The paper introduces Modality-Aware Expert Pruning (MAEP), a training‑free method that decomposes expert importance by modality in a single forward pass for Mixture‑of‑Experts multimodal large language models. MAEP uses a Delta‑H Output metric to measure the impact of removing an expert and leverages low‑attention vision tokens as a stable pruning signal, while preserving text‑side importance. Experiments on three MoE‑based MLLMs show that MAEP reduces the BF16 expert weight footprint by about 27 GB (≈47% of total memory), enabling a 58 GB model to fit on a single 40 GB GPU and outperforming modality‑agnostic baselines on most multimodal benchmarks.
By Donggeun Kim
LongBanana is the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis, featuring 90 tasks each with about 32 reference images, 20 constraints, and 50 evaluation items, totaling 4,670 atomic evaluation items. The benchmark uses a graph-grounded evaluation design that converts entity-relation graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. An accompanying agent-harness framework, LongBanana-Gen, demonstrates that minimal harnesses improve generation quality, with a ReAct-Plan-Exec-Reflect workflow further enhancing evidence grounding and repair.
By Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo, Mingzhe Zheng, Junwen Miao, Huaisong Zhang, Yang Yao, Yu Wang, Dongfu Jiang, Ping Nie, Wenhu Chen, Chaoqun Wang, Changqian Yu, Kelsey R. Allen
The paper presents a comprehensive evaluation of 11 memory substrates for long‑horizon LLM agents, using a unified set of 26 performance and efficiency metrics across three backbone models and four benchmark suites. It finds that no single substrate consistently outperforms others; retrieval‑heavy substrates aid long‑context factual QA but can impair sequential decision‑making, and scalability can shift performance trade‑offs. The study demonstrates that selecting a memory substrate tailored to the operating regime yields better results than any fixed choice, offering empirical guidance for designing efficient, reliable, and regime‑aware long‑term memory systems.
By Wei-Chieh Huang, Weizhi Zhang, Yuchen Wu, Yankai Chen, Eric Hanchen Jiang, Wooseong Yang, Yiwei Yang, Henry Peng Zou, Hanrong Zhang, Ying Nian Wu, Haolun Wu, Kai-Wei Chang, Philip S. Yu, Xue Liu, Aylin Caliskan
LITMUS is a benchmark for testing the safety of large language model agents in real operating system environments, focusing on behavior jailbreaks that can cause irreversible OS-level harm. It includes 819 high-risk test cases organized into a harmful seed subset and six attack-extended subsets covering jailbreak speaking, skill injection, and entity wrapping, and employs a semantic-physical dual verification mechanism with OS-level state rollback. Evaluation of frontier agents shows that even strong models can execute dangerous operations, exhibit execution hallucination, and are vulnerable to skill injection and entity wrapping attacks.
By Chiyu Zhang, Lu Zhou, Xiaogang Xu, Huiqin Yang, Bendong Jiang, Xiaolei Zhang, Yiran Zhao, Ruyi Chen, Jiafei Wu, Liming Fang, Zhe Liu
Open‑HumanScale presents the first fully open recipe for scaling egocentric human video pretraining to 50,000 hours, supported by the curated Open‑HumanScale‑50K corpus annotated with hand motion and natural language. Using this corpus, the authors develop HumanWAM‑Vanilla, a world–action model that achieves high success rates on RoboTwin 2.0 and RoboDojo after robot post‑training, and HumanWAM‑Goal, a visual‑goal‑conditioned model that further improves zero‑shot performance and demonstrates embodied in‑context learning on physical robots. The paper also releases the dataset, curation pipelines, pretrained checkpoints, and evaluation tools to enable reproducible scaling research in embodied AI.
By Juncheng Ma, Jiaxin Li, Jiankai Tu, Bo Liang, Guodong Lv, Yufan Deng, Junzhe Tang, Liangzhu Yang, Chen Bian, Zicong Cao, Baohao Zhu, Jianxin Bi, Tengfei Yan, Yinyuan Zhang, Kewei Zhang, Xinglong Chen, Kaiqi Chen, Duomin Wang, Xiaotian Tang, Liang Tan, Li Li, Yuan Xu, Haosen Yang, Yuqi Wang, Jingfei Xia, Songyan Huang, Zhiyang Dou, Zhen Dong, Enze Xie, Tat-Seng Chua, Daquan Zhou
SenseShift6D is a new RGB‑D benchmark that systematically varies 13 RGB exposures, 9 RGB gains, auto‑exposure, 4 depth‑capture modes, and 5 illumination levels for six household objects, producing 795.4k RGB‑D scenes. Experiments show that state‑of‑the‑art pose estimators, even when pretrained on large datasets, exhibit significant performance drops across different sensor and lighting settings, revealing a strong sensitivity to environmental changes. The benchmark also demonstrates that test‑time sensor selection can improve accuracy, but current practical methods offer limited gains, indicating room for research in sensor‑aware adaptation.
