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

In three linesHINT-SD proposes targeted self-distillation for training long-horizon LLM agents. The method uses full-trajectory hindsight to identify failure-relevant actions and applies feedback-conditioned distillation only on targeted action spans. On BFCL v3 and AppWorld, it improves over dense per-turn feedback baselines by up to 18.80% while achieving 2.26× lower time per training step.

Summary generated by Claude — human-verified