GalaxyWiki finds what papers actually say about individual galaxies

Finding a paper that mentions a galaxy is not the same as finding an explanation of what makes that galaxy interesting. An object might appear among thousands of catalog entries, with no accompanying interpretation. Astronomical databases such as NED and SIMBAD provide object bibliographies, but those links alone do not distinguish a detailed discussion from a table entry.

In an arXiv preprint posted on November 9, 2026, Nolan Koblischke and Jo Bovy describe GalaxyWiki, a resource designed to make that distinction easier. Their pipeline extracted 44,332 galaxy discussions containing 85,314 quotations from 18,641 papers. Grouping discussions by sky position produced 13,616 entries interpreted as individual galaxies, although the authors caution that the grouping does not guarantee distinct objects.

Looking for discussion, not just names

The researchers started with 156,006 astronomy papers dated June 1993 through May 2024. Their source was an astronomy subset of MINT-1T, a dataset containing text parsed from papers’ LaTeX source files. A preliminary filter kept papers containing the standalone word “galaxy” or “galaxies,” leaving 87,899 papers for model processing.

GPT-5.5 language-model agents, running in the Codex software environment, searched those papers for galaxies discussed in more than two sentences. Here, an agent means a model equipped with tools to inspect text and copy passages, rather than simply answer from its training.

For each galaxy, the agents were asked for one to three quotations totaling three to eight sentences, a short summary supported by those quotations, and topic labels from a fixed astronomy vocabulary. They were instructed to exclude the Milky Way and objects within it, and to extract at most 30 galaxies per paper.

The resulting wiki lets readers search by name, sky position, topic or text description. Each page includes paper-level summaries and quotations, source links, and a two-sentence overview generated using Gemini 3.5 Flash. The researchers programmatically checked that the quotations and papers referenced by those overviews belonged to their dataset. That checks source membership, not whether every summary accurately represents its evidence.

Matching different names to the same object

Collecting passages was only part of the problem. To combine discussions across papers, the team needed to determine which records referred to the same galaxy.

The initial extraction found explicit sky coordinates for just 8.3% of discussion records. The researchers therefore queried NED using recognizable galaxy identifiers. Agents investigated remaining records by checking the original papers, trying additional NED identifiers, and consulting external resources such as published catalogs when necessary.

Together, these methods supplied coordinates for 98.2% of the discussion records. The team then linked positions less than one arcsecond apart, also joining records connected through intermediate neighbors. An arcsecond is one three-thousand-six-hundredth of a degree.

This produced the 13,616 entries, of which 69.2% contained only one discussion record. These numbers describe the extracted literature, not a representative census of galaxies or their properties.

What the checks establish

The authors compared extracted quotations with their source text. About 77.8% matched character for character; allowing differences in Unicode, spacing and punctuation raised that figure to 96.5%. Many additional quotations consisted of separately matching fragments combined into one quote.

That distinction matters: authentic words can still lose context when separated or joined. As the authors emphasize, text matching does not establish correct attribution or preservation of meaning.

They also manually examined 120 distinct discussion records in two samples. Across those checks, 118 correctly identified an individual galaxy and extracted substantive discussion. Two records contained little substantive discussion or described another object. Four records in the second sample included source text with missing or corrupted LaTeX symbols.

The samples support the usefulness of retrieved records, but do not measure how many relevant discussions the pipeline missed. Coordinate checks also revealed errors reaching 12.2 arcseconds—larger than the grouping threshold—which can split discussions of one galaxy across separate entries.

An absence is not a discovery claim

Coverage remains a central limitation. The starting corpus contained 156,006 papers, compared with 299,824 astronomy postings the authors counted over the same period. Papers without an arXiv version are absent too, and coverage stops in May 2024.

Models may overlook discussions even within included papers. The 30-galaxy extraction limit was reached in 39 papers, guaranteeing omissions. Identifier matching also produced false positives: the team invalidated 112 records associated with an overly broad pattern intended for Abell designations.

GalaxyWiki is therefore best understood as a route into the literature, with inspectable passages attached. A search result can help a reader locate previous scientific discussion. No result cannot establish that a galaxy is unstudied—and a generated overview does not replace reading the cited paper.