arxiv

name	arxiv
description	Search, triage, and cite arXiv papers. Use when the user asks for arXiv/preprint/academic paper discovery, recent ML/AI/science papers, paper metadata, abstracts, or BibTeX. Prefer arxiv_* tools over ad-hoc curl.

arXiv Research

Use this skill for arXiv paper discovery and abstract-level triage. The companion extension provides deterministic metadata tools; this skill provides the research workflow and judgment.

Use the extension first

Prefer:

arxiv_search
arxiv_get
arxiv_bibtex

Human commands:

/arxiv:search <query>
/arxiv:get <id>
/arxiv:bibtex <id>

Do not hand-roll curl unless the extension is unavailable or the user explicitly needs a raw API check.

When to use

Use when the user asks to:

• find papers or preprints on a topic
• look up a specific arXiv ID
• compare recent work
• collect citations/BibTeX
• identify authors, categories, or abstracts
• save research leads into a chain or wiki

Do not use arXiv alone for:

• citation counts or influence metrics
• peer-review status
• exhaustive literature reviews
• full-paper claims not supported by the abstract/metadata

Search workflow

1. Convert the user request into 1-3 focused queries.
2. Use arxiv_search with small maxResults first, usually 5-10.
3. Use category filters when helpful, e.g. cs.LG, cs.CL, cs.CV, cs.AI, stat.ML.
4. Triage by title, abstract, date, authors, and category.
5. Use arxiv_get for exact IDs worth citing.
6. Use arxiv_bibtex only when the user needs references.
7. Distinguish abstract-supported claims from hypotheses that need full-paper reading.

Query tips

Common categories:

cs.AI     artificial intelligence
cs.CL     computation and language / NLP
cs.CV     computer vision
cs.LG     machine learning
cs.CR     cryptography and security
stat.ML   machine learning statistics
math.OC   optimization and control

Useful tool shapes:

{ "query": "test time scaling language models", "category": "cs.CL", "maxResults": 5, "sortBy": "submittedDate" }

{ "author": "Yann LeCun", "category": "cs.LG", "maxResults": 5 }

{ "ids": "1706.03762", "includeBibtex": true }

Reporting format

For paper discovery, return:

Top matches
1. Title — arXiv:id, year/category
   Why it matches: ...
   Caveat: ...
   Link: https://arxiv.org/abs/id

Notable misses / query limits
- ...

Suggested next query
- ...

For comparisons, use a table:

| Paper | Year | Main idea | Why relevant | Caveat |
| --- | --- | --- | --- | --- |

Provenance and persistence

• Cite arXiv IDs and abs URLs.
• If research affects project decisions, save a chain link with query terms and selected IDs.
• If findings become durable project/domain knowledge, ask whether to add them to the wiki; cite arXiv IDs or saved source notes.

Guardrails

• arXiv preprints may be wrong, superseded, or not peer reviewed.
• Abstracts are not enough for detailed method claims.
• Do not claim citation counts; the arXiv extension does not provide them.
• Do not download PDFs unless the user asks and an appropriate document/OCR workflow exists.
• Keep searches bounded; broaden only after explaining why.