The paper introduces Clinical System One (CS1), a 4‑billion‑parameter decision model that rapidly produces typed judgments from clinical records. Trained on 12.4 million Q&A pairs from 3.6 million patients, CS1 can answer multiple questions in a single forward pass, providing probability estimates for uncertain cases. In evaluations on 963,408 held‑out patients and 21 external datasets, CS1 outperformed larger open models and, when paired with a language‑model agent, efficiently identified trial‑eligible patients with far fewer tokens.
By Andrew Zhang, Trinh T. L. Vuong, Sophia J. Wagner, Dongmin Bang, Tong Ding, Yuchang Su, Mary Cipperman, Christina S. Chen, Judith Lefkes, Ramin Kahidi, Ming Y. Lu, Caiwei Tian, Tianhao Luo, Jingnan Ma, Rowland Pettit, Long Phi Le, Faisal Mahmood
EDMA: Entropy-Driven Multimodal Answering introduces a framework that formalises multimodal question answering as an iterative entropy‑reduction process. By quantifying each modality’s contribution through logical entropy, EDMA constructs measurable answering trajectories and employs entropy‑driven fusion and selection to minimise logical entropy. Across standard MMQA benchmarks, EDMA surpasses prompting baselines, achieving significant gains in F1 and LLM‑J scores, and demonstrates that higher entropy correlates with more false positives, validating entropy as a reliable performance proxy.
By Emanuele Mezzi, Gertjan Burghouts, Fabio Massacci, Mengyuan Zhang
Magnolia is a persistent‑memory AI agent framework designed for computational chemistry. It integrates protocols such as docking, quantum chemistry, and molecular dynamics with lab notebooks and LLM‑augmented chat logs, using a three‑tier memory system to capture, distill, and promote learnings into project‑tier entries and cross‑project rules. In four case studies, Magnolia tracked 317 runs, produced 116 project‑tier entries, and promoted four cross‑project rules, demonstrating that runs with memory passed all checks while memory‑less runs failed.
By Tao Jiang, Benjamin Navet, Yousouf Taghzouti, Fabien Gandon, Louis-F\'elix Nothias, \'Elise Dumont
ProMemAgent is a proactive agent that uses prospective memory to handle delayed user intentions over long horizons. It organizes past sensory perceptions into context-aware memory, generates prospective cues, and links earlier events to future developments for retrieval. The authors also introduce CPMBench, a benchmark of 1,000 instances across eight everyday categories, and show that ProMemAgent outperforms baselines in memory retrieval and proactive assistance tasks.
By Lilin Xu, Linyang He, Kaiyuan Hou, Yuang Fan, Minghui Zhao, Xuhai Xu, Xiaofan Jiang
The paper investigates how large language model (LLM) agents can improve performance by storing and reusing experience in external memory. It identifies a problem called error reproduction, where mistakes in stored abstractions propagate through cycles of abstraction, execution, and re-abstraction, leading to persistent errors. To mitigate this, the authors propose Anti-Persistence, a memory design that retains raw trajectories and outcomes but reconstructs task-specific guidance on demand, discarding it after use. Experiments across multiple benchmarks show that Anti-Persistence yields higher held‑out success rates and lower LLM‑call costs compared to other memory designs, supporting the idea of separating experience accumulation from abstraction persistence.
By Ruhan Wang, Kishan Panaganti, Dongruo Zhou
M-Plan is a closed‑loop multi‑agent system that combines long‑term and short‑term memory to improve large language model agents on long‑horizon, multi‑turn tasks. It uses task‑specific milestones and local hints to generate global‑local guidance, forming short‑term memory during execution, and then optimizes and stores relevant information in long‑term memory for future use. Experiments on ALFWorld, Webshop, and PDDL show that M‑Plan outperforms baseline methods in both accuracy and efficiency.
By Ziyang Liu, Jiakang Yuan, Jiayuan Fan, Tao Chen
The paper investigates how optimizing the harness of language model agents—without changing model weights—affects performance. It distinguishes between gains that improve success on tasks already within reach (elicitation) and gains that broaden the range of solvable tasks (expansion). By comparing optimized harnesses to their pre‑optimization versions across static and dynamic tasks, the study finds that benefits can reverse at higher attempt budgets and that aggregate gains may mask shortcomings on harder tasks, highlighting the need for detailed diagnostics.