By Yegyu Han, Taegyoon Yoon, Dayeon Woo, Sojeong Kim, Hyung-Sin Kim
The paper introduces MV-PVMapper, a training‑free framework that combines vision foundation models (SAM 3) with UAV photogrammetry to produce high‑fidelity maps of photovoltaic modules. By generating candidate contours, correcting for surface slope, and iteratively refining segmentation across multiple views, the method recovers 7,710 of 7,711 modules with no false positives and achieves macro‑averaged precision and recall above 98 %. The approach outperforms orthomosaic segmentation and multi‑view pixel‑voting baselines, reaching a Jaccard index of 97.55 % without any task‑specific training data.
By WenHao Duan, Qi Chen, Penglei Zhang, Ruizhe Deng, Zhiling Guo, Haoran Zhang, Yuntian Chen
MA‑SafeDiffuser is a multi‑agent extension of SafeDiffuser that incorporates diffusion‑based trajectory planners with finite‑time diffusion invariance guarantees for joint‑agent safety specifications. It formalizes joint safe sets as intersections of per‑agent and pairwise barrier sets, and derives both centralized and decentralized constrained denoising procedures that maintain invariance under mild assumptions. The approach also tackles local‑trap and deadlock issues using time‑varying specifications and liveness CBFs, and introduces a lightweight benchmarking suite including a multi‑agent Maze2D domain, showing reduced violation counts compared to unconstrained diffusion baselines while preserving planning quality.
By Kiran Ravish, Ankita Kushwaha, Preeti, Pawan Kumar
The study evaluates six large language models (LLMs) on a 102-article benchmark about the Spanish plume weather pattern, testing them across four review stages: article discovery, interpretation, synthesis, and future research proposals. Five tools collectively found 57 relevant articles, with institutional library access boosting discovery for some models. While three tools correctly identified all 18 substantial articles, they misclassified a large portion of passing mentions and produced syntheses that, though largely accurate, contained unsupported claims and quotation errors, and their research proposals were often broader than the benchmark scope.
By Abdullah Al Kalbani, David M. Schultz, Alexandra K. Anderson-Frey, Zhonghua Zheng, Catherine McGuire
Agent Harness Awareness addresses the gap between tool descriptions and real‑time operational suitability by having the host provide a verified, up‑to‑date state snapshot before an agent dispatches an action. In controlled benchmarks, this approach cut observable policy regret by nearly 97 % and boosted single‑dispatch completion from 69.4 % to 95.8 %. In live, fault‑injected tests, verified state eliminated excess operational cost and achieved 100 % completion across all model‑by‑fault scenarios.
By Manu Agrawal, Shrey Nagpal
The paper argues that the term "verified" in LLM‑generated code is ambiguous, identifying three distinct failure surfaces: verifier acceptance (Surface A), syntactic specification binding (Surface B), and semantic specification fidelity (Surface C). Empirical tests on Python and Rust benchmarks show that high verifier acceptance does not guarantee binding to a specification, and even fully‑verified Rust outputs can fail hidden tests, revealing gaps between formal proofs and real‑world correctness.
By Aditya Vempaty, Himaneesh Sompalle, Manan Tayal, Mowafak Allaham, Prasenjit Dey, Ashish Jagmohan, Ravi Kokku
AvalancheBench is a multimodal enterprise analytics benchmark that tests data agents’ ability to recover a latent world from observed data, using 58 questions across text, images, and documents. It evaluates agents on analytical fidelity, monetary cost, and a new capability‑adjusted token cost (CATC) metric, revealing that misinterpreting requirements and misuse of semantic operators are the biggest sources of error. The benchmark highlights how errors can propagate through dependent questions, affecting overall performance.
By Darek Kleczek, Fuheng Zhao, Alexander W. Lee, Julien Tissier, Pawel Liskowski, Ugur Cetintemel, Anupam Datta
The paper introduces OPSD, a method that uses on‑policy self‑distillation to train a context‑free student model from a teacher conditioned on user memory, combined with a verify‑then‑regenerate router that directs queries to either weight or text memory. OPSD enables weight memory to compete with or surpass traditional text‑based memory systems on the HaluMem benchmark, reducing token usage per query. The study also explores how different training recipes influence weight memory performance and offers guidance for integrating it with retrieval in persistent, multi‑session agent deployments.
