This paper introduces a five-category ASR error simulation framework, evaluating its impact on embodied AI safety; findings show error type significantly influences risk levels.

Sihan Jia, Oliver Lemon

CE-MoE reduces GPU hours by 33.3% at 31.5B parameters using layer re-configuration.

Simeng Sun, Roger Waleffe

SpikeOPD stabilizes on-policy distillation for autoregressive spiking language models, improving accuracy by 2.9 points.

Enqiao Lu, Xingrui Yu, Yiwei Fu et al.

CorporateBench is a large-scale, temporally evolving knowledge base benchmark for multi-task enterprise QA, with over 230,000 documents.

Sil Hamilton, Albert Yu Sun, Oscar J. Romero et al.

WM-R1 replaces real environments with world models, significantly enhancing GUI agents' reasoning capabilities.

Yu Han, Tianwen Qian

Addressing critical global challenges, from food security and disaster risk to disease outbreaks and socio-economic vulnerability, demands high-fidelity geospatial modeling. However, building predicti

Evelyn Ma, Rama Kumar Pasumarthi, Kishwar Shafin et al.

SwarmWorld employs decentralized LLM agents in a shared environment to evolve durable technologies, outperforming isolated search in diversity and resilience.

Subhadeep Pal, Fiona Y. Wang, Markus J. Buehler

Proposes MO-IKE, a multi-objective RL method, boosting knowledge editing reliability to 92% and retention to 63.4%, balancing generality and specificity.

Xuzhong Wang, Maiqi Jiang, Tejal Nair et al.

SPO++ aligns advantage normalization with action-token measure using event-time memory, boosting asynchronous RL efficiency with 19-point reward improvement.

Kai Ruan, Jinghao Lin, Qianshan Wei et al.

FedV-KGQA enables multi-hop reasoning over vertically partitioned knowledge graphs with near-centralized performance, using local graph enrichment and entity anchoring.

Md Saikat Islam Khan Bappy, Oshani Seneviratne

CES-PK framework uses lightweight symbolic constraints with three-valued semantics to verify LLM-generated answers over incomplete knowledge graphs, improving precision.

Emanuel Kitzelmann

Introduces FrontierChallenge, a cross-domain benchmark for scientific workflow completion, with a top model success rate of only 20.6%.

Liangcai Su, Zhaopeng Feng, Zhuo Chen et al.

StarHarness employs stratified search to evolve environment-specific harnesses, boosting performance by 20-35 points without changing model weights.

Esakkivel Esakkiraja, Denis Akhiyarov, Vikas Yadav et al.

CAFE framework enhances search agents via co-evolving feedback, excelling on seven benchmarks.

Boyang Liu, Senjie Jin, Peixin Wang et al.

MHA-CSP uses Mahalanobis distance for complex state propagation, with only 119K parameters.

Xiaohe Li

VideoHarness-RSI recursively optimizes context construction for frozen vision-language models to improve long-video understanding.

Guoyang Xu, Hao Chen

ReWorld uses split training and pose-indexed landmark bank to balance long-term memory and real-time control, achieving state-of-the-art fidelity.

Zhifei Chen, Luozhou Wang, Guibao Shen et al.

Prime Agent employs recursive language models and persistent tools to achieve 95.5% on ARC-AGI-3, enabling long-horizon autonomous evaluation and multi-agent coordination.

Seth Karten, Alex L. Zhang, Kevin Thomas et al.

Using Bayesian latent ability models, the study finds independent reasoning boosts skill development, while frequent AI help can cause dependency and performance decline.

Shang Wu, Catarina G Belem, Shuyuan Fu et al.

Proposes Safety-Direction Penalty (SDP) based on representation geometry to mitigate reasoning-induced safety misalignment, restoring safety in Qwen2.5-3B and 7B models.

Yipeng Zhao, Qishun Yang, Shenzhe Zhu et al.