ICML
Last reviewed
Apr 28, 2026
Sources
25 citations
Review status
Source-backed
Revision
v1 · 4,148 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 28, 2026
Sources
25 citations
Review status
Source-backed
Revision
v1 · 4,148 words
Add missing citations, update stale details, or suggest a clearer explanation.
ICML (the International Conference on Machine Learning) is a leading academic conference in [[machine_learning|machine learning]], organized annually by the [[imls|International Machine Learning Society]] (IMLS). Along with [[neurips|NeurIPS]] and [[iclr|ICLR]], it is widely regarded as one of the three top-tier venues for publishing machine learning research. The conference grew out of an informal series of workshops first held at [[cmu|Carnegie Mellon University]] in 1980 and has since become one of the largest academic gatherings in artificial intelligence, attracting tens of thousands of submissions and many thousands of attendees in recent years [1][2].
ICML covers the full breadth of machine learning research, including supervised, unsupervised, and reinforcement learning, statistical learning theory, optimization for ML, probabilistic models, kernel methods, [[deep_learning|deep learning]], representation learning, and the interfaces with natural language processing, computer vision, robotics, and computational biology. Proceedings have been published open access since 2010 through the Proceedings of Machine Learning Research (PMLR) [3].
The lineage of ICML traces back to the first International Workshop on Machine Learning (IWML), held in July 1980 at [[cmu|Carnegie Mellon University]] in Pittsburgh. The 1980 workshop was organized by [[ryszard_michalski|Ryszard S. Michalski]], [[jaime_carbonell|Jaime Carbonell]], and [[tom_mitchell|Tom M. Mitchell]], three figures who would go on to edit the foundational Machine Learning: An Artificial Intelligence Approach book series that helped define machine learning as a coherent subfield distinct from broader artificial intelligence [1][4]. The workshop format continued through the 1980s, with subsequent meetings in 1983 (Monticello, Illinois), 1985 (Skytop, Pennsylvania), and 1987 (Irvine, California). These were intimate gatherings of a few dozen to a few hundred attendees, with proceedings often published as edited volumes.
ICML adopted the conference name and numbering scheme starting with the 5th International Conference on Machine Learning in 1988, held in Ann Arbor, Michigan. The numbering reflects continuity with the four prior workshops [1]. The 1993 edition at the University of Massachusetts Amherst is sometimes cited as the point at which ICML solidified its modern conference format with refereed proceedings, although the conference designation had begun in 1988 [1].
For most of its early history, ICML was run informally, with each year's host institution managing logistics independently. At the ICML 2001 business meeting in Williamstown, Massachusetts, attendees approved the creation of a democratic governance structure, and the [[imls|International Machine Learning Society]] (IMLS) was established as a non-profit organization in 2001, with formal incorporation completed shortly thereafter [5][6]. The IMLS now governs ICML, manages the conference finances, and oversees affiliated publishing efforts. The IMLS is distinct from the Neural Information Processing Systems Foundation, which separately governs [[neurips|NeurIPS]].
The rise of [[deep_learning|deep learning]] after 2012 transformed ICML from a moderately sized academic conference into one of the largest events in computer science. Submissions roughly doubled every few years through the 2010s, and the conference expanded to include extensive workshop and tutorial tracks. By the mid-2020s, ICML had become a major commercial as well as academic event, with sponsorship from industry research labs playing a substantial role [2].
A typical modern ICML runs for five to seven days and has a layered structure that has evolved gradually since the 1990s.
Reviewing for the main track has been double-blind since the early 2010s. Different years have used different review platforms; recent editions have made use of [[openreview|OpenReview]] for portions of the review cycle, while submission and reviewer assignment have at times been managed through CMT (Microsoft's Conference Management Toolkit).
