Searching

The positional argument matches every indexed field, a word at a time, with a quoted phrase kept whole.

arxiv search "attention is all you need"

If what you type is already arXiv's grammar, with a field prefix or a bare AND, OR or ANDNOT in it, it goes out as written and the flags still compose around it.

arxiv search 'ti:"attention is all you need" AND cat:cs.CL'
arxiv search 'ti:transformer OR ti:attention' --cat cs.CL

The second one sends (ti:transformer OR ti:attention) AND cat:cs.CL. The parentheses are added because otherwise the trailing AND would bind to the last term alone and quietly answer a different question.

The field flags are arXiv's own prefixes under readable names, and they AND together.

arxiv search --title attention --author vaswani

That sends ti:attention AND au:vaswani. -v prints both the query that was built and the URL it went out on, which is the fastest way to learn the grammar.

Flag	Prefix	Matches
--title	ti:	the title
--author	au:	the author field
--abstract	abs:	the abstract
--comment	co:	the author comment
--journal	jr:	the journal reference
--cat	cat:	a category code
--report	rn:	the report number

Categories

--cat takes a code, repeats, and the codes are OR'd.

arxiv search "transformer" --cat cs.CL --cat cs.LG

A bare archive code such as hep-th or cs matches whether or not that archive was split into categories.

A code arXiv does not have is refused before any request goes out. That is on purpose: arXiv answers a wrong code with zero results and no error, which reads as an empty category rather than as a typo, and a silent zero is the worst possible answer to a misspelling.

arxiv categories          # every code
arxiv category cs.CL      # one, with its group and archive

Dates

The date flags take 2026, 2026-01 or 2026-01-01.

arxiv search "diffusion" --cat cs.CV --from 2025-01 --to 2025-06
arxiv search "diffusion" --updated-from 2026

A bound names a period rather than an instant, so --to 2026-01 means the end of January and not the start of it. That is the reading a person means when they say "up to January", and getting it the other way round silently drops a month.

The raw grammar

--raw sends a query through untouched, for the parts the flags do not cover: parentheses, OR, ANDNOT and quoted phrases. A grammar query typed as the positional argument composes with the flags; --raw is the one that promises nothing at all will be added to it, which is why it refuses to be combined with them.

arxiv search --raw 'abs:"large language model" ANDNOT cat:cs.CL'
arxiv search --raw '(ti:transformer OR ti:attention) AND cat:cs.CL'

arxiv grammar
arxiv grammar --kind operator

prints the whole grammar with examples that work.

One rule is worth memorising: a multi word phrase must be quoted, and the quotes have to survive your shell. ti:large language model is parsed as ti:large AND two loose terms, which returns a different set and does not look like an error.

Sorting and paging

arxiv search "attention" --sort submitted --order desc -n 50

Sort is by relevance unless you say otherwise, and by submission date under --all.

arXiv will not page past ten thousand results. --all walks the whole set anyway by cutting the query into date slices that each fit.

arxiv search --cat cs.CL --from 2026-01 --all -o jsonl > cs-cl-2026.jsonl

The walk runs in submission order rather than relevance order, and that is not a preference. Relevance is recomputed on every request, so a walk ordered by it both repeats and skips papers, and you would never know which.

A long walk says what it is about to cost before it starts, and -vv names each slice as it reaches it.

$ arxiv search cat:cs.CL --all -n 25000 -vv
116325 results in 18 slices, 35 count requests to plan and about 1170 to walk
slice 1 of 18, 199108010000 to 200902061159, 1585 results
slice 2 of 18, 200902061200 to 201711110559, 6000 results
slice 3 of 18, 201711110600 to 202001200429, 6000 results
...

Twenty five thousand ids took about fifty minutes on the three second pace, most of it waiting. arXiv answered 429 a few times along the way, and the tool held the whole plane and carried on rather than losing the walk.

To learn only how many results a query has, use arxiv count, which is one request.

arxiv count --raw 'cat:cs.CL AND abs:transformer'

The seven search UI fields

Seven fields exist in arXiv's search UI and have no prefix in the export API.

--acm-class, --msc-class, --doi, --orcid, --license, --author-id and --full-text.

Naming any of them sends the whole query to the search UI on the fifteen second plane instead, and -v says so before it starts.

arxiv search --doi 10.1038/nature14539
arxiv search --orcid 0000-0002-3300-2109
arxiv search --msc-class 11M06

That route cannot take --cat, --raw or the two --updated- date flags, because the search UI has no equivalent for them. Each refusal explains itself rather than quietly dropping the flag.

arxiv grammar --kind field

lists every field with the plane it lands on.