CoEM introduces a Commit-on-Evidence Memory system that learns when to compress source evidence into compact memory facts while preserving potentially useful excerpts verbatim in a pending set. The system uses a learned policy to decide whether to promote, retain, or discard each pending excerpt as new context arrives, and a frozen verifier ensures only supported facts are committed. Reinforcement learning trains this policy with step-level evidence rewards and final answer rewards, leading to consistent improvements in long-context reasoning, achieving 10.4–11.4 F1 points over the strongest baseline on 6,400-document inputs.
By Jingguang Li, Yebo Wu, Zuyi Guo, Kailang Ma, Xianjie Dai, Han Zheng, Benwang Chen, Li Li, Can Rong, Heye Huang
MatToolBench is a new benchmark that evaluates multimodal GUI agents on professional materials science software. It contains 204 tasks across 10 tools in three modalities—GUI operation, OriginPro scripting, and code-based database queries—executed inside a Windows 11 VM. The benchmark offers fine-grained, expert-decomposed scoring and a high-performing multimodal judge for aesthetic assessment, revealing that strong general benchmark performance does not transfer to scientific workflows.
By Mei Wu, Rui Xie, Runyu Zhang, Yuqiang Li, Tianfan Fu, Bo Chen, Kai Yu, Xin Chen, Lu Chen
SkillCome introduces a skill-evolution framework that uses group contrast optimization and a dual memory system to refine large language model skills. By generating multiple trajectories per question and contrasting successful versus failed ones, it identifies key behavioral divergences to guide skill edits. The dual memory accumulates historical evidence across groups, enabling more generalized and reliable optimization signals, leading to consistent performance gains on diverse benchmarks.
By Haolin Li, Feng Hong, Ang Li, Chilin Fu, Weichang Wu, Ya Zhang, Yanfeng Wang, Xiaolu Zhang, Jiangchao Yao
The paper introduces Milestone Viability Potential Policy Optimization (MVPO), a reinforcement learning algorithm designed for long‑horizon large language model agents. MVPO addresses the zero‑credit failure problem by learning from viable failure prefixes, estimating prefix potential over Union‑Find viability regions, and repairing zero‑credit groups with potential‑difference advantages. Experiments on Qwen2.5‑1.5B‑Instruct demonstrate that MVPO outperforms eight strong baselines, improving success rates on ALFWorld and WebShop with minimal overhead.
By Qi Zhou, Yuanfan Li
The paper investigates how the design of LLM-as-a-Judge protocols influences both the intrinsic quality of judgments and their downstream utility in open-ended tasks. By varying verdict granularity, critique usage, and evaluation batching, and by applying Judge guidance to test-time inference methods such as Best-of-N selection, revision, and beam search, the authors find that judgment quality and downstream performance do not always align and that protocol choices significantly affect outcomes. The study highlights the need for comprehensive evaluation of LLM Judges that considers both judgment quality and practical utility.
By Zheng Zhang, Lufei Li, Xinyue Tan, Yuanhao Zeng, Ziwei Shan, Yexin Li, Kan Ren
The paper introduces ToxicBench, a benchmark designed to evaluate how tool‑augmented data agents handle incorrect tool outputs. By pairing clean and poisoned observations across numerical, label, schema, and retrieval errors, the authors assess both the checking process and the final answer adoption. In a 118‑task GPT evaluation, poisoning reduces task success by 26–39 percentage points, revealing that repeated poisoning leads to wrong-answer adoption even after checking, while ordinary retries help only under one‑shot poisoning. Human annotations on 200 trajectories confirm the scoring system’s reliability, showing 96% agreement with task success and supporting the benefits of retries and audit‑based adoption.
By Zifu Tao, Changqing Yin
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
OptiCom introduces a unified framework for state-conditioned composition in large language model (LLM)-driven optimization. It models LLM optimizers within a shared configuration space (artifact, query, operator, evaluation, memory, strategy) and uses an Optimization Controller to dynamically compose mechanisms while a Strategy Adapter refines long-term preferences. Experiments on 32 benchmark groups show OptiCom outperforms 14 configurations, achieving the top score in 23 groups.
