AI Wiki

NeurIPS

AI EventsArtificial IntelligenceMachine Learning
14 min read
Updated
Suggest editHistoryTalk37 citationsFact-checked
RawGraph

NeurIPS is the short name used by the annual Conference on Neural Information Processing Systems. Founded in 1987, it is an interdisciplinary research meeting covering artificial intelligence, machine learning, and adjacent fields. Its program combines peer-reviewed papers with invited talks, tutorials, workshops, competitions, demonstrations, and an industry exposition.[1][2]

The Neural Information Processing Systems Foundation, a nonprofit corporation, hosts the conference as its principal activity.[6] The most recent completed edition before the July 28, 2026 research cutoff was NeurIPS 2025. Its organizers reported 21,575 valid main-track submissions, 5,290 main-track acceptances, and an acceptance rate of 24.52 percent.[17] The official fact sheet separately reported approximately 26,382 registrations across the edition's physical and virtual participation modes.[18] These submission and registration figures measure different populations and should not be combined.

Scope and organization

NeurIPS began at the intersection of neural networks, engineering, biological modeling, and information theory. Its subject range later expanded. The 2024 main-track call, for example, solicited work in algorithms, learning theory, infrastructure, evaluation, computer vision, natural language processing, neuroscience, reinforcement learning, and social or economic aspects of machine learning.[8] This breadth encompasses areas such as deep learning, foundation models, language models, large language models, online learning, diffusion models, scaling laws, and the design of a benchmark. Listing a topic in a call for papers does not imply endorsement of a particular method or result.

The Foundation has a board, while a separately listed organizing committee and program committee run each annual edition.[6][7] Roles include general chairs, program chairs, track chairs, workshop and tutorial chairs, ethics review chairs, and operational teams. The people and role names vary by year, so an edition's own organizer page is the authoritative record for that edition.

The conference program is broader than the archival paper tracks. The official description includes invited talks, demonstrations, symposia, oral and poster presentations, tutorials, workshops, and a professional exposition.[1] Competitions and affinity events have also appeared in recent editions. These activities have different selection procedures, and participation in a workshop, competition, tutorial, or expo should not be described as acceptance to the main paper track.

History

Founding and early proceedings

Yaser Abu-Mostafa, the founding program chair, identifies Ed Posner as the founding general chair of the first meeting. The inaugural conference was held in Denver, Colorado, in 1987 with sponsorship from the IEEE Information Theory Group. Abu-Mostafa also records that Dana Anderson edited the first proceedings and that Carver Mead and Terrence Sejnowski were the first invited speakers.[2]

The first proceedings were titled Neural Information Processing Systems. The bibliographic record lists Dana Z. Anderson as editor and the meeting as taking place in Denver in November 1987; the volume was published by the American Institute of Physics in 1988.[3] The conference's own proceedings archive catalogs that volume as NIPS 1987 and lists 90 papers.[4] Later volumes used Advances in Neural Information Processing Systems, while the official online archive indexes conference papers by edition.[15]

The early program deliberately connected engineering and biological approaches to neural information processing.[2] Later calls and proceedings also include statistical learning, optimization, probabilistic modeling, and application areas. The conference's current scope is therefore wider than the biological or hardware connotations that the original name might suggest.[1][8][15]

From NIPS to NeurIPS

For most of its history, the meeting used the abbreviation NIPS. In November 2018, the conference board announced that it would instead use NeurIPS as the abbreviated name. Reporting at the time described the change as a response to criticism of the older acronym and debate within the research community. The full name, Conference on Neural Information Processing Systems, did not change.[5]

The distinction matters when reading older sources. Papers and proceedings from 1987 through 2017 are ordinarily labeled NIPS in the official archive, while later conference material generally uses NeurIPS.[15] Historical titles should be preserved as published rather than retroactively rewritten.

