Submitted by Bingxiang He 83 Rethinking On-Policy Distillation of Large Language Models II: One Training Example Thinking Space 51 2