Rethinking OPD

This work provides an interesting point of view on analyzing LLM On-Policy Distillation. Probably migratable to multimodal RL.