Experience Distillation for Sample-Efficient Agent Learning
July 22, 2026
Experience Distillation allows agents to internalize knowledge from interaction histories via context distillation without requiring additional environment samples. Testing on 749 software engineering tasks shows agents can move in-context learning gains into model weights efficiently.
HOW THIS AFFECTS YOU
●
builderThis offers a path to reducing long-term inference costs by moving in-context experience into model weights.
●
researcherYou can optimize agent training by distilling interaction history without the cost of extra environment steps.