International Conference on Learning Representations

The International Conference on Learning Representations (ICLR) is an annual academic conference centered on representation learning and related areas of machine learning. Its first meeting was held in Scottsdale, Arizona, in May 2013 under general chairs Yoshua Bengio and Yann LeCun.^[3][4] The conference's scope has expanded with the field and includes deep learning, generative models, reinforcement learning, optimization, theory, computer vision, natural language processing, robotics, neuroscience, biology, and the social and ethical implications of machine learning.^[1][2]

ICLR is especially associated with open peer review. Its original publication model made submitted manuscripts, anonymous designated reviews, public comments, author responses, and revisions visible on a review site.^[5] That review philosophy should be distinguished from the software platform used to implement it: the conference used CMT for at least the 2015 and 2016 conference tracks, returned the conference track to OpenReview in 2017, and introduced double-blind submissions in 2018.^{[9][10][11][13]} The current process combines anonymous public reviews with author responses, revisions, discussion among reviewers and area chairs, and a final decision by the program committee.^[14][15][16]

The fourteenth ICLR took place at Riocentro in Rio de Janeiro, Brazil, from April 23 through April 27, 2026. The main conference occupied the first three days, followed by two workshop days.^[1] Organizers reported 19,525 valid, format-compliant submissions for that edition and 5,355 accepted papers.^[27] Those figures continued a rapid rise from 67 submissions and 23 conference-track acceptances in 2013.^[6]

Scope and program

Representation learning concerns methods that learn features or internal descriptions of data instead of relying entirely on features designed by hand. ICLR's current call for papers treats that subject broadly. It includes supervised, semi-supervised, and unsupervised learning; generative models; reinforcement learning and planning; optimization; theory; causal methods; computer vision; natural language processing; speech and audio; robotics; neuroscience; biology; software and hardware; and work on fairness, safety, privacy, and interpretability.^[2] This breadth means that papers need not propose a new neural-network architecture to fall within the conference's scope.

The main track publishes full research papers. Accepted papers are presented as posters, and a smaller subset may also receive oral or other highlighted presentations. The annual program also includes invited talks, workshops, social events, mentoring, and meetings organized by affinity groups.^[1][26] Workshops are selected separately and can use formats or review criteria that differ from the archival conference track. This distinction was present in the first edition: ICLR 2013 had separate conference and workshop programs, and rejected conference submissions were not automatically archival publications.^[5][7][8]

The event has combined research communication with community functions. The 2025 fact sheet, for example, listed 40 workshops, 23 socials, six invited talks, mentoring sessions, a blog-post track, and several affinity events in addition to the paper program.^[26] These activities change from year to year, so they are better understood as parts of an annual program rather than permanent tracks.

Origins and early development

The archived 2013 conference overview described a practical motivation for establishing ICLR. Representation learning affected applications in vision, speech, audio, and natural language processing, but the organizers believed researchers working on those problems lacked a common venue. Their stated goal was to fill that gap with a meeting devoted to how useful representations can be learned.^[3] This was an organizer rationale, not a formal finding that no relevant work appeared elsewhere.

The inaugural conference ran from May 2 through May 4, 2013, alongside AISTATS in Scottsdale. Bengio and LeCun served as general chairs; Aaron Courville, Rob Fergus, and Chris Manning were program chairs.^[3][4] Sixty-seven papers were submitted to the conference track, 23 were accepted there, and 20 were accepted to the workshop track.^[6][7] The workshop program included Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean's "Efficient Estimation of Word Representations in Vector Space," the early word2vec paper that described the continuous bag-of-words and skip-gram architectures.^[8][41]

The original publication plan separated dissemination from evaluation. Authors first posted manuscripts to arXiv and supplied the conference with links. A separate site hosted anonymous designated reviews and identified them as official reviews. Other readers could comment under their names, authors could answer, and papers could be revised during discussion. The program committee then selected oral and poster presentations. Papers assigned only to the workshop track, as well as rejected papers, were treated as non-archival.^[5]

This model was open, but the platform history was not continuous. The 2015 conference page directed authors to CMT.^[9] The 2016 call explicitly said the conference track would again use CMT instead of OpenReview, while the opening slides identified a newer OpenReview implementation for that year's workshop track.^[6][10] The 2017 conference page linked both tracks to OpenReview, and the organizers' opening presentation credited the platform with supporting extensive paper discussion.^[11][12] ICLR therefore helped develop open review from its first meeting, but it would be inaccurate to say that every conference track used the OpenReview platform from 2013 onward.