By Chenxing Wei, Sichen Liu, Lizhao Liu, Ningyuan Sun, Chen Bingzhou, Ying He, Bo Jiang, Fei Yu, Yao Shu
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
The paper introduces SaveRouter, a sparse‑supervision framework for large language model routing that selectively gathers informative model feedback and shares capability information across related queries. By using only about 33–41% of available training feedback, SaveRouter achieves competitive or superior routing quality while reducing the break‑even deployment volume by 1.9–9.5× compared to conventional routers. The study also shows that the supervision level that minimizes serving cost may differ from the one that yields the earliest payback.
By Guannan Lai, Gelin Bian, Hao-Xuan Ma, Jun-Peng Jiang, Long Chen, Jian-Dong Liu, Zhi-Hao Tan, Han-Jia Ye
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 two token-error supervision methods for sparse mixture-of-experts (MoE) large language models, aligning routing affinities with token-level cross-entropy loss. The first method predicts an error score per expert and uses it to adjust affinities before top‑K selection, while the second directly aligns router affinities to the model’s objective without extra heads or inference changes. Experiments on Granite and ARC‑Challenge show accuracy gains of about 2.3–2.94 percentage points over a parameter‑matched baseline, preserving the native sparse execution budget.
By Yury Nahshan, Nati Daniel, Jacob Goldberger, Yoli Shavit
Co-PiLOT is a latent optimization framework that maps candidate physical structures through a generative encoder-decoder, using the decoder as a learned validity prior and performing physics-informed black-box optimization in latent space. It is applied to the inverse design of magnesium alloy microstructure/texture, employing a vision transformer encoder paired with latent diffusion, diffusion transformer, and rectified-flow transformer decoders trained on an 80,000-sample EBSD dataset to produce a minimal bottleneck representation. The MERIDIAN optimizer, driven by a deep-kernel Gaussian process and failure-aware feasibility prediction, achieves the best target-driven objective score within 160 simulations, reducing relative target error by 3–22% compared to seven baseline methods.
By Mahish K. Guru, Mayank Nagar, Ayush vyas, Jan Bohlen, Roland Aydin, Noomane Ben Khalifa
The paper demonstrates that in open‑weight LLM inference markets, selecting a model is insufficient; clients must also choose a provider, as the same model can differ markedly in quality, latency, availability, and price across providers. The authors propose a market‑aware routing approach, including a measured‑map policy and an online router called FACET, which certifies provider feasibility for each task and safely falls back to a reliable anchor. Experiments show that this strategy yields cost savings while maintaining quality and avoiding degraded endpoints.
By Liang He, Jingbo Wen, Yixiong Chen, Yue Yang, Qizhen Lan, Kangning Cui, Xilu Wang
BrainNet Studio is a unified toolkit that enables the construction, analysis, and visualization of both static and dynamic brain networks. It integrates 27 algorithms—including deep learning, graph neural networks, and spatiotemporal sequence models—to support classification, biomarker identification, and the extraction of discriminative brain regions and connections. The toolkit also employs a large language model to generate researcher‑verifiable summaries of functional and structural connectivity, as well as structure‑function coupling, at individual and group levels.
By Xiwei Zeng, Shengrong Li, Yiheng Liu, Chunwei Tian, Daoqiang Zhang, Qi Zhu
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
TRACE is a deployable framework that enhances oncology language models by separating offline structure learning from online inference. It organizes oncology concepts and relations into an updatable tree‑relational structure, refines it with LM‑loss evidence, and retrieves compact prompt evidence during inference. The approach improves performance on ten classification tasks and a QA benchmark, outperforms vanilla RAG and generic GraphRAG, and provides interpretable evidence paths aligned with clinical reasoning.
By Jizheng Lai, Yingyun Li, Ying Qin, Haiyang Qian