model training 相关 AI 资讯

09:54

官方账号arXiv cs.AI@Justin Robert, Raheel Qader

On-Policy Self-Distillation (OPSD) removes the need for a second model as teacher, using the model itself with privileged information. Early results show accuracy comparable to reinforcement learning. However, the review identifies 'collapse' as a dominant failure mode, which is exacerbated by privileged information. The review analyzes collapse through three factors: signal application, teacher's privileged information, and signal change dynamics. The study focuses on mathematical reasoning, with no new experiments reported. The contribution is a shared vocabulary for discussing phenomena and clarifying settled and disputed points.

推荐理由：This review delves into the challenges of On-Policy Self-Distillation, highlighting the 'collapse' issue and offering insights into its causes. It's a must-read for those interested in understanding the nuances of model training with privileged information.

原文