paper-reads

A daily archive of structured close readings of trending machine-learning papers, written in both Chinese and English, and published as a static site on GitHub Pages.

Live site: https://yingwang.github.io/paper-reads/

Format

Each note follows the same four-section structure:

• The Problem — what the paper sets out to solve, why it matters in practice, and where prior approaches fall short.
• Method — the core design and the key technical components, with every symbol and equation explained in the order it first appears.
• Experiments — datasets, baselines, headline numbers, and the ablations that carry the most signal.
• Limitations — what the authors acknowledge directly, and what a careful reader can identify beyond the paper's own framing.

Notes are written in Chinese first and then translated into English paragraph-for-paragraph, so the two versions remain aligned in structure and level of detail.

Each note opens with a Reading Guide section that places the paper within its subfield, lists the concrete questions the reader will be able to answer after the read, declares the prerequisite knowledge assumed of the reader, and provides a glossary of every acronym used in the remainder of the note. Wherever structure benefits the reader, the note includes mermaid diagrams of architectures, training pipelines, and method comparisons.

The audience is a machine-learning engineer with general background, not a specialist in the paper's subfield. Technical terms receive a brief introduction on first use, and the prose deliberately avoids LLM-flavored closers ("this work demonstrates that...", "in conclusion", "to summarize") in favor of letting the facts speak.

Repository layout

paper-reads/
├── docs/
│   ├── index.md                          # Chinese landing page
│   ├── en/
│   │   ├── index.md                      # English landing page
│   │   └── papers/                       # English close readings
│   ├── papers/                           # Chinese close readings
│   │   ├── YYYY-MM-DD-<slug>.md
│   │   └── ...
│   └── javascripts/mathjax.js
├── mkdocs.yml                            # MkDocs Material configuration
└── .github/workflows/deploy-docs.yml     # Build and deploy pipeline

Each paper appears under both docs/papers/ and docs/en/papers/ with the same filename. The MkDocs navigation lists every note directly under the top-level 中文 and English tabs, sorted by date with the newest entry first.

Automation

The notes are generated by the paper-read skill in yingwang/claude-skills. The skill is invoked in two modes:

/paper-read --trending 1                       # close-read today's top trending paper
/paper-read https://arxiv.org/abs/2401.12345   # close-read a specific paper

A daily cron at 08:47 (local time) runs /paper-read --trending 1 automatically. After the Chinese close reading is drafted, an English translation is produced from the Chinese version paragraph-for-paragraph, both files are copied into the corresponding directories in this repository, the navigation in mkdocs.yml is updated, and the commit is pushed. GitHub Actions then builds the static site and deploys it to GitHub Pages.

Local build

pip install \
    "mkdocs-material>=9.5" \
    "mkdocs-awesome-pages-plugin>=2.10" \
    "pymdown-extensions>=10.7" \
    "mkdocs>=1.6" \
    "mdx-truly-sane-lists>=1.3"

mkdocs serve

The local site is served at http://localhost:8000/paper-reads/.

License

The prose in this repository is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. See LICENSE for details. Figures, equations, and excerpts quoted from the papers being summarized remain the property of their original authors and are reproduced for the purpose of educational commentary.