The table below lists ICML editions from the 1980 founding workshop through the most recent meetings. Conference numbering follows the IMLS convention of counting from the 1980 IWML as the first edition [1].
| Year | Edition | Location | Country |
|---|---|---|---|
| 1980 | 1st (IWML) | Pittsburgh ([[cmu | CMU]]) |
| 1983 | 2nd | Monticello, Illinois | United States |
| 1985 | 3rd | Skytop, Pennsylvania | United States |
| 1987 | 4th | Irvine, California | United States |
| 1988 | 5th | Ann Arbor, Michigan | United States |
| 1992 | 9th | Aberdeen | United Kingdom |
| 1993 | 10th | Amherst, Massachusetts | United States |
| 1996 | 13th | Bari | Italy |
| 1999 | 16th | Bled | Slovenia |
| 2001 | 18th | Williamstown, Massachusetts | United States |
| 2002 | 19th | Sydney | Australia |
| 2004 | 21st | Banff, Alberta | Canada |
| 2005 | 22nd | Bonn | Germany |
| 2008 | 25th | Helsinki | Finland |
| 2009 | 26th | Montreal | Canada |
| 2010 | 27th | Haifa | Israel |
| 2011 | 28th | Bellevue, Washington | United States |
| 2012 | 29th | Edinburgh | United Kingdom |
| 2013 | 30th | Atlanta, Georgia | United States |
| 2014 | 31st | Beijing | China |
| 2015 | 32nd | Lille | France |
| 2016 | 33rd | New York City | United States |
| 2017 | 34th | Sydney | Australia |
| 2018 | 35th | Stockholm | Sweden |
| 2019 | 36th | Long Beach, California | United States |
| 2020 | 37th | Vienna (virtual due to COVID-19) | Austria |
| 2021 | 38th | Vienna (virtual) | Austria |
| 2022 | 39th | Baltimore, Maryland | United States |
| 2023 | 40th | Honolulu, Hawaii | United States |
| 2024 | 41st | Vienna | Austria |
| 2025 | 42nd | Vancouver | Canada |
ICML 2014 in Beijing was the first time the conference was held in mainland China and marked an important step in broadening its geographic reach. The 2020 and 2021 editions were forced online by the COVID-19 pandemic. Both years are still nominally associated with Vienna, Austria, where they had been scheduled to take place; the conference returned to Vienna for an in-person meeting in 2024 [8].
From the 1990s through the early 2010s, ICML typically saw a few hundred to roughly a thousand submissions per year, with [[acceptance_rate|acceptance rates]] hovering between 20 and 30 percent. The deep-learning era pushed those numbers sharply upward.
| Year | Submissions | Accepted | Acceptance rate |
|---|---|---|---|
| 2015 | 1,037 | 270 | 26 percent |
| 2018 | 2,473 | 621 | 25 percent |
| 2020 | 4,990 | 1,088 | 21.8 percent |
| 2022 | 5,630 | 1,235 | 21.9 percent |
| 2024 | 9,473 | 2,609 | 27.5 percent |
| 2025 | 12,107 | 3,260 | 26.9 percent |
These figures are sourced from official ICML statistics pages and the IMLS fact sheets where available [9][10][11]. The 2024 acceptance rate, around 27.5 percent, was unusually high relative to recent precedent, in part because that year saw an enlarged program committee and a deliberate increase in the number of accepted papers to absorb the growing submission load [11].
ICML presents two principal classes of awards each year: paper awards for outstanding work submitted to the current conference, and a Test of Time Award recognizing influential papers from a decade earlier. The terminology has shifted over time. Through 2018 ICML primarily used the label "Best Paper Award"; since 2019 the conference has often used "Outstanding Paper Award" for the same recognition, with Best Paper Award occasionally retained for top selections among the outstanding set. The two terms are used interchangeably in much community discussion.
The table below summarizes selected paper award recipients across recent years. It is not exhaustive: most years select multiple papers and additional honorable mentions or runners-up.
