On-Policy Distillation (5 minute read)

TLDR AI Papers

Summary

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.

On-policy distillation trains a student model on trajectories sampled from its own policy while a teacher provides dense token-level supervision through KL-based regularization, closing the train-inference distribution mismatch that off-policy methods suffer. The canonical formulation unifies forward-KL, reverse-KL, and JSD losses with reverse-KL emerging as the default for mode-seeking smaller students, and a one-line code swap of the regularizer model on top of an RL stack like Tinker implements the technique.
Original Article

Similar Articles

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.

OPRD: On-Policy Representation Distillation

Hugging Face Daily Papers

OPRD proposes a new knowledge distillation method that aligns student and teacher hidden states across layers during on-policy rollouts, eliminating sampling variance from token-space KL estimation. Empirically, OPRD outperforms output-space baselines on math reasoning benchmarks (AIME 2024/2025, AIMO) while being 1.44x faster and using 54% less memory.

DOPD: Dual On-policy Distillation

Hugging Face Daily Papers

DOPD proposes a dual on-policy distillation paradigm that dynamically routes token-level supervision between privileged teacher and student policies based on advantage gaps and probabilities, addressing privilege illusion and improving capability transfer in LLMs and VLMs.