VCE-Skill lets AI agents learn from past mistakes
New method boosts AI agent performance by 3-5 points using version-control data
Researchers from China’s National University of Defense Technology and other institutions have developed VCE-Skill, a framework that enables AI agents to learn from historical skill upgrades rather than starting from scratch each time. The method distills noisy, implementation-specific version changes from public skill histories into reusable, structured evolution priors. These priors are then adaptively fused with trajectory-derived proposals from a base evolver, combining external experience with task-specific evidence.
In experiments, VCE-Skill improved skill self-evolution by 3.20 to 4.98 points on mean scores, demonstrating its effectiveness in leveraging version-control data. Additionally, the resulting skills achieved stronger cross-model transfer performance, suggesting broader applicability across different AI systems. The work highlights public skill version changes as an underexplored yet powerful source of prior knowledge for AI agents.
- VCE-Skill distills historical skill upgrades into reusable evolution priors
- Boosts AI agent performance by 3.20-4.98 points in experiments
- Enables stronger cross-model transfer performance for skills
Why It Matters
Unlocks faster, more efficient AI agent learning by reusing past improvements, reducing redundant training cycles.