The first four editions also show how quickly the program grew. Organizer slides for 2016 list 67 conference submissions in 2013, 87 in 2014, 143 in 2015, and more than 265 in 2016. Conference-track acceptances were 23, 35, 31, and 80, respectively. The same slides list 20, 38, 71, and 108 workshop acceptances and attendance rising from about 105 in 2013 to 449 in 2016.^[6] Because the 2016 submissions figure is only a lower bound and excludes the workshop track, it should not be used to calculate a precise acceptance rate.

ICLR introduced Best Paper Awards in 2016. The two recipients were "Neural Programmer-Interpreters" and "Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding."^[6][28] In 2017 the program gave three Best Paper Awards and 15 Best Review Awards, indicating that recognition of reviewing had become part of the conference's program as well.^[12]

Review and publication process

Open review and double blindness

ICLR 2018 marked a major procedural change. For the first time, conference submissions were double blind: reviewers could not see author names during review, and authors could not see reviewer names. Papers and official anonymous reviews remained on OpenReview, public discussion continued, and authors were still allowed to post their papers on arXiv. Author names were revealed after the review period.^[13] This design combined anonymity during evaluation with a public record of reviews and discussion.

The 2026 author guide describes the modern workflow. Authors uploaded anonymized submissions to OpenReview. Official reviews were released publicly without reviewer names. Authors could respond to reviews and revise papers during the discussion period, while OpenReview generated a PDF comparison showing changes from the original submission. Reviewers discussed the work with authors and one another, then supplied recommendations to an area chair.^[14][15] Area chairs considered the reviews, author responses, discussion, and their own assessment when writing meta-reviews and recommendations; senior area chairs and program chairs oversaw the final decisions.^[16]

The process is open in several different senses, none of which should be conflated:

Scroll sideways for more →

The exact visibility rules can change. In 2020, for example, program chairs disabled new public comments partway through discussion so that authors and reviewers could focus on a defined set of issues.^[18] In 2021, public participation was available during the first discussion stage, followed by a more restricted stage for reviewers and area chairs. Organizers reported that 84 percent of active submissions updated their manuscripts, 93 percent posted rebuttal messages visible to reviewers, and 32 percent of reviews were updated at least once after their initial release.^[21] These figures describe that cycle rather than a permanent guarantee for every year.

Proceedings and access

Accepted papers from 2013 through 2025 are linked from ICLR's proceedings page through OpenReview. The conference lists ISBNs for proceedings beginning in 2017 and states that it does not have an ISSN.^[17] OpenReview also preserves non-accepted and withdrawn submissions according to each year's policy. The 2018 call, for example, stated that a paper withdrawn after the submission deadline would remain publicly visible and that non-accepted conference papers were non-archival even though their reviews, comments, and submitted versions remained on the site.^[13]

Open review makes the evaluation record available for scrutiny, but it does not by itself guarantee consistent decisions or eliminate bias. Academic studies have used ICLR data to examine reviewer scores, rebuttals, author attributes, institutional associations, and later citations. Their results are observational and depend on the years and variables studied, so they should not be converted into claims that a single mechanism caused acceptance or rejection.^[38][39]

Annual meetings and growth

ICLR has met in North America, Europe, Africa, Asia, and South America. The sequence of host locations is documented by the conference's archived pages and later fact sheets.^{[3][6][9][11][24][26]}

Scroll sideways for more →

The planned physical ICLR 2020 would have taken place in Addis Ababa. As concern about COVID-19 increased, the organizers canceled the in-person event and moved the conference online.^[19] The virtual plan combined pre-recorded talks with live question sessions, multiple poster sessions scheduled across time zones, asynchronous paper discussion, workshops, social spaces, and downloadable material for participants with limited bandwidth.^[20] All 687 accepted papers received video and slide presentations.^[18][20]

