Thinking Machines Lab proposes On-Policy Distillation training method
Original title:On-Policy Distillation
AI summary
Thinking Machines Lab published a technical blog proposing on-policy distillation: trajectories are sampled from the student model, and the teacher model computes reverse KL for each token as a dense reward.
Quick Facts
- Published
- Updated
- Source
- Thinking Machines Lab:官方博客
- Category
- Papers & Research
- AI score
- 68
原文给出反向 KL 损失、成本对比和可复现实验,读者可以据此评估用稀疏 RL 换成密集蒸馏监督的落地空间。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Thinking Machines Lab proposes On-Policy Distillation training method. October 27, 2025. https://new.nat6.net/en/news/cmq75sy1p0006sl5wfu0bt7hs/