By Younghyun Kim, Wooseok Seo, Minseo Kim, Seunghyuk Oh, Minjae Lee, Donghoon Kim, Hyung Il Koo, Jaehyung Kim, Wonjun Kang
LatticeMind is a validation‑gated, agentic framework that translates natural‑language requests into executable density functional theory (DFT) workflows. It uses workflow contracts to specify required stages, files, and dependencies, deterministic validators to check projects before solver launch, and bounded repair to fix failures, ensuring traceable reports linked to project files. On benchmarks spanning 44 RESCU and 19 VASP cases, LatticeMind reproduced all reference values, with 119 of 132 runs succeeding on the first validated plan and the rest after a single repair cycle, outperforming a conventional trial‑and‑repair baseline.
whyItMatters":"The framework demonstrates that incorporating pre‑execution validation and repair can reliably prevent costly errors in multistage DFT workflows, making AI‑assisted scientific computing more trustworthy and efficient."
By Nathaniel M. Vegh, Yu Zhu, Rapha\"el J. Prentki, F\'elix Beaudoin, Shuaishuai Yuan, Hong Guo
MemT-1.0 is a memory-native evolution architecture that embeds memory directly into model computation, using Gated DeltaNet-2 for token-level updates and a Mixture-of-Storage scheme to scale capacity. It is trained in two stages with 34B tokens and released as MemT-1.0-Nano and MemT-1.0-Flash, which outperform parametric memory baselines and compete with retrieval-augmented methods while preserving general capabilities. At a 128K-token history, MemT-1.0-Nano achieves a 1.86× inference speedup and a 66× reduction in persistent storage compared to a full context KV cache.
By Zhiyu Li, Zeyu Zhang, Linyun Liu, Yiwei Bu, Chenyang Xi, Xichong Zhang, Yi Yu, Yu Yu, Zhichen Liu, Yuxiao Yi, Chunyu Li, Haoran Liu, Zhenkai Ding, Le Zhou, Juncheng Zhang, Boyu Chen, Zehao Lin, Guorui Li, Renyu Fu, Ji Li, Youyang Xia, Jiajun Shen, Harvey Li, Jingsong Zhou, Helan Hu, Tong Shen, Tianxiao Xu, Chenye Han, Weijie Li, Zhongquan Zhou, Hongming Xu, Yuekeng Li, Erenkai Zhu, Peng Yu, Wenhao Li, Jingyi Kang, Xixuan Hao, Xuzhou Zhu, Can Chen, Jiawei Liu, Chen Jiang, Tao Jiang, Fangwei Xiang, Zepeng Yao, Kai Chen, Guang Yang, Mingfei Lu, Chunhui Li, Zhengxuan Jiang, Bo Tang, Shiyu Zhan, Yihang Sun, Jie Chu, Wenjun Xiong, Zhenxiang Hua, Chen Lin, Detong Su, Dongyu Deng, Qiuping Ding, Guofeng Ding, Si Chen, Guohua Wen, Xu Chen, Zhi-Qin John Xu, Ying-Cong Chen, Ningyu Zhang, Haofen Wang, Muning Wen, Haotian Chen, Wentao Zhang, Xuanhe Zhou, Rita Zhang, Junchi Yan, Feiyu Xiong
The paper introduces StreamStop, a lightweight stopping policy for reasoning-enabled language model agents that learns offline from prefix-level action agreement. By encoding available reasoning text with a frozen prefix reader and a stop head, StreamStop can decide when to stop reasoning without needing internal signals from the model. Experiments on WebShop and ScienceWorld show that StreamStop achieves higher success rates than many fixed-budget baselines while using fewer reasoning tokens, notably reducing token usage by 42% on WebShop with Qwen3.8-27B.
By JunHyeok Oh, Byung-Jun Lee
The paper presents a method for creating LiDAR-based 3D object detectors that remain accurate under point‑cloud corruptions while meeting strict latency and memory constraints. It achieves robustness by generating corrupted samples that mimic extreme weather and training detectors to focus on hard samples, guided by perturbation sensitivity estimated from an ensemble of seven heterogeneous detectors. Efficiency is improved by narrowing the network and using offline neighborhood‑density measurements to reduce the number of points processed at each sampling stage, resulting in significant gains in AP under corruption and reductions in inference latency and GPU memory usage across multiple datasets and detector architectures.
