Reward-Aligned Distillation

Yilan Chen, Zian Su, Manish Prajapat +4

Reinforcement learning with verifiable rewards improves the reasoning capabilities of large language models, but remains computationally expensive and sample inefficient. On-policy…