On-policy distillation: the reading list frontier models actually use

What it does

A curated bibliography tracking 269 papers on On-Policy Distillation (OPD) for large language models. The repository sorts research into a systematic taxonomy—objective functions, signal sources, training stabilization, and applications—and anchors a living survey paper (V4) plus a companion site, OPDHub, with multi-axis filters and copy-ready BibTeX blocks.

The interesting bit

The maintainers treat the repo as an active research instrument, not a static list: it includes a teacher-student model atlas, loss-objective distribution charts, a Hall of Fame, and even a unified OPD framework called EasyOPD released as a companion toolkit. That level of infrastructure is unusual for an awesome-list.

Key highlights

• Tracks 269 papers with a systematic taxonomy spanning objectives, signal sources, and training stabilization.
• Pairs with a companion site (OPDHub) offering full-text search, multi-axis filters, and copy-ready BibTeX blocks.
• Includes a “Hall of Fame” and model atlas with loss-distribution visualizations for quick landscape scanning.
• Anchors a living survey (V4) that added 72 new papers in the latest revision, plus an AI-trace audit.
• Maintains a related unified framework, EasyOPD, covering 10+ distillation methods with YAML-configurable switching.

Caveats

• The table of contents is sprawling—objectives, signal sources, failure modes, theory—so browsing can feel as dense as reading the survey itself.
• The repository is strictly a curated reading list; ready-to-run training code lives in the separate EasyOPD project.

Verdict

Worth bookmarking if you are building or researching post-training pipelines for reasoning models and need a curated map of on-policy methods. Skip it if you are hunting for ready-to-run training scripts—head to the linked EasyOPD repo instead.