The paper investigates how few-shot demonstrations help large language models (LLMs) build and use in-context world representations. By providing demonstrations from different worlds, the authors show that LLMs can better encode graph structures in low-dimensional hidden states, leading to improved predictions across six models from four families. The study also demonstrates that these representations shift to nearly orthogonal subspaces and that targeted interventions on them more significantly impair performance, indicating their critical role in downstream tasks such as ARC-AGI, web agent tasks, and Othello.
By Kohsei Matsutani, Gouki Minegishi, Core Francisco Park, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo
arXiv:2512. 21066v4 Announce Type: replace Abstract: Explainable artificial intelligence (XAI) reveals how explanatory variables relate to a response variable, yet communicating XAI outputs to laypersons remains difficult, limiting trust in AI-based predictions.
By Tomoaki Yamaguchi, Yutong Zhou, Masahiro Ryo, Keisuke Katsura
ChemVTS-Bench is a domain-authentic benchmark that evaluates Visual‑Textual‑Symbolic reasoning in multimodal large language models for chemistry. It presents diverse chemical problems—organic molecules, inorganic materials, and 3D crystal structures—in three input modes: visual-only, visual‑text hybrid, and SMILES-based symbolic. The benchmark includes an automated agent workflow for inference, answer verification, and failure diagnosis, and shows that visual-only inputs and structural chemistry remain challenging for current models.
By Zhiyuan Huang, Baichuan Yang, Zikun He, Yanhong Wu, Fang Hongyu, Zhenhe Liu, Lin Dongsheng, Bing Su
ToolCUA is an end‑to‑end agent that learns to optimally orchestrate GUI actions and tool calls for Computer Use Agents. It introduces a staged training pipeline that first scales interleaved GUI‑tool trajectories from static GUI data, then bootstraps decision making with single‑turn reinforcement learning, and finally refines performance via online agentic RL guided by a tool‑efficient path reward. On the OSWorld‑MCP benchmark, ToolCUA achieves 46.85% accuracy, a 66% relative improvement over the baseline and a 3.9% gain over GUI‑only models, setting a new state of the art for comparable‑scale models.
By Xuhao Hu, Xi Zhang, Haiyang Xu, Kyle Qiao, Jingyi Yang, Xuanjing Huang, Jing Shao, Ming Yan, Jieping Ye
PTCG-Bench is a new benchmark that uses the Pokémon Trading Card Game to evaluate large language model (LLM) agents on two fronts: their decision‑making within a single complex game environment and their capacity to evolve through accumulated experience. The benchmark includes a modular harness ablation to isolate agent performance from model capability. Experiments show that while LLM agents can achieve non‑trivial gameplay, sustained self‑evolution remains difficult and performance depends on harness design.
By Dongdong Hua, Yifei Sun, Renhong Huang, Feng Gao, Chunping Wang, Yang Yang
The paper introduces NavSecure, a vision‑based navigation framework that combines world models with safety‑focused decision making for autonomous vehicles operating over 5G networks. By predicting potential threats and generating safer routes, NavSecure reduces reliance on extensive real‑world trial‑and‑error learning and enables continuous autonomous adaptation. Experiments in simulation‑to‑real driving scenarios demonstrate superior performance in collision prevention and risk reduction compared to other end‑to‑end methods.
By Hong Ding, Ziming Wang, Yi Ding, Hongjie Lin, SuYang Xi, Chia Chao Kang
MAGIC introduces a dense‑reward reinforcement learning framework for generating mixed‑granularity agent graphs in large‑language‑model based multi‑agent systems. The method sequentially selects functional roles, instantiates them as either single agents or reusable groups, and connects them to existing units, optimizing the construction policy with intermediate feedback from probe‑based utility and structural signals. Experiments show MAGIC outperforms state‑of‑the‑art baselines on eight benchmarks and achieves strong inference efficiency.
By Kairui Yang, Ziheng Yi, Xunkai Li, Minghao An, Zhanke Liu, Zekai Chen, Rong-Hua Li
The paper investigates how machine‑learning models can determine whether a claim (a test assertion) remains valid after a code change. It compares two questioning strategies: asking whether a diff preserves behavior versus asking whether a specific claim still holds. The authors find that the latter approach yields far higher precision (up to 0.974) across models of varying cost, while the former performs poorly (precision 0.291–0.329). They also benchmark against a regression‑test selector, showing that even near‑complete knowledge of a change’s reach does not reliably identify falsified claims. The study is grounded in 10,369 mined claims with 184 execution‑verified flips from 23 Python libraries.
By Atul Anand
ToolCompass is a post‑training framework that improves how large language model agents explore new tools by organizing tool‑call representations according to shared functions. It models each function class as a von Mises–Fisher distribution, reducing variation within a function while increasing separation between different functions, thereby guiding exploration toward functionally similar unseen tools. Experiments on AppWorld and FTRL show consistent gains, with up to a 10.71‑percentage‑point improvement in out‑of‑distribution task success over vanilla post‑training and outperforming competitive baselines.
By Junlin Fang, Chong Zhang, Do Nguyen-Thanh, Xiaogang Xu, Zhen Fang, Sean Du
The paper introduces AIDE^2, an AI research agent that recursively improves its own code by proposing, benchmarking, and selecting modifications. Over an eight‑day autonomous run, it achieved seven successive improvements—including new search policies and memory mechanisms—that transferred to four held‑out benchmarks in machine learning, algorithm engineering, and weather forecasting. The agent’s best version matched or outperformed a top human‑engineered production research agent and also reduced reward‑hacking rates, despite never optimizing for that metric.
By Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, Zhengyao Jiang
CliffCompaction is an autocompaction technique that reduces cost by up to 50% while maintaining or improving performance on benchmarks such as Terminal‑Bench and KernelBench. It achieves this by truncating or dropping content without rephrasing, ensuring compacted information remains faithful and preventing context drift. The method enables efficient test‑time scaling, matching or surpassing higher‑cost models like Opus 4.7 and GPT‑5.3 Codex, and delivers significant CUDA kernel speedups on KernelBench.
By Trang Nguyen, Eulrang Cho, Bingqing Chen, Tim Dettmers
The paper introduces ELEMENT, a framework that combines episodic and lifelong entropy maximization to drive reward-free exploration in reinforcement learning. It addresses two key limitations of existing entropy-based methods: the vanishing intrinsic reward after a state is visited and the computational cost of estimating entropy over large datasets. ELEMENT achieves this by deriving an average episodic state entropy reward and employing a k‑NN graph‑based estimator for lifelong entropy, leading to superior state coverage and unsupervised pre‑training performance compared to current baselines.
By Hongming Li, Zhao Yang, Xiaoxuan Liang, Shujian Yu, Jose C. Principe
The paper addresses the lack of system‑level confidence estimates in multiagent language model systems such as collaborative reasoning and debate. It introduces confidence composition methods, including confidence‑aware routing and log‑odds pooling, to combine agent confidences while maintaining selective utility and probabilistic reliability. Experiments on five benchmarks with diverse model pairs show that gated‑fusion techniques improve AUARC and Brier scores compared to single‑agent and standard debate baselines, and a shared dependence discount further enhances reliability.
By Ali Elahi, Michael J. Curry, Barbara Di Eugenio
Parallel coding agents can produce patches that work individually but fail when merged, a problem arising when one agent alters an interface or rule another agent depends on. The authors introduce *stale*, a benchmark for semantic coordination, evaluating patches both alone and in combination across synthetic tasks, real Django pull‑request pairs, and constructed tasks. While only one of 417 mined Django pairs showed interference after grading corrections, interference appeared in 97% of runs on constructed tasks, and a message describing the concurrent change recovered 82% of those runs.
By Haocheng Xia, Eugene Wu, Yongjoo Park
Qwen3.8-Omni-Flash is a natively multimodal agentic model designed for real‑world multimodal productivity, offering enhanced multimodal understanding, reasoning, and long‑horizon agentic task performance. It builds on a sparse mixture‑of‑experts architecture, extends its context window to one million tokens, and supports long‑context multimodal reasoning and planning. The release includes Qwen-MM-Plugins for native audio and video support and Qwen-Live-Harness for building responsive, real‑time multimodal agents, with extensive evaluations confirming strong performance across multimodal tasks.
By Qwen Team
Agensh is a new multi‑agent harness that eliminates a central orchestrator by letting workers self‑organize through a continuous cooperation loop. The system uses a shared workspace, message interface, and shared context to coordinate tasks, verify results, and merge progress asynchronously. Experiments on ProgramBench and pandoc show that scaling from 1 to 1,024 agents improves test‑pass rates by up to 49% relative, demonstrating that agent count is a viable scaling dimension for complex tasks.
By Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu Wei
The article surveys how large language models (LLMs) are being applied to mental health, outlining a three‑phase evolution: Phase I uses LLMs as passive information tools and pattern recognizers for assessment; Phase II employs them as empathetic conversationalists for stateless, in‑the‑moment interactions; Phase III aims to create longitudinal, personalized companions that act as stateful cognitive agents. It systematically reviews core technologies, agent architectures (Profile, Memory, Reasoning, Planning), and the datasets and benchmarks that support this progression, offering a coherent narrative and roadmap for future research. The survey also provides a curated resource list at https://github.com/Emo-gml/Awesome-Mental-Health-LLMs.
By He Hu, Yucheng Zhou, Qianning Wang, Yingjian Zou, Chiyuan Ma, Juzheng Si, Jianzhuang Liu, Zitong Yu, Laizhong Cui, Fei Ma, Qi Tian
WatchPoint is a simulated‑user system that generates and runs diagnostic scripts against a live web application, producing structured observations to guide coding model retries. Unlike prior methods that rely on screenshots or non‑executable metrics, WatchPoint operates on Web‑Bench—a benchmark of 50 multi‑file web projects with 1,000 sequential tasks verified by deterministic end‑to‑end tests. It recovers 57.6% of diagnosed tasks, matching a human tester’s 54.5% recovery rate, and identifies when such simulated feedback is beneficial or should be withheld.
By Guanqun Yang, Wei Yang, Xueqing Liu
Hy‑MultiTurn is a Chinese benchmark designed to evaluate deep multi‑turn dialogue understanding over long interactions. It introduces six controlled evaluation modes—constraint memory, precise execution, constraint synthesis, object localization, action suppression, and reference resolution—across 209 tasks ranging from 12 to 76 turns, incorporating dialogue length, irrelevant distractions, and colloquial phrasing. Testing 22 state‑of‑the‑art models shows the benchmark is highly challenging, with even the best model meeting all criteria only 41.1% of the time and no model excelling in every mode.
By Eileen Ye, Jiawen Tao, Yaoming Li, Chenxu Liu, Wenhan Yu, Yaxin Fan, Xiaokun Yuan, Mengzhou Wu, Yanbing Jiang, Maxm Pan
The paper presents a foundation-guided auto‑annotation pipeline that improves standard autonomous driving object detectors in adverse weather. By benchmarking YOLOv8, Co‑DETR, and SAM3 on a custom dataset of 25 operational scenarios, the authors find SAM3 to be the most robust and use it offline to generate pseudo‑labels. Fine‑tuning YOLOv8 on these labels boosts overall mAP by 16.04% and yields significant gains in specific conditions such as Residential Direct Sunlight (32.73%) and Highway Fog (28.65%).
By Sepideh Gohari, Goodarz Mehr, Azim Eskandarian