Paper tracks and program

Main track

The main track is an archival paper track. The 2024 call described a double-blind process in which author identities were hidden during review. Submissions and reviews were handled through OpenReview. For accepted papers, reviews, meta-reviews, and discussion were made public, while reviewer identities remained anonymous.[8] Exact rules can change between editions, so the applicable call, handbook, and reviewer guidelines should be checked for a specific year.

Main-track papers are assigned to reviewers and area chairs, with senior area chairs and program chairs involved in calibration and final decisions. The 2025 program chairs emphasized that ratings are inputs rather than an automatic decision rule: chair discussions, rebuttals, evidence raised during review, and integrity checks can affect an outcome.[17] This makes a paper's mean score an incomplete description of the decision process.

Datasets and Benchmarks

NeurIPS launched a dedicated Datasets and Benchmarks track in 2021. That first edition received 484 submissions and accepted 174 papers.[13] Its call used a single-blind process and evaluation criteria tailored to datasets and benchmarks, including documentation, accessibility, and the value of the artifact to a research community.[14] The 2021 papers appeared in a separate proceedings volume; from 2022, the conference archive includes Datasets and Benchmarks papers in the main annual proceedings index.[15]

Track names and criteria have continued to evolve. A result from the Datasets and Benchmarks track should be identified as such rather than counted as a main-track paper. The same rule applies when comparing submission totals or acceptance rates across years.

Position papers

NeurIPS introduced a Position Paper track in 2025. Its call sought well-supported viewpoints about the field rather than conventional reports of original research, and it used criteria designed for that contribution type.[16] The 2025 program chairs described the Main Program, Datasets and Benchmarks, and Position Paper tracks as distinct programs with their own chairs, committees, processes, and criteria.[17] A combined conference-wide acceptance rate would therefore obscure material differences among tracks.

Review, ethics, and reproducibility

Research review and ethics review

Recent main-track calls have used anonymous review, author responses, area-chair recommendations, and program-chair decisions.[8] A separate ethics process can be triggered when reviewers or chairs flag concerns. For 2024, ethics reviewers advised the main program committee in a second round; they did not independently decide acceptance or rejection.[9] Ethical assessment may consider research conduct, potential harms, consent, privacy, data rights, and whether risk mitigations are proportionate to the work.

The conference's paper checklist asks authors to address reproducibility, transparency, ethics, and societal impact. An omitted checklist can be grounds for desk rejection under the published guidance.[10] NeurIPS introduced a reproducibility program and checklist for the 2019 edition.[11] The 2020 call then required a broader-impact discussion and disclosure of funding sources, while also formalizing an ethics-review pathway for flagged submissions.[12] These policies are procedural safeguards, not certifications that an accepted paper is reproducible, safe, or free of error.

Decision consistency

Peer review contains judgment and is not perfectly reproducible. In both 2014 and 2021, NeurIPS ran an experiment in which a subset of submissions was evaluated by two independently assigned committees.[23] A later analysis of the 2021 experiment reported disagreement between the two committees' accept-or-reject recommendations for 23 percent of the duplicated submissions. Under the study's model, approximately half of the accepted-paper list could change if the review process were rerun.[24]

Those figures concern the experimental subset and the specified review procedure. They do not show that half of accepted papers were erroneous, nor do they measure the truth of the papers' claims. The experiment instead quantifies sensitivity of borderline conference decisions to the assigned review committee.

Scale and statistics

Submission counts grew substantially during the first half of the 2020s, but totals must be compared within the same track. The table below uses the official main-track figure reported for each selected edition. It excludes Datasets and Benchmarks, position papers, workshops, and other programs.

