paper_notes

[English | 日本語 ]

This repository is a human-curated collection of AI research papers and related resources, including blog posts. The maintainer follows discussions on X every day to keep up with developments in AI. Each item is managed as a GitHub Issue.

• Curated from What's Being Discussed on X: The collection focuses primarily on papers and resources that attract broad interest on X, while the maintainer's own interests also influence what gets added.
• Tagged for Discovery: Tags are manually assigned to reflect the maintainer's interpretation of each item. Filter by tag to find related work and follow a topic across issues.
• Notes Based on the Maintainer's Understanding: Some items include notes in Japanese reflecting the maintainer's understanding of the X posts and papers. These notes are tagged by content and focus.

All content stored in GitHub Issues is exported daily as XML files under agent_docs/. An AI agent can use these files to answer questions across the collection. See Conversational Interface.

The primary purpose of this repository is to serve as a personal study log and memory aid for the maintainer.

---

Issue Contents

Each issue mainly contains the following information:

• 🔗 Links to papers, blog posts, or related resources
• 📄 Paper metadata and abstracts
• 🌐 Japanese translations and summaries of abstracts
• 📝 Notes based on the maintainer’s understanding
• 🏷️ Manually assigned labels reflecting the maintainer’s interpretation
• 📚 Useful external references

Paper metadata and abstracts are retrieved via the arXiv API[^1]. Since the maintainer is a native Japanese speaker, Japanese translations and summaries of abstracts are automatically generated using GPT.

---

Actively Collected Research Areas

The repository currently focuses on the following four areas:

• Natural Language Processing (NLP)
• Computer Vision (CV)
• Speech Processing
• Recommender Systems (RecSys)

---

Topics of Recent Interest

Recent areas of focus include (but are not limited to):

• Large Language Models (LLMs) / Vision-Language Models (VLMs) / Diffusion Models
• Proprietary / Open-Weight / Open-Source Models
• Multimodal Models / Unified Multimodal Models (UMMs)
• AI Agents / Context Engineering / Memory / Evaluation Harnesses / Scaffoldings
• (Synthetic) Datasets / Evaluation / Benchmarks
• Pre-training / Mid-training / Post-training / Continual Learning
• Reinforcement Learning
• Test-time Scaling / (Memory-based) Test-time Learning / Test Time Training (TTT)
• Representation Learning / Embeddings
• Self-Improving AI / Recursive Self-Improvement (RSI)
• Transformer architectures (Attention, Positional Encoding, Residual Streams, etc.)

---

Issue Notes and Granularity

Each Issue may contain:

• The maintainer’s understanding or interpretation of the work [^2]
• Initial impressions
• Links to explanatory SNS posts or discussions

The level of detail varies depending on available time and interest at the time of reading.
Notes are managed using the following labels (not all Issues have been fully labeled yet because the maintainer have manually labeled it):

• : first impressions based on the abstract, related posts, or skimming
• : a single key takeaway
• : a broad summary of the paper
• : a notes on multiple important technical points
• : notes containing the maintainer's random thoughts, opinions and impressions after reading
• : details of the paper and multiple important technical points with the maintainer's own opinion
• : a collection of related links (e.g., explanatory posts)
• : Explanatory Posts and Threads by the Authors

---

Label Taxonomy

Labels are broadly categorized and color-coded as follows:

• Medium: , , , ...

• Research Field: , , , , ...

• Task: , , , ...

• Method: , , [``](https://github.com/AkihikoWatanabe/paper_notes/issues?q=is%3Aissue%20state%3Aopen%20label%3AReinforcementLearning), ...

• Fine-grained Category: , , , ...

• Properties / Characteristics: , , , ...

• Conference / Journal: , , , ...

The above labels are just a few examples, and there are currently 600+ different labels in total. You can filter issues using these labels.

---

Wiki: Personal Trend Notes

A Wiki is maintained to record personal impressions of trends observed while casually browsing papers and discussions on X.

• These notes reflect subjective observations and interpretations
• They may contain inaccuracies or biases
• Updates are irregular and informal

The Wiki is intended as a lightweight place to capture emerging ideas, trends, and intuitions that may not yet be fully formed.

Conversational Interface

Recommended: Clone the repository locally and use an agent to chat with the repository based on the contents of its issues.

git clone https://github.com/AkihikoWatanabe/paper_notes.git
cd paper_notes

Here is the example from VSCode + Codex (GPT-5.6-luna, reasoning_effort=xhigh):

Please summarize the pretraining of large language models in chronological order. Explain it while citing the references you used and the original issue links.

The agent can use the repository's AGENTS.md instructions and agent_docs/ to answer questions based on the collected Issues and study notes. The agent_docs directory contains daily dumps of the issue contents.

In addition, the repository can be explored through the following complementary interfaces.

---

DeepWiki/CodeWiki

This repository can also be explored via DeepWiki/CodeWiki, which provides a chat-based interface for interacting with the contents of this repository.

Using DeepWiki/CodeWiki, it may be easier to:

• Interactively extract information from Issues and notes
• Ask questions about specific papers, topics, or trends
• Navigate the repository in a conversational manner rather than browsing Issues manually

Here is an example query, run in codemap mode on DeepWiki:

Survey the literature on pretraining.

It returns a structured document explaining the literature on pretraining, based on the materials accumulated in this repository, with references to their sources.

---

Updates

2026.10.14

• Fixed an issue that had prevented agent_docs from being updated since the September 19 update.

2026.09.19

• Implement full text search in the blog.
• The repository history was rewritten to remove generated files under docs/, including previously committed static site artifacts.
  • Commit hashes before the rewrite are no longer valid.
  • Existing clones should be deleted and freshly cloned.
  • The generated website is now built and deployed through GitHub Actions as a Pages artifact instead of being committed to the repository.

2026.03.07

• For agentic use cases, I have added AGENTS.md and agent-specific documentation to the agent_doc directory. This prevents AI agents from wasting context on unnecessary HTML strings.

2026.01.20

• Update README.md

2025.10.17

• I created a Wiki to jot down personal impressions of trends observed while casually browsing papers on X.
• These are purely subjective notes and may contain inaccuracies. Updates will be irregular and informal.

2025.08.05

• At present, the most convenient way to use this repository may be to extract knowledge interactively via its DeepWiki interface.

2025.07.19

• The following improvements were made:
  • Fixed degraded blog response time
  • Updated displayed blog content
  • (A full redesign is still pending 🫠)

2024.12.28

• Several months have passed since I first planned to redesign the blog.

2024.09

• I reorganized Issue notes into a blog-style format.
• The blog is available here.

2022–Early 2023:

• Beginning to Systematically Track AI Research on Twitter (X).

2017.12.28

• I started this repository and created the first issue.

---

Star History

[^1]: Metadata and abstracts are used in accordance with the CC0 1.0 license specified in the arXiv API terms of use.
[^2]: Images and screenshots included in Issues or blog posts are, in principle, cited from the original sources (papers, articles, slides, etc.) linked within each Issue.