Choosing your primary category (cs.AI vs cs.LG vs cs.CL and friends)

Every arXiv submission has one primary categoryPrimary categoryThe single category that best fits a paper's main contribution. It decides which endorsement you need and where the paper is announced., the place where your paper is announced and where its main readers will find it. You may also add a small number of cross-listsCross-listingAlso listing a paper in secondary categories so readers in related fields can find it. Use it only when the paper is genuinely relevant there. (see Cross-listing). Picking the primary well helps readers find you and reduces the chance that moderatorsModerationarXiv's check of every submission for fit and basic quality before it is announced. It is separate from endorsement and from peer review. need to reclassify your paper or hold it.

This article focuses on the categoriesCategory (subject class)arXiv's topic label for a paper, such as cs.AI (Artificial Intelligence) or cs.CL (Computation and Language). Moderators and readers use it to route and find papers. that most first-time authors in AI ask about. ResearchGuild is not affiliated with arXiv; the arXiv Category Taxonomy is the authoritative list, and it is worth reading the description of every category you are considering.

The key idea: who is your main reader?

arXiv's Machine Learning Classification Guide gives a useful principle: if the primary domain of your application is available as another arXiv category, and readers of that category would be the main audience, that application category should be primary.

So before you look at category codes, finish this sentence: "The people who most need to read this paper are researchers in ____." That answer usually points to your primary.

The common AI-adjacent categories

The summaries below paraphrase arXiv's subject class descriptions. Read the originals before deciding.

Category	What arXiv says it covers (paraphrased)
cs.AI (Artificial Intelligence)	All areas of AI except vision, robotics, machine learning, multiagent systems, and computation and language, which have their own subject classes.
cs.LG (Machine Learning)	All aspects of machine learning research (supervised, unsupervised, reinforcement learning, bandits and so on), including robustness, explanation, fairness and methodology. Also an appropriate primary for applications of ML methods.
cs.CL (Computation and Language)	Natural language processing and computational linguistics. Work on artificial languages (programming languages, logics, formal systems) that doesn't address natural-language issues is not appropriate here.
cs.CV (Computer Vision and Pattern Recognition)	Image processing, computer vision, pattern recognition and scene understanding.
cs.IR (Information Retrieval)	Topics such as recommender systems, document classification, topic modelling and computational advertising.
stat.ML (Machine Learning, Statistics)	Machine learning with a statistical focus.

cs.AI is not the "default" for anything AI-related

A frequent mistake is choosing cs.AI because a paper involves "AI". arXiv's description explicitly excludes areas that have their own classes. A language-model paper whose main readers are NLP researchers usually belongs in cs.CL; a new training method usually belongs in cs.LG.

cs.LG vs stat.ML

arXiv's guide says the choice often depends on the authors' home field, and that ML papers with a statistical focus (new statistical methodology or derivations) are more appropriate for stat.ML. It also notes that papers with stat.ML as primary are automatically cross-listed to cs.LG, but not vice versa.

Reinforcement learning

Per the guide, reinforcement learning papers should generally have cs.LG as primary, unless they are more appropriate in math.OC (optimisation and control), econ.GN, eess.SY (systems and control), or stat.ML (statistical grounding).

Applied ML papers

The guide says papers that use computer vision or machine learning methods to tackle a specific domain application (its example is detecting diseases in medical images) should be submitted to the topic dealing with that application, with cs.CV as a secondary. The same logic applies to other applied fields.

A short decision checklist

1. List two or three candidate categories.
2. Read each one's full description on the taxonomy page.
3. Look at recent listings in each category: do papers like yours appear there?
4. Check where the papers you cite most were posted.
5. Pick the one whose readers are your main audience. Add at most one or two cross-lists if others genuinely care.

What if a moderator disagrees?

arXiv's moderation pages say moderators may reclassify submissions so they appear in the most relevant category, and questions about classification are one reason a submission can be put on holdOn holdA submission status meaning arXiv moderators are taking a closer look before announcing the paper.. That is normal and not a judgement on your work. See After you submit.

Note also that endorsementEndorsementA statement, made on arXiv's own form by an established arXiv author, that you are someone who can reasonably contribute to a subject area. It is not peer review and doesn't guarantee acceptance. is tied to endorsement domainsEndorsement domainA group of related arXiv subject areas that share one endorsement. Being endorsed in one domain doesn't automatically cover another., so your category choice affects which endorserEndorserAn arXiv author who is eligible to endorse in a subject area, which depends on their own recent papers there. On ResearchGuild, endorsers are volunteers who opt in to read abstracts. can help you. See What is arXiv endorsement?.

Get a second opinion

If you are torn between two categories, ask a reviewer on ResearchGuild. People who read those listings regularly can often tell you in a sentence where your paper will feel at home. Start a paper to request feedback.