Paper Hub

July 20, 2026 · View on GitHub

A static, zero-cost search engine for top ML conference papers.

顶会论文搜索网站 — 纯静态、零成本、客户端全文检索。

Live site: https://yupu-wang.github.io/paper-hub/

What it does

Index every accepted paper from selected ML conferences and let you search them by keyword, author, conference, year, and presentation type — all in your browser, with no backend.

Currently indexed: 26,579 papers

  • ICLR 2025 — 3,703 papers
  • ICLR 2026 — 5,352 papers
  • ICML 2025 — 3,257 papers
  • ICML 2026 — 6,341 papers
  • NeurIPS 2025 — 5,286 papers
  • AAAI 2026 — 2,375 papers
  • NDSS 2026 — 265 papers (security)

Each result links back to the paper's OpenReview page where you can read or download the PDF.

Features

  • Full-text search across title, abstract, authors, and keywords (powered by MiniSearch)
  • Filters: conference, year, author
  • Presentation tags: oral / spotlight / poster (parsed from OpenReview venue field)
  • Incremental updates: each <conference>-<year> is its own shard — adding a new conference only ships one new file, all existing shards stay browser-cached
  • Zero infrastructure: static HTML/CSS/JS on GitHub Pages, no server, no database, no JS framework

Architecture

[Python scrapers] → data/raw/<conf>-<year>.json     (one per conf+year)
       ↓ build
[Build scripts]   → data/shards/<conf>-<year>.<hash>.json
                    data/manifest.json
       ↓ git push
[GitHub Pages]    serves web/ + data/
       ↓
[Browser]         loads manifest, fetches shards in parallel,
                  builds in-memory MiniSearch index, runs queries

Full design rationale: docs/specs/2026-04-26-paper-hub-design.md

Local development

Requires Python 3.11+ and a Conda environment.

# One-time setup
conda create -n paper-hub python=3.11 -y
conda activate paper-hub
pip install -r scrapers/requirements.txt

# Run tests
make test

# Local preview (creates a symlink web/data → ../data)
make serve
# → open http://localhost:8000

Adding a new conference / year

make update CONF=neurips YEAR=2026
git add data/ && git commit -m "Add NeurIPS 2026" && git push

GitHub Actions auto-deploys to Pages within a few minutes. Browser caches old shards via content-hash filenames, so users only download the new one.

Currently supported sources:

  • ICLR / ICML / NeurIPS via OpenReview API (2023+ for ICML/NeurIPS)
  • Older years (ICML 2021–2022, NeurIPS 2021–2022) require a separate scraper that hasn't been written yet — would target proceedings.mlr.press and papers.nips.cc respectively.

Project structure

paper-hub/
├── scrapers/                # Python data pipeline
│   ├── fetch_iclr.py        #   one fetcher per conference
│   ├── fetch_icml.py
│   ├── fetch_neurips.py
│   ├── build_shard.py       #   raw → shard (with content-hash filename)
│   ├── build_manifest.py    #   regenerates manifest.json from shards/
│   └── common/
│       ├── schema.py        #   pydantic Paper model
│       └── openreview_client.py
├── data/
│   ├── raw/                 # source-of-truth, committed
│   ├── shards/              # served to frontend
│   └── manifest.json
├── web/                     # static site
│   ├── index.html
│   ├── main.js              #   loads shards + builds index
│   ├── search.js            #   query + filter
│   ├── render.js            #   result list + infinite scroll
│   └── styles.css
├── tests/                   # 42 pytest tests
└── docs/
    ├── specs/               # design docs
    └── plans/               # implementation plans

Unified paper schema

Every paper, regardless of source, conforms to:

{
  "id": "iclr-2025-1",
  "title": "...",
  "authors": ["...", "..."],
  "abstract": "...",
  "keywords": ["...", "..."],
  "conference": "ICLR",
  "year": 2025,
  "url": "https://openreview.net/forum?id=...",
  "presentation": "oral"
}

Validated by scrapers/common/schema.py — invalid entries are dropped with a warning during build.

Future ideas (not implemented)

  • AI-generated Chinese abstract translations
  • AI one-line paper summaries
  • Backfill 2021–2024 historical years
  • Add security conferences (CCS / USENIX Security / S&P)
  • Saved/favorites list (browser localStorage)