On-policy distillation – Definition & Meaning

Definition of on-policy distillation

: a form of model distillation in which the student generates its own outputs or trajectories and learns from teacher feedback on those student-visited states
- The team used on-policy distillation so the student could receive corrections on mistakes it actually made during generation.
- On-policy distillation reduced the mismatch between the examples seen during training and the sequences produced by the deployed student.

Origin & history

first known use2023

A 2023 paper on distilling language models from self-generated mistakes formalized the approach through Generalized Knowledge Distillation, training a student on its own generated sequences with feedback from a teacher. The name on-policy distillation subsequently became common in research on language-model post-training.

Test yourself

Which of these is the meaning of on-policy distillation?

Related terms

distillation training a smaller "student" model to imitate the outputs of a larger "teacher" model… post-training everything done to a model after pre-training to make it useful and safe, such as instruction… reinforcement learning a way of training in which a model tries actions, receives rewards or penalties, and gradually… fine-tuning further training of an existing model on a smaller, focused dataset so it gets better at a… adversarial distillation the systematic, unauthorized use of a capable AI model's outputs or reasoning traces to train… reasoning model a language model trained to work through a problem step by step before giving its final…

Cite this entry

"On-policy distillation." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/on-policy-distillation/. Accessed 6 Oct. 2026.