Submitted by Bingxiang He 104 Rethinking On-Policy Distillation of Large Language Models II: One Training Example Thinking Space 92 2