ExpVoyager: Direct Experience Navigation for Dynamic Agent Skill Synthesis

This paper introduces ExpVoyager, a framework that reframes LLM agent skill synthesis as a dynamic navigation problem over past experience. Instead of abstracting past experience into fixed procedural knowledge prematurely, a skill curator actively explores accumulated raw trajectories on demand, tailoring access to current task needs. Experiments demonstrate consistent improvements in downstream task performance, continual gains as the experience space scales, and practical compatibility with existing skills under efficient experience access.

ExpVoyager reframes agent skill synthesis as a dynamic navigation problem over past experience rather than relying on fixed procedural knowledge.

A skill curator navigates raw experience across different views and resolutions to identify reusable procedural knowledge on demand.

Extensive experiments demonstrate consistent downstream task performance improvements and continual gains as the experience space scales.