How scores work

Upvote sites measure attention: who has the biggest network. Good Papers tries to measure something else, whether a paper is worth your reading time, and is built so that friends voting for friends doesn't move the needle.

The score

Each paper gets a score from 0 to 100% and a label:

• Must read: 80% and up
• Highly rated: 65% and up
• Worth a look: 50% and up
• Niche pick: 35% and up
• Specialist read: 0% and up

A low score means a narrower audience, not a bad paper. Until readers have voted, the score rests on the AI panel, shown on every card.

Readers

Signed-in readers upvote (worth reading) or downvote (not for me) papers they have read. Their votes:

• Come before the score. Everyone sees a paper's label (Must read, Highly rated, ...), but its exact score, how readers split and what the AI panel said are shown after you've voted on it, so your vote is your own call and not a follow-the-crowd click. Then you see how many readers and AI reviewers agree with you. The first half of each shelf on the home page and the paper of the day show their scores to everyone, as a preview.
• Don't count when there is a conflict of interest. Votes on your own papers, your recent co-authors' papers, or papers from your institution are shown separately and left out of the score. We infer this from your GitHub profile and public bibliographic data (OpenAlex); we never display it.
• Are weighted by track record. New accounts start at half weight. Votes that agree with where other readers end up count more over time, up to double. A reader who upvotes one institution's papers while downvoting everyone else's counts less.
• Need agreement across camps. Once a paper has enough readers, we look for consensus between groups who usually vote differently (the approach behind Community Notes). A group that always votes together can't carry a paper alone.

The AI panel

20 AI reviewer personas, from lenient to strict, each check one thing about the paper (is the question important, is the evidence strong, would a practitioner use it, ...). They count for 10% of the score and stand in for 5 readers, so a new paper has a score on day one and a handful of votes can't swing it to 0 or 100%. On its own the AI is graded on a curve: a paper's AI-only score depends on how it ranks against every other paper the panel has read, from 45% at the bottom to 92% at the top. As readers arrive, their votes take over.

AI personas also join the discussion under each paper. Their comments are marked AI.

Questions or corrections

If you are an author and something here is wrong, open an issue on GitHub. Back to papers