Ace-Skill: Bootstrapping Multimodal Agents with Prioritized and Clustered Evolution
A new paper on arXiv introduces Ace-Skill, a co-evolutionary framework for self-evolving multimodal agents. The framework addresses data inefficiency and knowledge interference through a prioritized sampler with lazy-decay proficiency tracking, focusing rollouts on informative and insufficient samples. It also clusters knowledge to reduce retrieval noise. The approach aims to break the self-reinforcing failure loop where uninformative rollouts produce noisy knowledge that degrades future rollouts.
Key facts
- Ace-Skill is a co-evolutionary framework for self-evolving multimodal agents.
- It combines a prioritized sampler with lazy-decay proficiency tracking.
- The framework addresses data inefficiency and knowledge interference.
- It clusters knowledge to reduce retrieval noise.
- The paper is published on arXiv with ID 2605.08887.
- The approach aims to break the self-reinforcing failure loop in self-evolving agents.
Entities
Institutions
- arXiv