The OPD field guide: 414 papers and a decision tree

What it does

This repository is an opinionated bibliography of on-policy distillation (OPD) research. It catalogs 414 papers, technical reports, frameworks, and tools that explore training student models on their own trajectories while a teacher provides dense, token-level supervision. The list spans foundational theory, failure-mode analyses, and industrial recipes from labs like DeepSeek, Alibaba, and NVIDIA.

The interesting bit

Rather than dumping links, the README structures the field as a practitioner’s decision tree: do you have teacher logits (white-box), API-only access (black-box), or no teacher at all (self-distillation)? It also treats OPD as a mainstream post-training primitive, not an academic curiosity, citing its use in shipping models such as Qwen3, DeepSeek-V4, and GLM-5.

Key highlights

• 414 entries organized from foundational papers (MiniLLM, GKD) to production technical reports
• “Start Here” section with a fast-path syllabus and a white-box/black-box/self-distillation decision framework
• Explicit coverage of failure modes: instability, diversity collapse, tokenizer mismatch, and horizon coverage decay
• Industrial recipes from Qwen3, DeepSeek-V4, MiMo, Nemotron-Cascade 2, and GLM-5
• Dedicated section for frameworks and implementations, including TRL’s DistillationTrainer

Caveats

• This is a reading list, not a framework; you will still be reading papers and wiring code yourself
• The document is sprawling and assumes fluency in RL, knowledge distillation, and tokenizer arcana
• Many entries link to X threads and arXiv preprints that may shift or disappear over time

Verdict

Worth bookmarking if you are an LLM post-training engineer or researcher trying to navigate the OPD literature without reading 400 abstracts. Skip it if you need a single, pip-installable distillation library.