Beyond Imitation: Reflective On-Policy Self-Distillation for LLM Reasoning

Name: Beyond Imitation: Reflective On-Policy Self-Distillation for LLM Reasoning
Uploaded: 2026-10-03T14:24:57.860039+00:00
Duration: 282 s
Description: Reflective On-policy Self-Distillation (ROSD) improves LLM reasoning by shifting from imitating correct answers to correcting errors. It uses a self-reflector to identify key fixes and then applies distillation loss only to the erroneous parts of the response. This approach leads to stronger performance both in known tasks and when applied to new, unseen data.

Reflective On-policy Self-Distillation (ROSD) improves LLM reasoning by shifting from imitating correct answers to correcting errors. It uses a self-reflector to identify key fixes and then applies distillation loss only to the erroneous parts of the response. This approach leads to stronger performance both in known tasks and when applied to new, unseen data.

Open the full episode