Yu Lin

The paper demonstrates that using on-policy distillation from a strong teacher model significantly improves the performance of compact Automatic Speech Recognition (ASR) models, achieving competitive results with a much smaller audio dataset compared to supervised fine-tuning.

The paper introduces Lookahead Group Reward (&) to combat Supervision Fidelity Decay (SFD) in on-policy distillation, significantly improving student model performance on long reasoning tasks.

The paper introduces Atomic Decomposition and Recombination (ADR), a novel framework that generates genuinely novel and challenging verifiable code tasks, significantly improving the scalability of Reinforcement Learning with Verifiable Rewards (RLVR) for LLMs.