| Year | Paper | Authors | Award |
|---|---|---|---|
| 2017 | Understanding Black-box Predictions via Influence Functions | Pang Wei Koh, Percy Liang | [[best_paper_award |
| 2018 | Obfuscated Gradients Give a False Sense of Security | Anish Athalye, Nicholas Carlini, David Wagner | Best Paper [13] |
| 2018 | Delayed Impact of Fair Machine Learning | Lydia Liu, Sarah Dean, Esther Rolf, Max Simchowitz, Moritz Hardt | Best Paper [13] |
| 2019 | Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations | Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Ratsch, Sylvain Gelly, Bernhard Schoelkopf, Olivier Bachem | Best Paper [14] |
| 2019 | Rates of Convergence for Sparse Variational Gaussian Process Regression | David Burt, Carl Rasmussen, Mark van der Wilk | Best Paper [14] |
| 2020 | On Learning Sets of Symmetric Elements | Haggai Maron, Or Litany, Gal Chechik, Ethan Fetaya | Outstanding Paper [15] |
| 2020 | Tuning-free Plug-and-Play Proximal Algorithm for Inverse Imaging Problems | Kaixuan Wei, Angelica Aviles-Rivero, Jingwei Liang, Ying Fu, Carola-Bibiane Schoenlieb, Hua Huang | Outstanding Paper [15] |
| 2021 | Unbiased Gradient Estimation in Unrolled Computation Graphs with Persistent Evolution Strategies | Paul Vicol, Luke Metz, Jascha Sohl-Dickstein | Outstanding Paper [16] |
| 2022 | Stable Conformal Prediction Sets | Eugene Ndiaye | Outstanding Paper [17] |
| 2022 | Causal Conceptions of Fairness and their Consequences | Hamed Nilforoshan, Johann Gaebler, Ravi Shroff, Sharad Goel | Outstanding Paper [17] |
| 2022 | G-Mixup: Graph Data Augmentation for Graph Classification | Xiaotian Han, Zhimeng Jiang, Ninghao Liu, Xia Hu | Outstanding Paper [17] |
| 2023 | Learning-Rate-Free Learning by D-Adaptation | Aaron Defazio, Konstantin Mishchenko | Outstanding Paper [18] |
| 2023 | A Watermark for Large Language Models | John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, Tom Goldstein | Outstanding Paper [18] |
| 2024 | Stealing Part of a Production Language Model | Nicholas Carlini, Daniel Paleka, and others | Best Paper [19] |
| 2024 | Genie: Generative Interactive Environments | Jake Bruce, Michael Dennis, Ashley Edwards, and others | Best Paper [19] |
| 2024 | Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution | Aaron Lou, Chenlin Meng, Stefano Ermon | Best Paper [19] |
| 2024 | Position: Considerations for Differentially Private Learning with Large-Scale Public Pretraining | Florian Tramer, Gautam Kamath, Nicholas Carlini | Best Paper (position track) [19] |
| 2024 | Position: Measure Dataset Diversity, Don't Just Claim It | Dora Zhao, Jerone Andrews, Orestis Papakyriakopoulos, Alice Xiang | Best Paper (position track) [19] |
| 2024 | Debating with More Persuasive LLMs Leads to More Truthful Answers | Akbir Khan, John Hughes, and others | Best Paper [19] |
| 2024 | VideoPoet: A Large Language Model for Zero-Shot Video Generation | Dan Kondratyuk, Lijun Yu, and others | Best Paper [19] |
| 2025 | CollabLLM: From Passive Responders to Active Collaborators | Shirley Wu, Michel Galley, Baolin Peng, and others | Outstanding Paper [20] |
| 2025 | Train for the Worst, Plan for the Best | Jaeyeon Kim, Kulin Shah, Vasilis Kontonis, Sham Kakade, Sitan Chen | Outstanding Paper [20] |
| 2025 | Conformal Prediction as Bayesian Quadrature | Jake Snell, Thomas Griffiths | Outstanding Paper [20] |
| 2025 | The Value of Prediction in Identifying the Worst-Off | Unai Fischer Abaigar, Christoph Kern, Juan Perdomo | Outstanding Paper [20] |
The 2024 awards drew particular attention because three of the recognized works fell under the new position-paper track (the differentially private pretraining and dataset-diversity papers being the most discussed), which gave the inaugural year of that track an immediate seal of legitimacy [7][19].
ICML introduced its Test of Time Award in 2009 with retrospective recognition of papers presented a decade earlier. The award was put on hold for several years and then restarted in 2017 in its current form, recognizing one paper from the conference held ten years prior that has had outstanding lasting impact. Recent recipients include the following.
