Contents
You find an intriguing AI paper, save its title and link, and plan to read it later. A few weeks later, when you start researching a similar topic, you cannot remember where you saved it—or what caught your attention.
Akihiko Watanabe’s “paper_notes” is a repository designed to reduce those moments when you know you have read something but cannot put it to use in your next research task. It manually collects AI papers, blog posts, code, slides, social media posts, and other materials in GitHub Issues. More than a list of paper links, it is an evolving research log that records the maintainer’s own understanding and insights.
The collection covers natural language processing, computer vision, speech processing, and recommender systems. Recent interests span LLMs, VLMs, diffusion models, AI agents, evaluation, reinforcement learning, and more. The accumulated issues are exported daily to agent_docs/, and the repository also supports having an agent read its contents and answer questions.
Start by opening the GitHub repository ↗ and browsing the issues and README to get a sense of what it contains.
Adding a Reader’s Understanding to Paper Links#
Each issue includes some combination of links to papers and related articles, paper metadata and abstracts, Japanese translations and summaries, the maintainer’s own notes, and manually assigned labels. Some also include links to related external articles or discussions. The arXiv API is used to retrieve paper information and abstracts, while GPT automatically generates Japanese translations and summaries.
The issues are not all organized at the same level of detail, however. Some contain only first impressions based on a quick look at an abstract or related posts; others explore several technical points in depth. As the README explains, the level of detail varies with the maintainer’s interests and available reading time.
That variation makes sense for a personal learning log. Leave a short note when you discover a paper, then expand it when time allows. The aim is to record your understanding at that moment, rather than produce an equally polished explanation of every paper. Readers can use the amount of detail as a guide and consult the paper or original article when needed.
The maintainer also assigns labels based on their own interpretation. Issues can be filtered along several dimensions, including publication format, research field, task, method, detailed category, characteristics, and conference or journal. There are more than 600 labels, and some have not yet been applied to all relevant issues. These categories provide useful starting points, but neither their completeness nor their consistent application is automatically guaranteed. When researching a particular topic, it is best to look at related issues and original papers as well, rather than draw conclusions from label searches alone.
Turning GitHub Issues into Materials You Can Query Later#
One distinctive feature of paper_notes is that it makes its accumulated issues accessible to agents. The README explains how to clone the repository locally and ask an agent questions. Issue contents are exported daily to agent_docs/, and AGENTS.md and other agent instructions help agents answer from the research notes without loading large amounts of HTML.
For example, you can ask for a chronological overview of LLM pretraining, complete with links to the sources consulted and the original issues. This lets you build a foundation for research across multiple issues, instead of opening saved notes one at a time. It is best used while checking the cited papers and issues, rather than treating the answers as established facts.
The basic steps are simple.
- Clone the repository.
- Open it with an agent that can access local files, such as Codex.
- Review the guidance in
AGENTS.mdandagent_docs/. - Specify the topic you want to investigate and the sources you want included in the answer.
- Open the issues and papers cited in the answer to verify the relevant details.
For example, instead of simply asking, “Summarize pretraining chronologically,” ask, “Arrange the papers by publication date, and include links to the papers you used as evidence and their original issues.” This makes the answer easier to check later. The key is to ask questions within the scope of the information collected in the repository. It is not a system for exhaustively researching papers that have never been recorded there.
Alongside reading issues directly on GitHub, the repository introduces ways to explore its contents through conversational interfaces such as DeepWiki and CodeWiki. These offer a convenient starting point for finding material, after which you can move to the issues or original papers for a closer read.
Read Personal Trend Observations Separately from Paper Notes#
The repository also has a Wiki, where the maintainer records personal impressions of trends noticed while browsing papers and discussions on X. Updated irregularly, it is a place for brief notes on ideas that are still taking shape. It explicitly acknowledges that these notes may contain errors or bias.
This serves a different purpose from the issues that collect paper metadata and summaries. It can offer interesting clues for thinking about the direction of research, but it should not be treated as an objective survey of an entire field. If an observation catches your attention, follow the related papers and discussions to assess it yourself.
This distinction is also useful when questioning an agent. Stating whether you want a synthesis of paper contents or a summary of personal trend observations makes your expectations clearer. For example, “Summarize the development of this research based on the collected issues” calls for a different answer from “Summarize the personal impressions of trends recorded in the Wiki.”
What the Update History Tells Us About Using It Today#
According to the update history, AGENTS.md and documentation for agents were added in March 2026. The recorded update date is March 7, 2026—not July. On September 19, 2026, full-text search was added to the blog, and the repository history was rewritten to remove the generated docs/ directory. Old commit hashes became unusable, and users were instructed to delete existing clones and clone the repository again. The site now builds through GitHub Actions and is published as a Pages artifact.
Later, on October 14, 2026, a fix addressed an issue that had prevented agent_docs/ from updating since the September 19 change. If you want to clone the repository and try using it with an agent, follow the latest instructions in the README. In particular, if you are still using an older clone, check the guidance about the history rewrite first.
The blog view of the issues was introduced in September 2024, and the blog now supports full-text search as well. You can choose between browsing GitHub Issues, searching the blog, and asking an agent questions across the collection, depending on what you need. For a close reading of one paper, turn to the original paper and its individual notes; for a starting point on a broader topic, use search or a conversational interface.
Bringing It into Your Own Research#
paper_notes is not a comprehensive database of all AI research. Its contents are manually collected according to one person’s interests, and the labels and depth of the notes vary. Even so, bringing together paper links, Japanese reading notes, and related materials makes it a useful starting point for discovering unfamiliar research and revisiting things you have read before.
Start by opening two to four issues in a field that interests you, and look at the paper links, the level of detail in the notes, and how labels are used. Then try the blog’s full-text search or ask an agent a question to see how the collection fits your research process. To try it locally, make a fresh clone of the repository, follow the README, and ask an agent one small question. From there, follow the cited issues to start putting the accumulated research notes to use.
Source: https://github.com/AkihikoWatanabe/paper_notes
Contents