# AwesomeOPD

A curated list of open-source repositories and papers on training language, vision-language, and agent models with on-policy distillation and on-policy self-distillation, organized by teacher source, supervision signal, and training-pipeline stage.

AwesomeOPD is a curated list of research papers and open-source repositories on On-Policy Distillation and On-Policy Self-Distillation, two approaches to training language and vision-language models where a student model learns from its own sampled trajectories under a teacher's supervision.

Entries are annotated along four axes: where the teacher's supervision comes from, what kind of signal it provides, how the student's rollouts are used, and where the technique sits in a training pipeline. Sections cover foundational surveys, white-box and black-box distillation, iterative self-bootstrapping, hybrids with reinforcement learning, applications to reasoning, multimodal, and embodied agents, speculative-decoding distillation, and frameworks and toolkits.

The list was built by reading papers, project pages, and source code with the help of coding agents, then reviewed by hand.

Website: https://github.com/thinkwee/AwesomeOPD