| Year given | Paper | Year published | Authors |
|---|---|---|---|
| 2017 | Combining Online and Offline Knowledge in UCT | 2007 | Sylvain Gelly, David Silver |
| 2018 | A Unified Architecture for Natural Language Processing: Deep Neural Networks with Multitask Learning | 2008 | Ronan Collobert, Jason Weston |
| 2019 | Online Dictionary Learning for Sparse Coding | 2009 | Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro |
| 2020 | Gaussian Process Optimization in the Bandit Setting: No Regret and Experimental Design | 2010 | Niranjan Srinivas, Andreas Krause, Sham Kakade, Matthias Seeger |
| 2021 | Bayesian Learning via Stochastic Gradient Langevin Dynamics | 2011 | Max Welling, Yee Whye Teh |
| 2022 | Poisoning Attacks Against Support Vector Machines | 2012 | Battista Biggio, Blaine Nelson, Pavel Laskov |
| 2023 | Learning Fair Representations | 2013 | Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, Cynthia Dwork |
| 2024 | DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition | 2014 | Jeffrey Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, Trevor Darrell |
| 2025 | Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift | 2015 | Sergey Ioffe, Christian Szegedy |
The selection of "Combining Online and Offline Knowledge in UCT" by Gelly and Silver in 2017 was widely viewed as fitting given the role that Monte Carlo tree search later played in [[deep_learning|deep reinforcement learning]] systems including AlphaGo. The 2025 recognition of Batch Normalization underlines how thoroughly that technique reshaped deep network training [21].
Many influential works in modern machine learning first appeared at ICML. Selected papers that genuinely originated at ICML and have proven highly influential include the following.
It is worth noting which famous papers are commonly but incorrectly associated with ICML. The original AlexNet paper (Krizhevsky, Sutskever, Hinton, 2012) was at NeurIPS, not ICML. The first Generative Adversarial Networks paper (Goodfellow and colleagues, 2014) was at NeurIPS. The Adam optimizer paper (Kingma and Ba, 2015) was at ICLR. Word2vec (Mikolov and colleagues, 2013) appeared at ICLR and NeurIPS. The neural attention paper for machine translation (Bahdanau, Cho, Bengio, 2015) was at ICLR. The original AlphaGo paper appeared in Nature, not at any of the top ML conferences. The official PMLR proceedings and DBLP listings are authoritative for venue attribution.
ICML is one of three flagship conferences for machine learning research. The table below compares it with [[neurips|NeurIPS]], [[iclr|ICLR]], and other adjacent venues. Numbers are approximate and reflect recent multi-year averages.
| Conference | Founded | Governance | Scope | Typical submissions | Typical [[acceptance_rate|acceptance rate]] | | --- | --- | --- | --- | --- | --- | | [[icml|ICML]] | 1980 (workshop), 1988 (conference) | [[imls|IMLS]] | All of [[machine_learning|machine learning]] | 9,000 to 12,000 | 22 to 28 percent | | [[neurips|NeurIPS]] | 1987 | NeurIPS Foundation | ML, neuroscience, statistics, applied probability | 12,000 to 17,000 | 23 to 27 percent | | [[iclr|ICLR]] | 2013 | Open Review (community) | Representation learning, [[deep_learning|deep learning]] | 7,000 to 12,000 | 27 to 32 percent | | [[aaai|AAAI]] | 1980 | AAAI | Broad AI including ML, KR, planning | 8,000 to 14,000 | 20 to 25 percent | | [[ijcai|IJCAI]] | 1969 | IJCAI | Broad AI | 4,000 to 7,000 | 13 to 17 percent | | [[cvpr|CVPR]] | 1985 (formal) | IEEE / CVF | Computer vision | 9,000 to 13,000 | 22 to 28 percent | | [[eccv|ECCV]] | 1990 | ECVA | Computer vision | 5,000 to 9,000 | 25 to 30 percent | | [[iccv|ICCV]] | 1987 | IEEE / CVF | Computer vision | 7,000 to 12,000 | 23 to 27 percent | | [[acl|ACL]] | 1962 | ACL | Computational linguistics | 4,000 to 7,000 | 20 to 25 percent | | [[emnlp|EMNLP]] | 1996 | ACL SIGDAT | NLP, empirical methods | 4,000 to 7,000 | 20 to 25 percent | | [[naacl|NAACL]] | 2000 | ACL | NLP, North American chapter | 3,000 to 5,000 | 22 to 28 percent |
A common point of confusion is the relationship between ICML and [[neurips|NeurIPS]]. Both are top-tier general machine learning conferences with substantial overlap in scope and authorship, but they are governed by entirely separate non-profit bodies. ICML is run by the [[imls|IMLS]] founded in 2001 to 2002, while NeurIPS is run by the Neural Information Processing Systems Foundation, which traces its origin to 1987. [[iclr|ICLR]], first held in 2013, has historically focused more narrowly on representation learning and pioneered fully open peer review through [[openreview|OpenReview]].
Several trends have shaped ICML in the last several years.
Like other top machine learning conferences, ICML has been the subject of recurring criticism from the community.