EditionReported main-track submissionsAccepted main-track papersReported acceptance rateSource note
20209,4671,89820.0%Official 2020 fact sheet.[22]
20219,1222,33425.6%Official 2021 fact sheet; the new Datasets and Benchmarks track was reported separately.[21]
202312,3433,21826.1%Official 2023 fact sheet.[20]
202415,6714,03725.8%Official 2024 fact sheet.[19]
202521,5755,29024.52%Main-track program-chair report.[17]

Registration totals require similar care. Virtual and hybrid editions count participation differently from fully in-person editions, and official fact sheets sometimes revise historical totals. For 2025, the fact sheet gave a top-line estimate of approximately 26,382 total registrants and listed San Diego, Mexico City, and virtual figures whose sum does not reconcile to that top line.[18] The top-line number can be reported as the organizer's approximation, but its components should not be added into a supposedly more exact total.

The 2025 main track used 20,518 reviewers, 1,663 area chairs, and 199 senior area chairs according to the program chairs.[17] Those role counts describe that track. The broader 2025 fact sheet reports a different, conference-wide pool that includes other tracks, so the two sets should not be treated as contradictory or merged without a common definition.[18]

How to read edition statistics

An edition can publish several defensible totals for different stages of the process. Initial abstracts, completed papers, papers that pass administrative checks, valid submissions sent to review, withdrawn papers, and accepted papers are not the same population. A reported acceptance rate is interpretable only when its numerator and denominator refer to the same track and stage. The 2025 program-chair report explicitly calls its denominator valid main-track paper submissions.[17] Earlier fact sheets use shorter labels, so their values are reproduced here under the labels supplied by the organizers rather than harmonized into a new series.[19][20][21][22]

Presentation designations create another common error. An oral is a format assigned within a program, not an additional acceptance outside the accepted-paper total. The 2025 fact sheet listed 87 orals and 5,290 posters in its program summary.[18] The number of invited talks, tutorials, workshops, competition sessions, and position-paper panels likewise describes program events, not main-track acceptances.

Acceptance and attendance also answer different questions. Paper totals count submissions or accepted records, while registrations count participation credentials. A paper can have multiple authors and a registrant can participate without authoring a paper. Neither statistic establishes the number of distinct researchers who contributed to the proceedings, and neither alone is a measure of research quality or influence.

Proceedings and representative papers

The official proceedings archive provides volumes from 1987 onward and is the primary record for whether a paper appeared in the conference proceedings.[15] Appearance in NeurIPS does not imply that every idea in a paper originated at the meeting, especially when a preprint or related version circulated earlier. The examples below are therefore described only as papers that appeared in the proceedings and by the contribution stated in their abstracts.

YearPaperContribution described by the proceedings record
2001Latent Dirichlet AllocationPresented Latent Dirichlet allocation, a hierarchical Bayesian model for collections of discrete data and a method used in topic modeling.[25]
2012ImageNet Classification with Deep Convolutional Neural NetworksDescribed AlexNet, a large convolutional neural network trained for ImageNet classification using graphics processors and regularization techniques.[26]
2014Sequence to Sequence Learning with Neural NetworksUsed a multilayer LSTM to encode an input sequence and decode an output sequence for a sequence-to-sequence task.[27]
2014Generative Adversarial NetsFormulated the generative adversarial network framework as a game between a generative model and a discriminative model.[28]
2017Attention Is All You NeedProposed the Transformer, an encoder-decoder architecture based on attention rather than recurrence or convolution.[29]
2020Language Models are Few-Shot LearnersReported GPT-3 and evaluated task performance from text interaction without gradient updates or task-specific fine-tuning.[30]
2022FlashAttention: Fast and Memory-Efficient Exact Attention with IO-AwarenessPresented FlashAttention, an exact attention algorithm designed to reduce memory traffic between levels of GPU memory.[31]
2023Direct Preference Optimization: Your Language Model is Secretly a Reward ModelDerived Direct Preference Optimization, an objective for preference fine-tuning that avoids first fitting an explicit reward model and then running the separate reinforcement-learning stage used in common reinforcement learning from human feedback pipelines.[32]

NeurIPS also gives edition-specific awards. Award categories and names can change. As one documented example, the 2022 organizers selected the 2012 AlexNet paper for that edition's Test of Time Award.[33] An award recognizes the decision of the named committee and year; it is not an independent replication of every result in the paper.

