Given the papers you've read, which unread papers should you read? That's a named problem — vertex nomination — from the random-graph statistics literature. The principle: a candidate is relevant if it sits close to any paper you've already read. We score each candidate by its nearest-read cosine in embedding space (the simplest consistent member of the spectral-nomination family, robust to a reading list that spans several distinct threads) and reveal the ranking in order. The slider above is a window onto this list; the rows below are the same list, with a relative-relevance score and the paper that pulled each one in.

We also give well-known work a gentle lift: the ranking blends that cosine relevance with a small, log-scaled citation term, so discovery leans toward papers worth knowing — without letting citation counts override relevance (the most relevant paper still wins; citations only lift papers of comparable relevance).

Aim it at one paper — or several. Click any node to focus discovery on that paper; the slider then nominates the unread papers closest to it in particular (its own Semantic Scholar neighbours, cosine-ranked and tethered to it). Shift-click more papers to build a selection, and the slider finds the papers closest to all of them at once (ranked by their similarity to the farthest one selected, so a paper has to be near every pick). Press Escape, click empty space, or close the panel to return to your whole set.

Hover any node and the graph shades every other paper by predicted relatedness:

w(i, j) = 0.65 · cos(eᵢ, eⱼ) + 0.25 · shared-references + 0.10 · direct-citation

The embedding cosine is the only term that's always defined, so it sets each paper's nearest neighbour; the citation terms can only sharpen the ranking, never invent it. One wrinkle worth being honest about: SPECTER2 cosines between in-field papers sit in a narrow, high band — everything is somewhat similar — so the structure comes from k-nearest-neighbours (scale-invariant) and the hover shading is stretched across the range you're actually looking at, not an absolute scale.

Papers, embeddings, and recommendations from Semantic Scholar. Papers you add persist in your browser; export writes a papers.json you can commit to make them part of the public graph.