Dissecting Embodied Abilities in Multimodal Language Models through Skill-level Evaluation and Diagnosis
通过技能级评估与诊断解构多模态语言模型的具身能力
机构 * Northeastern University, Boston, MA, USA ; The Chinese University of Hong Kong, Hong Kong, China ; Peking University, Beijing, China ; Westlake University, Hangzhou, China ; Harvard University, Cambridge, MA, USA ; Purdue University, West Lafayette, IN, USA ; University of Oxford, Oxford, United Kingdom
AI总结 本文提出BEAR基准,通过分解具身任务为14个原子技能进行细粒度评估,发现感知能力是推理失败的主要瓶颈,并提出BEAR-Agent多模态对话代理,显著提升具身技能性能。
Comments Accepted to ICML 2026