Locations and editions

The conference was held in Denver for its first edition and later met in several countries, including the United States, Canada, and Spain, with virtual or hybrid formats during the COVID-19 period.[2][15][18] In 2025, NeurIPS operated in San Diego and Mexico City. The organizers described Mexico City as the conference's first second official physical location and as a pilot intended to address travel demand and visa barriers.[34] The completed-edition fact sheet should be used for attendance reporting rather than the pilot's advance estimates.[18]

As of the research cutoff, the fortieth edition was scheduled for December 2026 with its main meeting in Sydney and official satellite events in Atlanta and Paris.[35] The 2026 call continued to describe the main conference as interdisciplinary and set out that edition's submission scope and rules.[36] Because those events had not yet occurred at the cutoff, this article does not report 2026 acceptance, attendance, award, or outcome figures.

The structured identity associated with this page is Wikidata item Q1961016, which describes the Conference on Neural Information Processing Systems.[37]

See also: International Conference on Machine Learning and International Conference on Learning Representations.

References

  1. ^NeurIPS. "About the Conference." neurips.cc/About
  2. ^Abu-Mostafa, Y. "The first NIPS (now NeurIPS)." California Institute of Technology. work.caltech.edu/nips
  3. ^Anderson, D. Z., editor. *Neural Information Processing Systems*. American Institute of Physics, 1988. Springer bibliographic record. link.springer.com/...9780883185698
  4. ^NeurIPS. "Neural Information Processing Systems 0: NIPS 1987." proceedings.neurips.cc/...1987
  5. ^Else, H. "AI conference widely known as 'NIPS' changes its controversial acronym." *Nature*, November 19, 2018. nature.com/...d41586-018-07476-w
  6. ^NeurIPS. "NeurIPS 2025: Mission Statement and About the Conference." neurips.cc/...2025
  7. ^NeurIPS. "Neural Information Processing Systems Foundation Board 2025." neurips.cc/...Board
  8. ^NeurIPS. "NeurIPS 2024 Call for Papers." neurips.cc/...CallForPapers
  9. ^NeurIPS. "2024 Call for Ethics Reviewers." neurips.cc/...CallForEthicsReviewers
  10. ^NeurIPS. "Paper Checklist Guidelines." neurips.cc/...PaperChecklist
  11. ^NeurIPS. "NeurIPS 2019 Call for Papers." neurips.cc/...CallForPapers
  12. ^NeurIPS. "NeurIPS 2020 Call for Papers." neurips.cc/...CallForPapers
  13. ^NeurIPS Datasets and Benchmarks Chairs. "Announcing the NeurIPS 2021 Datasets and Benchmarks Track Papers." November 19, 2021. blog.neurips.cc/...ets-and-benchmarks-track-papers
  14. ^NeurIPS. "2021 Call for Datasets and Benchmarks." neurips.cc/...CallForDatasetsBenchmarks
  15. ^NeurIPS. "Advances in Neural Information Processing Systems: Proceedings." proceedings.neurips.cc
  16. ^NeurIPS. "NeurIPS 2025 Call for Position Papers." neurips.cc/...CallForPositionPapers
  17. ^NeurIPS Program Committee Chairs. "Reflections on the 2025 Review Process from the Program Committee Chairs." September 30, 2025. blog.neurips.cc/...om-the-program-committee-chairs
  18. ^NeurIPS. *NeurIPS 2025 Fact Sheet*. media.neurips.cc/...NeurIPS2025-Fact_Sheet.pdf
  19. ^NeurIPS. *NeurIPS 2024 Fact Sheet*. media.neurips.cc/...NeurIPS2024-Fact_Sheet.pdf
  20. ^NeurIPS. *NeurIPS 2023 Fact Sheet*. media.neurips.cc/...NeurIPS2023-Fact_Sheet.pdf