By Zeyu Fang, Jun Yin, Yongsheng Mei, Xu Han, Chaitali Joshi, Zhongkai Sun, Jie Hao, Bin Guo, Chengyuan Ma, Benjamin Z. Yao, Chenlei Guo
The paper introduces CHARM+, a Cross‑Hardware Attention framework that reorganizes sequential communications into multiple levels to enable intra‑level parallelism and overlap computation with communication. It decouples cross‑hardware from intra‑hardware communication, removing synchronization overhead caused by bandwidth imbalance. Experiments on dual‑NUMA PCIe and dual‑node NVLink systems show consistent speedups over state‑of‑the‑art methods, achieving 1.3×/1.7× and 2.8×/2.9× improvements for forward and backward passes respectively.
By Xinhang Zhang, Boning Zhang, Chengchun Liu, Chunpu Li, Lei Sun, Weifeng Zhang, Limin Xiao
The paper introduces a method to control reasoning behaviors in Mixture-of-Experts (MoE) language models by labeling sentence-level segments with six distinct reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion). Linear probes on MoE router logits can predict these behaviors nearly as accurately as probes on the full hidden state, enabling the authors to edit router scores during inference to reduce overthinking. This intervention cuts generated tokens by a median of 30.4% (up to 63.3%) across multiple models and benchmarks without significant loss in accuracy, and often improves it.
By Shengjie Wang, Mohsen Fayyaz, Nanyun Peng
Reglos is a training‑free personal memory agent that continuously updates an append‑only record of user interactions, allowing newer preferences to override older ones and enabling selective forgetting of facts. Its memory harness evolves through recursive self‑improvement: a coding agent generates patches from failure traces, a critic from a second model family screens them, and a fixed evaluator decides which patches to accept. Using the same backbone across systems, Reglos surpasses state‑of‑the‑art memory agents such as Mem0, MemOS, and EverOS, achieving 61.3 % and 84.0 % on PersonaMem‑v2 and LongMemEval‑S with Qwen3.8‑27B, and 57.5 % and 87.2 % with GPT‑5.6 Luna.
By Sen Yang, Jiaduo Lin, Zehong Huang, Wen Liu
The paper introduces MPRBench, a benchmark designed to evaluate multimodal personalized reasoning in multimodal large language models (MLLMs). It comprises 12 sub-tasks and allows for interpretable analysis of error cases. Experiments show that current MLLMs struggle with personalized reasoning due to inaccurate concept recognition and sensitivity to irrelevant personal information.
By Dohyun Kim, Junyong Lhim, Kwanyong Park
The paper introduces a new framework for multi-resident data association in smart homes, framing the problem as a Markov Decision Process and solving it with Behavior Cloning. Two policy architectures are proposed: a lightweight Transformer treating association as multi-label classification over event histories, and a fine-tuned Large Language Model that generates conditional text to leverage semantic knowledge. Experiments on three datasets show the approach outperforms existing baselines, especially for concurrent multi-resident events, and the LLM policy hints at cross-environment transfer potential.
By Xi Chen (M-PSI), Julien Cumin (LIG, M-PSI), Fano Ramparany (LIG, M-PSI), Dominique Vaufreydaz (LIG, M-PSI)
AmbiHTMLBench is a new benchmark that evaluates how well models can convert real-world documents into executable HTML while preserving content, organization, and visual appearance. It contains 1,115 pages with structured annotations and uses both objective metrics (table fidelity, text error, layout consistency) and a reference-anchored subjective protocol (LLM judges aligned with human preferences). Results show that different systems excel in objective versus subjective measures, and fine‑tuning a small expert model improves performance, highlighting the benchmark’s role as a diagnostic tool for advanced document parsing.
By Boyu Niu, Yingjie Xiong, Shurui Li, Linhan Pang, Yuntong Zhang, Han Wang, Hao Wu, Kai Zuo, Xu Tang, Xuanhe Zhou, Fan Wu
ElecLoad is a context‑aware agent that improves electricity load forecasting by adaptively organizing the forecasting process around multiple events. It analyzes each event’s relevance, timing, and interactions with historical load, then decides whether to forecast the entire horizon or partition it into segments. Sub‑agents generate initial forecasts for each segment, and the main agent refines them based on feedback, achieving state‑of‑the‑art performance on several datasets.
By Huajian Zhang, Mingyue Cheng, Xingpeng Gao, Xiaoyu Tao, Zhiding Liu, Jiqian Yang, Zhi Li