At each training round, Qwen3-8B or Qwen3-14B generated one on-policy interaction trajectory per task.
Agents improve by distilling hindsight from their own interactions
Retrospective Self-Distillation converts completed trajectories, including failures, into training signals without ground-truth solutions or verifier rewards for policy updates.
Big Tech
Yuhang Zhou · Yifan Wu · Bo Peng · Mingyi Wang · Jiayi Liu · Xiangjun Fan · +2 more
Meta AI
Research Digest··2 min read
Zhou et al.
§