The paper introduces NaviHydra, a trajectory‑scoring planner that addresses a navigation compliance gap in end‑to‑end autonomous driving. It proposes a new binary Navigation Compliance metric (NAVI) and a Controllability Measure (CM), and releases the NavControl dataset with 14,918 intersection scenarios and over 34,000 direction samples. NaviHydra achieves high performance on NAVSIM and NavControl benchmarks, and training with NavControl improves controllability across various architectures.
By Hanfeng Wu, Marlon Steiner, Michael Schmidt, Alvaro Marcos-Ramiro, Christoph Stiller
PerMIT is a memory architecture that provides provable integrity guarantees for persistent memory used by large language model agents. It limits trust to user‑typed inputs and values confirmed at a gate, mediating untrusted environment‑derived values through a finite interface while preserving multi‑turn planning and tool choice. Two machine‑checked integrity theorems in Lean 4 ensure that only gate‑confirmed data can influence sensitive actions, and community endorsements can accumulate to reduce verification effort as the user base grows.
By Tanqiu Jiang, Min Bai, Sadat Shahriar, Wanyu Du, Sandesh Swamy
The paper introduces A-B-D, a method that infers personality traits of large language model agents from their behavioral data, combining functional actions and linguistic outputs across 345,667 real-world trajectories of 80 models. After filtering for stability and discriminability, 79 features are reduced to six reliable factors—two functional and four linguistic—revealing differences such as Claude‑Sonnet‑4.6’s planfulness versus GPT‑5.5’s low energy. The study also demonstrates a weak correlation between these inferred traits and agents’ self-reported Big Five scores, highlighting a knowledge‑action gap.
By Haokai Zhao, Jie Gao, Yunze Xiao, Xintao Wang, Weihao Xuan, Aditya Joshi, Mark Dredze, Jen-tse Huang
PAUSE is a steering framework for coding agents that intervenes once before the agent’s first source‑code edit, eliminating the need for learned monitors, checkpoint searches, or extra model calls. By having a Planner generate a repair plan from the task and pre‑edit transcript, the same agent can resume execution with this plan, preserving context and workspace. Experiments on SWE‑bench Verified and Multilingual show that PAUSE improves resolution rates by up to 16.8% online and recovers 35.4% of failed tasks offline, while reducing Executor token usage by 33‑41% compared to restarts.
By Viral Patel, Boyin Tan, Jason Xue, Youcheng Sun
The paper introduces a closed‑loop framework to evaluate frontier AI models on the LIBERO‑10 robot manipulation benchmark, allowing models to control a robot via visual inputs, proprioception, and tool calls without task‑specific fine‑tuning. Six models (GPT‑6 Astra, GPT‑6.1 Sol, GPT‑6 Luna, Claude Opus 5, Claude Opus 5.5, Gemini‑4 Argon) were tested across 18 configurations and 1,800 episodes, with GPT‑6 Astra achieving the highest success rate of 44%. Results show that higher reasoning effort generally improves performance, yet benefits vary by model and task, and many models struggle with precise spatial arrangements and complex multi‑step interactions.
By May Espinola, Qifan Zhang, Yu Xiang
AdaGuard is an adaptive guard model that evaluates language‑model agent trajectories against user‑defined policies. The authors created the AdaptiveSafety dataset, containing 10,939 training and 1,000 test examples with policy rules, counterfactuals, explanations, and violated rules. Using supervised pre‑training followed by the SafePO reinforcement‑learning algorithm, they built 0.6B, 4B, and 8B guard models, with the 4B version achieving 89.30% accuracy on AdaptiveSafety and 71.82% on DynaBench.
By Yunhao Feng, Yifan Ding, Yuxiang Xie, Zheng Li, Mingrui Lao, Zeyuan Wang, Yanming Guo
The paper introduces Conservative Meta-Agent Search (CMAS), a method for adapting agent architectures in non‑stationary environments. CMAS retrieves archived designs, proposes typed graph edits, and evaluates them through paired shadow replay with variance‑penalized scoring and a simulated canary stage before promotion. Using the MetaInvest‑Bench simulator, CMAS reduces surrogate dynamic‑oracle regret and improves held‑out utility across 96 financial streams, while maintaining low false‑promotion rates and modest inference cost.
