@neural_avb: This article actually explains all the components of On Policy Distillation Loss functions (forward vs reverse KL), sup…

X AI KOLs Timeline News

Summary

This article explains components of On Policy Distillation loss functions, including forward vs reverse KL divergence, supervision granularity, privilege types, and privilege advantage estimation. Additional resources from Thinking Machines, Hugging Face, and others are provided.

This article actually explains all the components of On Policy Distillation Loss functions (forward vs reverse KL), supervision granularity, types of privilege, and this concept of privilege advantage estimation. Additional resources I like for OPD: Thinking Machines/Reverse KL: http://thinkingmachines.ai/blog/on-policy… TRL: http://huggingface.co/spaces/Hugging… wh: http://x.com/nrehiew_/statu… Will Brown: http://x.com/willccbb/statu…
Original Article
View Cached Full Text

Cached at: 07/11/26, 07:21 AM

This article actually explains all the components of On Policy Distillation

Loss functions (forward vs reverse KL), supervision granularity, types of privilege, and this concept of privilege advantage estimation.

Additional resources I like for OPD:

Thinking Machines/Reverse KL: http://thinkingmachines.ai/blog/on-policy…

TRL: http://huggingface.co/spaces/Hugging…

wh: http://x.com/nrehiew_/statu…

Will Brown: http://x.com/willccbb/statu…

Similar Articles

On-Policy Distillation (5 minute read)

TLDR AI

This paper introduces on-policy distillation, which trains a student model on its own trajectories with teacher token-level KL supervision to fix train-inference mismatch, unifying forward-KL, reverse-KL, and JSD losses, with reverse-KL favored for smaller students.

The Many Faces of On-Policy Distillation: Pitfalls, Mechanisms, and Fixes

Hugging Face Daily Papers

This paper presents a comprehensive empirical study on on-policy distillation for large language models, identifying failure mechanisms like distribution mismatch and optimization instability, and proposing fixes such as stop-gradient objectives and RLVR-adapted teachers.