H^2SD: Hybrid Hindsight Self-Distillation
H²SD changes the teacher’s role depending on whether the model’s reasoning attempt succeeded or failed.
For correct trajectories, the method uses the verified response as context and has the teacher re-evaluate the same tokens, refining credit assignment without replacing the student’s path. For failed trajectories, it supplies a verified reference hint and applies reverse-KL distillation as corrective guidance. The authors report stronger overall results than representative RLVR and self-distillation baselines on reasoning benchmarks, with stable optimization and a favorable accuracy-efficiency trade-off. HF Daily Papers' note
For correct trajectories, the method uses the verified response as context and has the teacher re-evaluate the same tokens, refining credit assignment without replacing the student’s path. For failed trajectories, it supplies a verified reference hint and applies reverse-KL distillation as corrective guidance. The authors report stronger overall results than representative RLVR and self-distillation baselines on reasoning benchmarks, with stable optimization and a favorable accuracy-efficiency trade-off. HF Daily Papers' note
score 4