New paper rethinks offline on-policy distillation with distributional supervision
A new arXiv paper proposes moving on-policy distillation beyond scalar advantages to distributional supervision, addressing the cost of repeatedly evaluating teacher models on student-generated trajectories.
Read the full story at arXiv cs.LG ↗
Timeline · 1 report
- 2026-11-17 05:00 · arXiv cs.LG
Rethinking Offline On-Policy Distillation: From Scalar Advantages to Distributional Supervision