Sample-Efficient Learning from Agent Experience
Experience Distillation keeps much of an agent’s learned trial-and-error behavior without more environment runs.
The paper frames costly interaction as the bottleneck for real-world agent learning. Its method distills collected interaction histories into model weights, so the agent does not need those histories in context at test time. On 749 software-engineering tasks and six text-adventure games, it preserves at least 64.8% of in-context learning gains, compared with 3.8% for direct supervised fine-tuning. The authors also report matching classical reinforcement-learning baselines with at least 9.6x fewer environment samples. HF Daily Papers' note
The paper frames costly interaction as the bottleneck for real-world agent learning. Its method distills collected interaction histories into model weights, so the agent does not need those histories in context at test time. On 749 software-engineering tasks and six text-adventure games, it preserves at least 64.8% of in-context learning gains, compared with 3.8% for direct supervised fine-tuning. The authors also report matching classical reinforcement-learning baselines with at least 9.6x fewer environment samples. HF Daily Papers' note
score 5