The authors introduce OPDArena, an evaluation framework covering five pretrained language models and four types of on-policy distillation teachers.
Teacher training determines whether on-policy distillation preserves prior language skills
Across five pretrained models, teachers trained with verifiable rewards produced more selective student updates and less capability loss than supervised fine-tuned teachers.
Big Tech
Yuanze Xu · Tianzhu Ye · Hui Sun · Wenhui Wang · Li Dong · Furu Wei
Microsoft Research · Nankai University
Research Digest··2 min read
Xu and colleagues compare supervised fine-tuning, reinforcement learning with verifiable rewards, and on-policy distillation under a shared mathematical post-training setup.
§