Awesome-LLM-OPD: A Visual Knowledge Graph for On-Policy Distillation Papers

Awesome-LLM-OPD is a searchable online graph that contains over 175 On-Policy Distillation papers, accompanied by the ICML 2026 survey paper 《A Survey of On-Policy Distillation for Large Language Models》 (arXiv:2604.00626). It provides intelligent retrieval, categorized browsing, and trend analysis functions, aiming to address the pain points of traditional static paper lists.

Access link: https://nick7nlp.github.io/awesome-llm-opd/

The field of large language model research is developing rapidly, with the OPD subfield accumulating over 175 papers in just a few years. Traditional static paper lists (Awesome Lists) have four major pain points: difficulty in quickly locating papers on specific methods, inability to intuitively show domain trends, lack of cross-paper correlation analysis, and high update and maintenance costs. Awesome-LLM-OPD was created to address these issues.

1. Intelligent Fuzzy Search: Implemented based on Fuse.js, supporting retrieval of title keywords, author names, and method components, including spelling tolerance and approximate matching;
2. Combined Filtering System: Multi-dimensional cross-filtering (chapter classification/loss function/publication year), allowing precise location of papers such as "2024 papers using FKL loss";
3. Visual Display: Provides model graph heatmaps, loss function distribution charts, monthly evolution trends, method timelines, etc.;
4. Daily Automatic Updates: A 7-stage pipeline (pre-check → reconnaissance → deep reading → filtering → update list → refresh index → loss classification → website refresh) ensures new papers are searchable within 24 hours.

• Pure Static Design: Uses HTML + Bulma CSS + Fuse.js architecture, hosted on GitHub Pages, with zero operation and maintenance costs, fast loading, and high availability;
• Data Pipeline: Integrates 4 types of data sources (paper_notes.json / loss classification json / chapter grouping README / paper abstract tex), and renders via arXiv ID association;
• Idempotent Design: Repeated pipeline runs have no side effects, ensuring data consistency.

Academic Value: The accompanying survey paper provides a systematic domain overview, unified mathematical framework, experimental comparisons, and future directions; the website serves as an online companion to the paper. Citation format must include arXiv:2604.00626.

Application Scenarios:

• Researchers: Literature research, method comparison, trend analysis, paper tracking;
• Engineers: Technology selection, implementation reference, performance benchmarking;
• Students: Introductory learning, paper reading, direction exploration.

Current Limitations: Covers only the OPD field, relies on arXiv/Semantic Scholar data sources, and automated classification may have errors.

Future Directions: Expand to other distillation methods such as Offline Distillation, add paper influence metrics (citation count/Stars), support user favorites and note-taking functions, and introduce a paper correlation knowledge graph.

Awesome-LLM-OPD is a new paradigm for academic resource management, shifting from static lists to dynamic graphs and from manual maintenance to automatic updates, providing an efficient tool for large model distillation researchers.

Contribution Guidelines: Do not directly edit the automatically generated index.html or papers.json; to add/correct papers, submit an Issue/PR to the Awesome-LLM-On-Policy-Distillation repository; for website template issues, submit a PR directly to this repository.