language models (LLMs) can handle character attacks—ad hominem arguments—in political debates. By analyzing natural political dialogues and comparing LLM-generated responses to a corpus of U.S. presidential debates, the study finds that most LLMs favor logical defenses and rarely use ethos-based counterattacks. The authors suggest that safety fine‑tuning limits LLMs’ strategic options, preventing them from fully engaging in realistic political discourse.
By Ewelina Gajewska, Katarzyna Budzynska, Jaroslaw Chudziak