Salesforce paper: RL expert distillation saves more compute than external large-model distillation
The study compares distillation against GRPO based on retained accuracy per training FLOP, covering varying task difficulties, model sizes, and domains. When using an externally trained teacher, GRPO generally outperforms distillation, which often leads to longer responses.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original