MapHub turns the daily preprint flood into a shared research heatmap

About ninety new astronomy preprints arrive with each arXiv announcement, according to the researchers behind MapHub. Their proposed response is not to summarize everything, but to show how the daily listing connects to a research community: one row per paper, one column per participant, and a color-coded relevance score at every intersection.

In a November 4 preprint, Kaili Cao, Camille Avestruz and Feige Wang of the University of Michigan describe MapHub, the MIDAS arXiv Preprint Hub. Their early evaluation suggests that its rankings can concentrate papers readers consider especially relevant. The evidence remains limited, however: the authors themselves supplied the human ratings, and the strongest result depends on a small number of highly ranked papers.

Matching interests, not judging quality

MapHub starts with research portfolios of roughly 1,500–3,000 words. These describe participants’ topics, methods, scientific targets and plans. A public prompt can help draft a portfolio from publication records, but participants check and correct it themselves.

The system then uses Claude Opus 5.5 to assess each new paper’s title, authors and abstract against those portfolios. Despite the language of an AI “reading” papers, the described scoring input is this identifying information and abstract—not the complete manuscript.

Its scale runs from zero to 100. A score of 30 means a paper may broaden a participant’s view; 60 means they would cite it; 100 means they wrote it, or would. These are intended as measures of fit to a portfolio, not assessments of scientific quality.

The distinction matters. A low-scoring paper could still be excellent science, while a high score does not establish that its conclusions are correct.

Portfolios remain among participants and the model provider, while scores are public with participants’ agreement. The authors also publish prompts, code and tables, label the scores as AI-generated, and record how each table was produced.

A shorter list to inspect

The preprint reports results from 70 listings containing 11,932 new submissions, scored for one individual, two groups and two collaborations. Across astronomy listings averaging 87 papers a day, about 6% reached a score of 30 for a given participant. Only 0.4% reached 60—roughly one paper every three days.

For the individual, whose interests span multiple subfields, the system flagged about seven astronomy papers per day at 30 or above. Meanwhile, 8% of astronomy papers reached that threshold for at least two participants. Those overlaps are central to the heatmap’s community purpose: they could identify papers worth discussing together and people who might contribute to the conversation.

To reduce computational cost, MapHub first screens batches against short portfolio outlines, then rescores promising matches against full portfolios. On two test listings totaling 748 papers, the authors report that this approach cost about 36% of a full scoring pass and missed none of the 56 paper–participant matches that the full pass scored at least 40. That is a result from those test listings, not a guarantee for future submissions.

How well did the rankings work?

The authors conducted a blind rating exercise using a shared selection that included 45 astronomy papers, with five shown twice, and ten AI papers. Selection favored higher-scoring papers; the analysis weighted ratings to account for those unequal sampling probabilities. Participants’ own papers were excluded, and the portfolios cited none of the rated papers.

In the pooled astronomy results, MapHub’s highest-ranked 10% contained about 70% of the papers in the raters’ own top 5%. A term-frequency text-matching baseline captured about 33% when inspecting the same share. For the broader target of each rater’s top 15%, MapHub’s advantage was smaller, and the methods converged after roughly 40% of the listing was inspected.

Several caveats constrain that comparison. The first author built MapHub and rated three of its five columns. Human ratings were not a gold standard, the authors acknowledge, and could be influenced by previously seen papers. The top-5% result rests on few papers; percentile-based evaluation also assumes every rater has something of interest in the listing.

Publication-based portfolios can additionally miss emerging interests. MapHub is therefore best understood as an early, inspectable aid to choosing what to read and discuss—not a replacement for reading papers or deciding their scientific value.