  21. ^NeurIPS. *NeurIPS 2021 Fact Sheet*. neurips.cc/...NeurIPS_2021-Fact_Sheet.pdf
  22. ^NeurIPS. *NeurIPS 2020 Fact Sheet*. neurips.cc/...NeurIPS_2020-Fact_Sheet.pdf
  23. ^NeurIPS Program Chairs. "The NeurIPS 2021 Consistency Experiment." December 8, 2021. blog.neurips.cc/...ips-2021-consistency-experiment
  24. ^Beygelzimer, A., Dauphin, Y., Liang, P., and Wortman Vaughan, J. "Has the Machine Learning Review Process Become More Arbitrary as the Field Has Grown? The NeurIPS 2021 Consistency Experiment." arXiv:2306.03262, 2023. arxiv.org/...2306.03262
  25. ^Blei, D. M., Ng, A. Y., and Jordan, M. I. "Latent Dirichlet Allocation." *Advances in Neural Information Processing Systems 14*, 2001. proceedings.neurips.cc/...8d543508116cbc-Paper.pdf
  26. ^Krizhevsky, A., Sutskever, I., and Hinton, G. E. "ImageNet Classification with Deep Convolutional Neural Networks." *Advances in Neural Information Processing Systems 25*, 2012. proceedings.neurips.cc/...8436e924a68c45b-Abstract
  27. ^Sutskever, I., Vinyals, O., and Le, Q. V. "Sequence to Sequence Learning with Neural Networks." *Advances in Neural Information Processing Systems 27*, 2014. proceedings.neurips.cc/...97f410bb7eca942-Abstract
  28. ^Goodfellow, I. et al. "Generative Adversarial Nets." *Advances in Neural Information Processing Systems 27*, 2014. proceedings.neurips.cc/...9a61f95710dbe25-Abstract
  29. ^Vaswani, A. et al. "Attention Is All You Need." *Advances in Neural Information Processing Systems 30*, 2017. proceedings.neurips.cc/...bd053c1c4a845aa-Abstract
  30. ^Brown, T. B. et al. "Language Models are Few-Shot Learners." *Advances in Neural Information Processing Systems 33*, 2020. proceedings.neurips.cc/...18bfb8ac142f64a-Abstract
  31. ^Dao, T., Fu, D. Y., Ermon, S., Rudra, A., and Re, C. "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness." *Advances in Neural Information Processing Systems 35*, 2022. proceedings.neurips.cc/...40d5-Abstract-Conference
  32. ^Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. "Direct Preference Optimization: Your Language Model is Secretly a Reward Model." *Advances in Neural Information Processing Systems 36*, 2023. proceedings.neurips.cc/...6ce7-Abstract-Conference
  33. ^NeurIPS Program Chairs. "Announcing the NeurIPS 2022 Awards." November 21, 2022. blog.neurips.cc/...ouncing-the-neurips-2022-awards
  34. ^NeurIPS Communications Chairs. "NeurIPS Announces Second Physical Location in Mexico City." July 16, 2025. blog.neurips.cc/...hysical-location-in-mexico-city
  35. ^NeurIPS Communications Chairs. "NeurIPS Newsletter: April 2026." April 22, 2026. blog.neurips.cc/...neurips-newsletter-april-2026
  36. ^NeurIPS. "NeurIPS 2026 Call for Papers." neurips.cc/...CallForPapers
  37. ^Wikidata. "Q1961016: Conference on Neural Information Processing Systems." wikidata.org/...Q1961016

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

5 revisions · v6 · 2,834 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check completed against 37 primary, official, academic, and independent sources; all 65 citation calls, 37 references, 31 canonical internal links, 32 material claim groups, and 12 visually inspected pages across six official fact sheets were separately reviewed.

Cite this page: AI Wiki. "NeurIPS." aiwiki.ai, updated 29 Jul 2026, fact-checked 29 Jul 2026. CC BY 4.0. https://aiwiki.ai/wiki/neurips

Suggest edit