Search spans all of arXiv (1991–present) — physics, mathematics, computer science,
quantitative biology, quantitative finance, statistics, electrical engineering &
systems science, and economics — plus OpenAlex's Computer Science, Mathematics
& Social Sciences literature, reaching back to the 1800s. First-use datings, coiners and
acronyms are drawn from the arXiv corpus. A term's earliest citation may be an earlier
sense of the same spelling — the corpus remembers that “LLM” meant
local linear model first.
Keys: / search
Under the hood: each of the ~23 million abstracts is distilled into a 768-dimensional vector by a neural embedding model, every dimension squeezed to a single byte, and the whole collection is bound into one HNSW graph — a small-world web of nearest neighbors. Your query walks that graph greedily, touching a few thousand candidates rather than all 23 million; a parallel BM25 keyword index catches the exact phrases that embeddings blur; the two ballots are fused by reciprocal rank, and a cross-encoder re-reads the finalists against your words to settle the final order.