ExSkill: Continual Learning from Experience and Skills in Multimodal Agents

摘要

Multimodal agents demonstrate impressive problem-solving capabilities but typically operate in isolated episodes without leveraging past experiences. Recent methods address this through dynamic retrieval of textual insights or predefined skill documents, yet face critical challenges: visual modalities are neglected during knowledge extraction, stored insights lack executable structure, and manually crafted skills fail to scale. We propose \textsc{ExSkill}, a framework combining task-level Skills (structured workflows and tool templates) with action-level Experiences (context-specific tactical insights) through automated accumulation from agent trajectories. Our approach employs visually-grounded summarization to extract knowledge integrating visual observations and textual reasoning, hierarchical consolidation to maintain quality and diversity, and context-aware adaptation to tailor knowledge to current visual contexts. Evaluated on five diverse benchmarks spanning visual tool use and multimodal search, \textsc{ExSkill} achieves average gains of 4.1-6.5 points over strong baselines across different backbone models, with superior zero-shot transferability and strategic improvements in tool selection and execution accuracy. These results demonstrate that our framework enables transferable continual learning for multimodal agents in real-world scenarios without parametric training, offering broad applicability for practical deployment.