By Tingxu Han, Zijie Shen, An Guo, Lyuke Wang, Chunrong Fang, Zhenyu Chen, Wei Song
LADBench is a new benchmark comprising over 1,000 synthetic images that contain logical anomalies across four domains—Residential, Urban, Collaborative, and Nature. It introduces a Tiered Prompting Protocol that progressively discloses information to test how much explicit help a model needs to locate and reason about these faults. Evaluation of leading foundation models shows that even the best only reaches 70.11% accuracy, indicating that implicit logical fault detection remains largely unsolved.
By Sahasra Kondapalli, Lara Radovanovic, Aadi Palnitkar, Mingyang Mao, Xiaomin Lin
CLOSER is an evidence‑aware active visual verification framework for embodied object‑goal navigation. It fine‑tunes a vision‑language model to label each view of a tracked candidate as target, non‑target, or uncertain, then selects complementary views using an embedding‑based diversity gate and fuses their probabilities with a lightweight network. When uncertainty remains, a diffusion policy directs the agent toward the candidate using only the tracked mask and depth until a decisive judgment is reached.
By Pengying Wu, Mingye Jiang, Yunze Hu, Zhejia Wen, Kangjie Zhou, Zongru Yang, Zhe Peng, Chang Liu
arXiv:2610.28201v1 Announce Type: new
Abstract: Recursive self-improvement (RSI) of agent harnesses has emerged as an effective approach to enhancing the capabilities of large language model (LLM) ag...
By Yu Cui, Hong He, Ruiqing Yue, Sicheng Pan, Zhuoyu Sun, Yifei Liu, Ting Li, Wenzhuo Zhu, Baohan Huang, Haibin Zhang, Cong Zuo
arXiv:2608.21755v2 Announce Type: replace
Abstract: Long-horizon agents require more than retrieval over stored conversations: they must preserve source evidence, revise outdated knowledge, and disti...
By Yu Qian, Hong Miao, Boyang Guo, Tingyi Jiang, Shan Zhao, Tianxing Le, Lintian Li, Meng Liu
arXiv:2610.24611v2 Announce Type: replace
Abstract: Evaluating software agents on a full benchmark is expensive. A small task subset can preserve agent rankings while yielding inaccurate scores. Hist...
By Wanyi Chen, Wanqin Shen, Yuanchen Shi, Zhensu Sun, Fang Kong
OptionWorld is a new benchmark that tracks how large language models handle long‑term personal assistant tasks, focusing on information gathering, decision making, actions, and outcomes over simulated daily life scenarios. The study finds that in 35.5% of failed cases the key facts are still present in the model’s last input, and identifies three main failure modes: missing or uncombined information, unused action opportunities, and unnecessary restrictions imposed by the agent. Follow‑up experiments show that providing missing facts, adding action reminders, and correcting erroneous restrictions can improve task success, highlighting gaps in state maintenance, deadline adherence, and assumption handling.
By Ruoyu Zhang, Zhaocen Liu, Weiyi Tian, Melissa Feng, Shilong Liu
PROVISO is a controller that selects which action an LLM agent should commit after the agent proposes an action and before execution. It encodes trusted and external observations, the proposed action, and an alternative into provenance-typed memory, uses role‑specific queries to relate them, and commits the action with the highest calibrated value learned from outcomes. In four agent security tasks, PROVISO outperforms all baselines in both security and utility, improving F1 by 13.84 points on SkillsBench and 7.07 points on ToolSafe, and achieving 88–99 F1 on three additional skill detection benchmarks.
By Bo Yan, Qi Zhang, Ziwen Kan, Fnu Suya, Song Wang
arXiv:2610.28010v1 Announce Type: new
Abstract: The integration of Large Language Models and Reinforcement Learning offers a promising direction for complex sequential decision-making tasks, such as...
By Artem Mikhaylov, Pavel Konoplev, Nikolay Efanov, Artyom Sorokin, Evgeny Burnaev
arXiv:2610.27621v1 Announce Type: cross
Abstract: Deep learning (DL) models are increasingly being deployed into radiation-challenging environments in spacecraft, particle accelerators, and, soon, sp...
By Peter Xiangyuan Ma, Joshua S. Bloom