general family · filed under post-training
On-Policy Distillation
Distillation based on student-induced states or responses, without evidence specifying a more particular variant.
Also called on-policy distillation (OPD).
- sources
- 7
- models
- 5
- labs adopt it
- 3
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
Documented in
Further reading
Picked by hand, not extracted: where to read more, not evidence for anything on this page.
- On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes (Agarwal et al., 2023) paper arxiv.org
- On-Policy Distillation - Thinking Machines Lab explainer thinkingmachines.ai
Evidence
9 spans quoted from the sources, strongest treatment first.
post-training pipeline replaces reinforcement learning with On-Policy Distillation from 10+ domain specialists
a single unified model is trained through on-policy distillation
we employ multi-teacher On-Policy Distillation (OPD; lu2025onpolicydistillation, minillm) as the primary technique for merging expert capabilities into the final model.
the mixed Reinforcement Learning (RL) stage was entirely replaced by On-Policy Distillation
followed by reinforcement learning (RL) and on-policy distillation (OPD; Gu et al., 2024; Lu and Lab, 2025)
we introduce a second-stage training procedure based on on-policy distillation (OPD), with the goal of distilling the model’s stronger response capabilities under text inputs into the audio-input setting.
unified model consolidation via on-policy distillation
unified model consolidation via on-policy distillation
on-policy distillation using dense, token-level rewards from teacher logits
Filed alongside
Other methods under post-training :: policy distillation.