Tag
The paper proposes Reasoning-Progress-Aware Reward Filtering for On-Policy Distillation (R2-OPD) to address the mismatch between teacher-derived rewards and genuine reasoning progress, improving reasoning performances in language model training.