AI researchers mostly publish in conference proceedings, not journals, and the biggest venues now screen tens of thousands of papers a year. NeurIPS 2025 received 21,575 valid main-track submissions and accepted 5,290, a rate of 24.52 percent, while ICLR 2026 accepted 5,355 of 19,525 valid submissions, or 27.4 percent. Preprints on arXiv come first, and since October 31, 2025 arXiv has required peer review for review articles and position papers in its computer science category.
Key takeaways
- NeurIPS main-track submissions grew from 9,467 in 2020 to 21,575 in 2025, with a 24.52 percent acceptance rate and 20,518 reviewers.
- ICLR 2026 reported 76,139 reviews from 18,054 reviewers and treated undisclosed or fabricated LLM use as a code of ethics violation.
- Original research papers on arXiv are moderated but not peer reviewed; review articles and position papers in computer science now need prior peer review.
- An acceptance at a selective conference is evidence that reviewers found a paper sound enough to present, not that its result has been reproduced.
- Four questions place any claimed "leading researcher" in this system: what was published, where, how it was reviewed, and who has built on it.
How a result travels from lab to record
A typical machine learning result takes three steps. The authors post a preprint, often on arXiv, so others can read it at once. They submit the same paper to a conference, where program committee members assign reviewers. If accepted, the paper appears in the conference proceedings and is presented at a meeting. Journals matter in the field too, but computer science treats a selective conference paper as a primary publication, which is why researchers often list venues such as NeurIPS and ICLR before journal titles.
Scale at NeurIPS in 2025
The program chairs' review-process report of September 30, 2025 gives the numbers for the Conference on Neural Information Processing Systems main track: 21,575 valid submissions, 5,290 accepted, and an acceptance rate of 24.52 percent, which the chairs call on par with earlier years. The review pool included 20,518 reviewers, 1,663 area chairs, and 199 senior area chairs. Submissions in 2020 numbered 9,467, so volume more than doubled in five years while the acceptance rate stayed near one in four.
Volume strains the system. To keep it running, the chairs describe a responsible-reviewing policy under which papers whose co-authors were confirmed as grossly negligent reviewers had to be desk-rejected, and 11 such cases are reported. With 20,518 reviewers for 21,575 submissions, the process relies on authors also serving as reviewers, and the policy exists to enforce that.
NeurIPS 2025 is the latest main-track report available as of October 2026. The 2026 edition has not published comparable review statistics yet. Its organizers have so far announced the workshop program: 477 proposals (454 valid) and 102 accepted, hosted in Sydney, Paris, and Atlanta, per the NeurIPS workshop announcement of August 10, 2026. A main-track review report, if it follows the 2025 pattern, would appear around the conference.
ICLR 2026: public numbers and a security incident
The International Conference on Learning Representations published its retrospective on March 31, 2026. It counted 19,525 valid submissions, 779 desk rejections, 5,042 withdrawals, 13,763 papers with a decision, and 5,355 acceptances. The stated acceptance rate of 27.4 percent equals accepted papers divided by all valid submissions. Divided by papers that received a decision, the same counts give about 38.9 percent, so a quoted rate depends on the denominator. The post reports 76,139 reviews by 18,054 reviewers, which works out to about 5.5 reviews per decided paper by simple division.
ICLR also describes two problems. Authors and reviewers had to disclose LLM use, and papers with confirmed hallucinated references were desk rejected. During the discussion period, a user exploited the OpenReview platform's interface to release the identities of authors, reviewers, and area chairs for many submissions. Organizers reset review scores to their pre-rebuttal state and reassigned area chairs. The episode shows that anonymous review is a procedure that can fail.
What arXiv does and does not check
arXiv is a preprint server, so posting there is not publication in the journal sense. An arXiv blog post of October 31, 2025 says all submissions are moderated to confirm that the work is science of potential interest, but original research papers are not required to be peer reviewed. For the computer science category, review articles and position papers must have been accepted by a journal or conference after peer review, with a journal reference and DOI supplied. Workshop review does not count. arXiv gave two reasons: a surge in submissions linked partly to generative AI, and volunteer moderators who could not vet hundreds of such papers a month.
Where a researcher's claim sits in this system
| Evidence | What it shows | What it does not show |
|---|---|---|
| Preprint on arXiv | The work exists and passed moderation | That experts judged it correct |
| Accepted conference paper | Reviewers judged it sound enough to present | That others reproduced the result |
| Review or position paper on arXiv (computer science) | Peer review took place at a journal or conference | That the review was thorough |
| Prize or named position | Past influence recognized by an organization | Current research quality |
Questions for any "top researcher" claim
- Which papers are cited, and are they listed with venue and year?
- Was the venue selective, and what was its acceptance rate that year?
- Does the claim come from the researcher's institution, a prize body, or an anonymous list?
- Is the ranking based on a stated measure, such as citations or awards?
FAQ
Is a preprint on arXiv peer reviewed?
Not for original research. arXiv moderates for scope and basic standards, but referees do not assess the results. Review articles and position papers in computer science are the exception since October 31, 2025.
Does a 25 percent acceptance rate mean three quarters of papers were bad?
No. An acceptance rate says how many submissions were accepted, not how many were poor. The ICLR 2026 counts give 27.4 or about 38.9 percent depending on whether withdrawn papers are included in the denominator.
Are conference papers less rigorous than journal papers?
The two follow different procedures, and the sources used here describe conference volume and review process only. They do not compare the quality of conference and journal papers.
How can the authors of an AI paper be checked?
Proceedings pages and the arXiv listing show names, affiliations, and dates. Matching a claimed paper to a venue and year is the first step, and a missing match is a reason to ask for a source.
Related reading
- The careers of renowned AI researchers
- How artificial intelligence is developed
- AI and creativity in art
- AI technology in education
More in AI Basics.
Last verified: October 2026
Sources
- Announcing the NeurIPS 2026 Workshops, NeurIPS blog, August 10, 2026
- Reflections on the 2025 Review Process from the Program Committee Chairs, NeurIPS blog, September 30, 2025
- A Retrospective on the ICLR 2026 Review Process, ICLR blog, March 31, 2026
- Attention Authors: Updated Practice for Review Articles and Position Papers in arXiv CS Category, arXiv blog, October 31, 2025