nick7nlp/Awesome-LLM-On-Policy-Distillation

A curated collection of papers and resources on On-Policy Distillation for Large Language Models.

What it solves

Researchers and practitioners often struggle to navigate the rapidly growing body of work on on-policy distillation for large language models (LLMs) — a training technique where a student model learns from a teacher using the student's own generated text. This project organizes 362 relevant papers (134 with code) into a structured, searchable collection with concise summaries, topic groupings, and guided reading paths, making it easier to find the right paper, understand key innovations, and identify code for implementation.

How it works

The project acts as a reading companion to the survey "A Survey of On-Policy Distillation for Large Language Models" (available on arXiv). It groups papers into thematic sections such as "Objectives and update rules", "Where does supervision come from?", and "How multiple agents learn". Each entry includes a one-line takeaway, publication date, a category tag (e.g., Direct OPD, Hybrid OPD), and links to the paper and code when available. The companion offers multiple entry points: you can browse all entries, expand a topic to read summaries directly, or follow curated reading paths that lead from foundational papers to advanced variations. An interactive website (OPDHub) additionally supports searching and filtering across the collection.

Who it’s for

This is aimed at LLM researchers, graduate students, and engineers working on model distillation, alignment, or post-training. It is especially useful for anyone who wants to quickly survey the state of the art, compare different distillation objective designs, or find reusable code for their own experiments. The reading paths also help newcomers start with core papers before diving into niche variants.

Highlights

  • 362 papers catalogued, 134 with code, updated as of 2026-10-13
  • Papers organized into thematic topics with crisp one-line summaries under each
  • Curated reading paths that answer specific questions, e.g., "How does the student learn?" or "What transfers and what does not?"
  • Every entry links to the arXiv paper, and many include direct code links
  • Full-text search and filtering available through the companion website OPDHub

Related

  • Project
    fudan-generative-vision/Remory

    Remory is a plug‑in that adds a learned “residual memory” to large language models, enabling them to compact long conversation histories while preserving important information. It works with frozen base models (Qwen 3.8‑27B or GLM‑5.3‑Flash) and provides a SGLang server, a Codex UI integration, and a Python API for summarising, encoding residual memory, and continuing generation from a compacted checkpoint.

  • Project
    cropflre/nowen-note

    A self-hosted, private knowledge base and task management platform that features AI RAG capabilities, MCP support for AI agents, and multi-device synchronization.

  • Project
    ganarajpr/awesome-dspy

    A curated list of DSPy resources — projects, tutorials, articles, videos, and papers — for building self-improving language model pipelines with DSPy.

  • Project
    changwookjun/nlp-paper

    A curated list of natural language processing research papers, organized into categories from BERT and Transformers to large language models and retrieval-augmented generation. It helps researchers quickly find seminal and recent works in specific NLP areas.