HINT-SD: Targeted Hindsight Self-Distillation for Long-Horizon Agents

Researchers introduced HINT-SD, a hindsight-based self-distillation framework for long-horizon LLM agent training.

It identifies failure-relevant action spans in full trajectories and applies feedback-conditioned distillation only to those parts.

On BFCL v3 and AppWorld, HINT-SD beat a dense per-turn feedback baseline by up to 18.8%.

It also cut training-step time by 2.26x, showing that selective supervision can improve both accuracy and efficiency.