Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs
Echo-GRPO addresses a problem where using strong teacher traces in reinforcement learning for video models causes gradient clipping on important reasoning steps. The solution is Echo-GRPO, which rewrites these privileged traces into the student model's own style while keeping the meaning intact. This method stabilizes training and allows the model to learn better justifications.