By Ali Atiah Alzahrani
METTA: Memory-Enhanced Test-Time Adaptation for Meta-Reinforcement Learning proposes a new approach that splits fast adaptation into two temporal scales: a causal neural memory for fine-grained within-episode adaptation with linear complexity, and compact episode-level summaries for long-range cross-episode inference. This design reduces the quadratic memory and computation costs of transformer-based agents while retaining expressive capacity for test-time training. Experiments on unseen Meta-World and PointMaze tasks show METTA achieves over 50% success on Meta-World and outperforms transformer baselines on long-context sparse-reward environments.
By Ali Najar, Hosein Hasani, Amir Kooshan Fattah, Yasamin Rajabi, Mahdieh Soleymani Baghshah
EpiSEM-Bench is an executable benchmark designed to assess how language‑model agents handle missing safety‑relevant evidence. It comprises 300 matched unsafe/safe task pairs across six domains, systematically suppressing critical facts to measure unsafe execution and failure cardinality. An independent audit tool evaluates whether agents seek missing facts and withhold prohibited actions while completing permissible tasks, revealing variation among ten models in omission robustness, evidence acquisition, and appropriate action.
By Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu, Zhongxiang Dai, Shuang Qiu
ULME is a post‑training framework that unifies text embedding and language modeling within a single causal decoder‑only LLM by inserting learnable slots for semantic and reasoning embeddings. It trains three objectives—contrastive loss for semantic alignment, language‑modeling loss for answer generation, and a stop‑gradient predictive loss linking reasoning to the answer—so that generation actively shapes the embedding. Across Qwen2.5 and Qwen3 models (0.5B–8B), ULME improves MTEB scores by 3.3–4.1 points, tightens BRIGHT and AdvBench‑IR, and matches generation‑only fine‑tunes on standard accuracy benchmarks.
By Yang Liu, Zhiqiang Yuan, Shu-Dong Huang, Jiancheng Lv, Liwei Yang, Qiwei Yan, Kailin Lv, Jinchao Zhang
AIRO is a streaming-based machine learning framework designed to detect and mitigate cyber threats in Industrial Internet of Things (IIoT) environments. It integrates Apache Kafka for real-time data ingestion, attack-specific feature extraction, and ML inference using models such as Gradient Boosting and LightGBM. The system employs confidence-based decision logic to trigger risk reduction actions like IP blocking, account suspension, and traffic limiting, achieving F1-scores up to 0.9229 on specialized datasets for C2 beaconing, reconnaissance, SQL injection, and phishing.
By Raima Deb, Pranav M Pawar, Mithun Mukherjee, Raja Muthalagu, Mian Guo
PROACT-Agent is a framework designed to provide proactive runtime oversight for large language model agents, aiming to prevent irreversible environmental harm by enabling real‑time guardrails. It introduces three key components: Progressive Trajectory Unrolling, Reasoning‑Augmented Causal Rectification, and Culturally‑Aware Data Localization, which together address safety drift and enforce temporal consistency. The accompanying PROACT‑Bench, a bilingual safety benchmark with 155,780 labeled states, demonstrates the framework’s effectiveness, achieving high unsafe‑class F1 scores and significantly reducing attack success rates in AgentDojo.
By Ding Jia, Wei Liu, Xianglong Du, Yingjie Li, Yingqing Yang, Huili Yu, Zhangsong Zhan, Chu Zhou
The paper investigates whether diffusion language model agents can exchange hidden messages via their continuous denoising states. By compressing a sender’s state into four dense vectors, the receiver can incorporate this information while preserving its own context. Experiments on tasks such as Tiny-A2D, Dream-7B, and QASC show that removing or substituting the message significantly degrades performance, and that a single late transfer can outperform multiple exchanges.
By Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov, Ilya Makarov
UserSWE is a new benchmark that evaluates coding agents’ ability to fix repository-level bugs through multi‑turn conversations with dynamic users, providing execution‑grounded verdicts after each turn. The benchmark transforms repository histories and existing single‑turn datasets into multi‑turn tasks, using a persona chain derived from 20,859 development sessions to simulate user interactions. Eleven agents were tested on 72 tasks from 42 repositories, revealing that agent rankings shift with the interaction budget and that single‑turn performance alone does not capture interactive coding capability.
