How PaperRank Decides What to Read First
PaperRank is Feynman’s clearest example of an inspectable research decision. Its question is narrower than “Which paper is best?” and more useful in practice: given a bounded set of candidate papers, which one deserves attention first, and what evidence explains that order? Understanding PaperRank means learning to read a score as a proposed allocation of attention—not as a verdict about scientific truth. The score, local citation graph, evidence spans, sensitivity tests, and follow-up records together show how Feynman turns prioritization into something a researcher can inspect and revise.
The decision is read-first triage
A literature search can return more plausible papers than a person can read immediately. A ranking system is therefore valuable even when it cannot settle the underlying science. It can reduce the first decision from “read everything” to “start here, for these visible reasons.” That is the decision PaperRank is designed to support.
The distinction controls how every later signal should be interpreted. If a paper ranks highly because its title and abstract closely match the query, that is evidence of topical fit. If it ranks highly because it sits in a well-connected local citation neighborhood, that is evidence of bibliometric context. If its available text contains markers of experiments, baselines, limitations, code, or data, that is evidence that the paper may be easier to inspect. None of those observations alone proves that the paper’s central claim is sound.
Feynman’s implementation makes the separation explicit by combining six signals: topical relevance, citation impact, local graph prestige, citation velocity, methodology quality, and reproducibility. The default weights give the largest share to topic match, followed by citation impact and local graph position, with smaller contributions from velocity, methodology, and reproducibility. Those weights are not presented as a universal theory of paper quality. They are a transparent starting hypothesis that a researcher can examine against their own reading priorities.
That transparency matters more than the existence of a single number. A black-box rank can be convenient but difficult to challenge. PaperRank retains the component signals, their availability, confidence, explanations, and supporting evidence. The researcher can therefore ask not only “What is the rank?” but “Which assumptions produced it, and which of those assumptions should I trust for this question?”
Six signals, six different questions
Topical relevance asks whether the candidate resembles the question being asked. PaperRank uses token overlap among the query and the paper’s available title, abstract, concepts, and topics, with contributions from title hits and the source result position. This is a practical approximation to relevance, not a semantic judgment of the paper’s contribution. It works best when the query vocabulary and the paper’s metadata share terms; it becomes less informative when a field uses synonyms, when abstracts are missing, or when the most important idea is not named in the title.
Citation impact asks how much attention a work has received relative to comparable works. When OpenAlex supplies a normalized citation percentile, PaperRank prefers it over a raw count. When that percentile is unavailable, the implementation falls back to a candidate-local logarithmic transformation of citation count. The fallback avoids allowing a very large count to dominate linearly, but it remains a property of the candidate set and the available metadata. Citation impact is influence evidence, not evidence that the cited work is methodologically reliable.