From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution
Researchers have explored new methods for representing reusable experience in AI systems, focusing on how this experience can be used for test-time control and iterative evolution. Their study, involving over 4,500 trials across 45 code-solving scenarios, found that a compact "Gene" representation significantly outperformed a documentation-oriented "Skill" package. The Gene representation proved more stable, provided stronger overall performance, and was a better carrier for accumulating experience, particularly when failure history was distilled into compact warnings. AI
IMPACT Suggests a more effective method for encoding AI experience, potentially improving model adaptability and learning efficiency.