By Zhan-Yu Wu, Ji Luo, Alyssa Wu, Junjielong Xu, Richeng Xuan, Zhichao Hu, Chongming Gao, Xinting Hu
arXiv:2610. 30072v1 Announce Type: new Abstract: Data agents built on ReAct-style loops now execute complex analytical workflows over structured data, yet their intermediate reasoning still runs almost entirely through text.
By Xinzhe Wang, Xudong Weng, Zang Li, Shaolei Zhang
The paper introduces MixAgent, an agentic pipeline for music post‑production that emulates audio engineers’ workflow to create interpretable and editable mixing plans guided by a reference song. It employs a grouping agent to cluster related tracks, a planning agent to select and order audio effects, and a detector‑critic pair to refine the mix by identifying needed corrections and proposing parameter adjustments. The authors also present MixCritic, a dataset of 45,657 repair critiques, and demonstrate that supervised fine‑tuning followed by preference optimization improves mix quality, achieving a 19% error reduction on MedleyDB and outperforming several large language models in human listening tests.
By Jaywon Koo, Junghyun Koo, Marco A. Martinez-Ramirez, Wei-Hsiang Liao, Seungheon Doh, Qiyu Wu, Vicente Ordonez, Yuki Mitsufuji
The paper reports how an AI-driven system, AI-greement, was used to discover a new consensus protocol called A‑agreement for geo‑replicated state machines. A‑agreement improves mean latency over existing baselines by 12.2% on write‑heavy and 33.3% on read‑heavy workloads while matching performance on read‑only workloads, using lease‑authorized local reads, direct client learning for leader‑ordered writes, and a fast fixed‑quorum per‑key write path. The AI‑agreement harness integrates a protocol knowledge base, a modular implementation skeleton, and a Lean verification framework, revealing vulnerabilities in existing protocols and demonstrating how domain‑specific harnesses can transform general‑purpose coding agents into effective tools for distributed computing research.
By Cl\'ement Burgelin, Rachid Guerraoui, Zicong Hong, Yinan Hu, Anne-Marie Kermarrec, Rafael Pires, Rasmus Moorits Veski, Martijn de Vos
The paper proposes that quantum entanglement underlies the synchronization of millions of neurons, introducing the Liu‑Chen‑Ao (LCA) hypothesis of entangled nerve fibers. Using a nonlocal many‑body wave equation, the authors model collective behaviors among adaptive agents and apply this framework to Chinese intraday stock market data, finding that adaptive entangled game modes explain 82‑94% of decision patterns and that only a small fraction of actions are truly independent. These empirical results support the LCA hypothesis and suggest that incorporating probability‑wave‑based entangled modules could improve artificial general intelligence by creating more efficient, human‑like processing units.
whyItMatters":"The study provides empirical evidence for the LCA hypothesis and highlights the limitations of conventional artificial neural networks, pointing toward more efficient and robust AGI architectures."
By Haochen Li, Xinshuai Guo, Jingdong Ouyang, Wei Zhang, Leilei Shi
CrashGym is a deterministic, replayable limit‑order‑book simulator that lets language‑model agents decide between a deleveraging intervention and a liquidity backstop before a market crash. Agents can query intervention experiments in a laboratory version of the market, and their choices are scored by simulating every feasible intervention on an independently seeded target realization, allowing direct regret calculation. In a study of 24 new worlds, laboratory access reduced mean regret by 57.4% for one deployment, and exploratory results suggest lower regret across all evaluated deployments, while an evolving strategy mix offers no clear advantage over a fixed mix.
By Haocheng Lu, Haoxuan Cheng, Mingyang Chen, Minjun Zhu
The paper introduces ReMAS, a reinforcement learning-based framework for black-box red-teaming of multi-agent systems (MAS) that fine-tunes attacker LLMs to craft adversarial prompts. ReMAS operates in two stages: a rewrite stage that refines base attack prompts to boost extraction success, and a template generation stage that builds attack templates to better invoke specific sub-agents and expose their system prompts. Experiments demonstrate that ReMAS significantly outperforms existing black-box attacks, showing high success rates and transferability across different backend LLMs and MAS structures.
By Yuqi Jia, Hangsheng Zhang, Chenxi Liu, Junjie Liang, Xun Chen