EfficientPaper

EfficientPaper is a MkDocs-based paper collection site for efficient AI research. It currently indexes 513 papers on Pruning, Quantization, KV Cache, Speculative Decoding, Efficient Inference/Training, and related system optimization topics.

The main experience is the Home page at docs/index.md: a searchable paper workspace with local editing, graph navigation, arXiv import, PDF lookup, and GitHub sync.

What the UI Supports

1. Searchable paper workspace

The Home page combines search, filters, stats, and paper actions in one place.

Keyword search supports plain terms, quoted phrases, AND, and negative terms like -kv.
Filters include year, venue, keyword, rating, and sort order.
Result cards support selection, copy/share, note jumping, graph jumping, local PDF lookup, and quick rating.
A right-side detail drawer shows cover, authors, institutions, tags, note preview, and paper links.

2. Interactive method graph

The graph page links baseline methods and derived work, and can jump back to the corresponding paper on Home.

3. Local paper metadata editor

The Home page Edit action opens an embedded editor modal backed by docs/edit.html. The editor supports .prototxt metadata including title, abbreviation, venue, authors, institutions, keywords, code URL, rating, and update time. docs/edit.html remains available as a standalone editor for local use.

The same editor also supports cover upload, preview, baseline method linking, save, and guarded deletion.

4. Add papers from arXiv

When the local editor server is running, the Home page can search arXiv by ID, inspect the detected paper, and create a new paper entry with an optional custom abbreviation.

5. Upload local changes to GitHub

The site can trigger the local refresh and upload flow from the browser. This is intended for local use and depends on the editor server.

Quick Start

1. Clone and install dependencies

git clone https://github.com/hustzxd/EfficientPaper
cd EfficientPaper
pip install protobuf==5.27.2 pandas arxiv openai mkdocs mkdocs-glightbox mkdocs-literate-nav mkdocs-macros-plugin watchdog

If protoc is not installed on your machine, install Protocol Buffers first.

2. Optional MiMo API key

Adding papers can call Xiaomi MiMo LLM (mimo-v2.5) to auto-generate Chinese summaries and keyword suggestions:

export MIMO_API_KEY="your-mimo-api-key"

If this variable is not set, paper creation still works, but auto summarization and keyword suggestion are skipped.

3. Start the local site and editor server

./start_editor.sh

This script will:

regenerate derived data with refresh_and_upload.sh
start MkDocs at http://localhost:8000
start the editor API at http://localhost:8001
watch meta/ and notes/ for changes and auto-refresh generated data

Typical Workflow

Add from arXiv ID or UI

./add_paper_info.sh 2512.01278v1

This wraps scripts/add_paper.py, looks up the paper by arXiv ID, and creates a new .prototxt paper entry plus a note directory under notes/<year>/<paper_id>/.

You can also open the Home page and use Add from arXiv when the local server is available.

Edit metadata and notes in browser

Visit http://localhost:8000
Use the paper card Edit action to open the metadata editor in a modal
Open the paper note page to edit notes/<year>/<paper>/note.md in browser

Refresh generated assets

./refresh_and_upload.sh

This regenerates README/about pages, protobuf templates, split metadata, graph data, and the search dataset.

Commit, push, and deploy

./refresh_and_upload.sh "update_paper_info"

With a commit message, the script additionally runs:

git add .
git commit -m ...
git push
mkdocs build
./build_and_deploy.sh

Repository Layout

readme_raw.md                       # Source for README.md and docs/about.md
docs/index.md                        # Main searchable home page
docs/baseline_methods_graph_interactive.md
docs/edit.html                       # Local metadata editor
docs/js/papers.json                  # Frontend search dataset
docs/js/paper_graph_map.json         # Home <-> graph mapping
meta/<year>/*.prototxt               # Structured paper metadata
notes/<year>/<paper>/note.md         # Paper notes
notes/<year>/<paper>/cover.*         # Paper cover assets
scripts/paper_editor_server.py       # Local editor / upload / pull / PDF API
scripts/generate_readme_pages.py     # README/about generator
scripts/generate_search_data.py      # Search dataset generator

Local-server-aware Features

Several UI actions depend on http://localhost:8001 and are intentionally disabled when the server is unavailable:

Add from arXiv
Upload to GitHub
Pull from GitHub
Set PDF Path
card-level PDF
card-level Edit
card-level delete

This graceful degradation is part of the intended local workflow.

Contributing

To add or update a paper:

Run ./add_paper_info.sh <arxiv_id> or use Add from arXiv.
Start the local tools with ./start_editor.sh.
Edit metadata, note content, cover image, keywords, and baseline links in the browser.
Run ./refresh_and_upload.sh to regenerate derived data.
Submit a Pull Request, or use the local GitHub upload flow if you are maintaining your own deployment.

Conference Timeline

招聘

如果您对论文涉及到的研究内容感兴趣，同时有求职意向（实习生/校招/社招），可以发送简历到 zhaoxiandong27@gmail.com，欢迎沟通交流。