About PaperDance

PaperDance (纸上跳舞, "dancing on paper") is a free web app for browsing arXiv papers the way you browse short videos. Describe your topic in plain words and swipe through full-screen cards, each with the paper's key figure, a TL;DR, the abstract, affiliations, first and corresponding authors, venue badge and citation count. Like, save, comment, and export your library as BibTeX, Markdown, CSV or JSON. Open paperdance.org in any browser; nothing to install.

What it does

• Describe a topic in natural language (English or Chinese), e.g. "video generation, especially controllability and faster inference". No categories to pick, no query syntax.
• One full-screen card per paper: key figure pulled from the paper, TL;DR and abstract, affiliations, first and corresponding authors, venue badge, citation count and tags. Chinese cards add a one-line takeaway and three bullets.
• Cards in English or Chinese; density from "figure + TL;DR" to "full abstract"; adjustable font size.
• Like, save, comment, or "not interested": bookmarks that also shape your ranking. Works without an account; sign up with an email to sync across devices.
• Export liked or saved papers from "My library" as BibTeX, Markdown, CSV or JSON, ready to hand to an agent such as Claude Code or Cursor for deep reading.
• Every paper has a stable link, paperdance.org/p/<arXiv id>, with title and figure previews on X, WeChat and other apps.

Data sources and coverage

Metadata comes from arXiv (CC0) for cs.AI, cs.LG, cs.CV, cs.CL, cs.RO, cs.IR, cs.SD, eess.AS and eess.IV: 249,000+ papers from September 2024 to today (latest published 2026-09-09), updated twice a day right after arXiv announcements.

Citation counts, venues and English TL;DRs come from Semantic Scholar. Figures are taken from the official arXiv HTML version of each paper. Affiliations are extracted from the author block and shown only when the paper itself supports them.

How the cards are made

Chinese cards are generated by a large language model (Qwen) under anti-fabrication rules: it may only restate what the abstract says, never add numbers or conclusions. Each paper is summarised once and shared site-wide.

The original English abstract and the arXiv link are always one tap away; when in doubt, the paper is the source of truth.

How ranking works

Your topic is rewritten into retrieval keywords and matched with both full-text search (BM25) and embedding similarity, with time decay on top. Likes, saves and "not interested" update your profile Rocchio-style, anonymously too. A couple of "explore" cards from neighbouring topics are mixed into every batch so the feed does not narrow over time.

Four ranking preferences are adjustable: recency (newest / fresh / any), topic fit (strict / balanced / broad), paper quality (all / accepted / highly cited / classics) and novelty (stable / balanced / surprise).

Accounts and privacy

No sign-up needed to swipe, like or save; those live in your browser first. Register with an email to keep them in an account and sync your topic across devices.

Only recommendation-related events are recorded (views, skips, likes, saves, comments) and used to improve your own feed. No ads, no selling data.

FAQ

Is PaperDance free?

Yes, completely free. No paywall, no invite code.

Do I need to install an app?

No. Open paperdance.org in any phone, tablet or desktop browser; on mobile you can "Add to Home Screen" to use it as a PWA.

Where do the papers come from and how many are there?

From arXiv, categories cs.AI, cs.LG, cs.CV, cs.CL, cs.RO, cs.IR, cs.SD, eess.AS and eess.IV: 249,000+ papers from September 2024 to today, updated twice daily. Citations and venues come from Semantic Scholar.

Are the summaries accurate?

Chinese cards are LLM-generated under anti-fabrication rules and only restate the abstract; English cards show the Semantic Scholar TL;DR and the abstract. The original abstract and arXiv link are always available.

Can I use it in English?

Yes. Interface language and card language are set separately; English cards show the key figure, TL;DR, abstract, affiliations, venue and citations.

How do I tell it what I want?

Type one sentence describing your topic, optionally add sub-topics, and set the four ranking preferences: recency, topic fit, paper quality and novelty.

How do I export my saved papers?

From "My library", export likes or saves as BibTeX, Markdown, CSV or JSON in one click.

How do I share a paper?

The "Share" button on the right rail gives you paperdance.org/p/<arXiv id>. Opening it pins that paper first, followed by the normal feed.

Why do some papers have no figure?

Figures come from the official arXiv HTML version. About a quarter of papers have no HTML version, so those show a text-only card.

Is PaperDance affiliated with arXiv?

No. It is an independent project, not an official arXiv product. Metadata comes from arXiv under CC0. Thank you to arXiv for use of its open access interoperability.