[논문리뷰] Revisiting On-Policy Distillation: Empirical Failure Modes and Simple FixesLarge Language Model (LLM)의 Post-training에 있어 On-policy Distillation (OPD)은 student-generated rollouts에 대한 teacher feedback을 활용하기 때문에 매력적이다.#Review#On-policy Distillation#LLM Post-training#Sampled-token OPD#Variance Reduction#Local Support Matching#Truncated Reverse-KL#Top-p Rollout Sampling#Special Token Masking2026년 3월 26일댓글 수 로딩 중