The 2023 press release described Kigali as the first in-person ICLR since the pandemic and a hybrid conference with live streaming.^[24] ICLR remained hybrid in Vienna in 2024 and Singapore in 2025.^[25][26] The 2026 home page identified Rio de Janeiro as the site of the fourteenth conference, while a separate dates page divided April 23-25 between the main meeting and April 26-27 between workshops.^[1]

Submission and acceptance figures

Submission totals are not always directly comparable. Some organizer documents count all full submissions; others count only valid or reviewed papers after withdrawals and desk rejections. The table below uses the denominator stated by the cited organizer source and avoids filling gaps from secondary statistical compilations.

Scroll sideways for more →

Sources for the early figures are the 2016 and 2017 opening presentations.^[6][12] The 2020 figure comes from the program chairs' review retrospective.^[18] Later figures come from organizer acceptance announcements, press releases, fact sheets, and the detailed 2026 review retrospective.^{[21][23][24][25][26][27]}

Two accounting differences are important. The 2021 acceptance announcement used 2,997 full submissions, while the conference fact sheet reported 3,014 total papers submitted.^[21][22] The two values appear to reflect different processing stages, so the 860 acceptances should be paired with the 2,997 denominator when reproducing the decision announcement. For 2026, the detailed retrospective reported 5,355 acceptances, whereas a later fact sheet listed 5,357. The table uses the retrospective's figure and its published 27.4 percent rate, but the official sources do not resolve the two-paper difference.^[27][47]

The retrospective's 2026 component totals also do not fully reconcile. It reports 19,525 valid submissions, 779 desk rejections, 5,042 withdrawals, and 13,763 papers receiving an accept-or-reject decision, of which 5,355 were accepted and 8,408 rejected. Accepted plus rejected equals 13,763, but valid submissions minus desk rejections and withdrawals equals 13,704, a difference of 59 papers. The retrospective does not explain that mismatch. Its published 27.4 percent rate divides 5,355 acceptances by all 19,525 valid submissions, not by the 13,763 papers reported as receiving a final decision.^[27]

Selected papers and awards

ICLR programs have included work that later became widely used across machine learning. The inaugural workshop presented the early word2vec paper.^[8][41] "Auto-Encoding Variational Bayes," presented at ICLR 2014, developed the variational autoencoder formulation and its reparameterized estimator.^[42] "Adam: A Method for Stochastic Optimization," presented in 2015, proposed an adaptive first-order optimizer.^[43] ICLR's 2025 Test of Time announcement described Adam as one of the most widely adopted optimization algorithms in deep learning.^[30] The 2015 program also included "Neural Machine Translation by Jointly Learning to Align and Translate," which described a learned alignment mechanism for encoder-decoder translation.^[44]

The 2016 program included "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks" and "Continuous Control with Deep Reinforcement Learning."^[31][45] In 2021, "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" presented the Vision Transformer, applying a transformer directly to sequences of image patches.^[46] These examples illustrate the conference's range across language, generative modeling, optimization, reinforcement learning, and computer vision.

Awards provide a more defensible measure of enduring recognition than retrospective claims that any one paper created a field. ICLR began its Test of Time program in 2024, selecting work from approximately a decade earlier.^[29] The award has since been given annually:

Scroll sideways for more →

The 2024, 2025, and 2026 selections were announced by ICLR's award committees.^[29][30][31] Outstanding Paper awards serve a different purpose: they recognize work in the current year's program, while Test of Time awards evaluate influence after years of use and follow-up research.

Workshops and community activities

Workshops have existed since the inaugural meeting and are distinct from the archival conference track. The 2013 publication model allowed the program committee to direct some work to the workshop track and treated that material as non-archival.^[5] In 2016, conference and workshop submissions had separate deadlines and review periods, and the workshop track used OpenReview even though the conference track used CMT.^[6][10] Modern workshops are selected through a proposal process and occupy separate days after the main conference.^[1]

Official guidance for the 2026 workshop program encouraged discussion, presentation and poster slots for contributed work, and tiny or short-paper tracks for less-than-full-conference submissions. It also required proposals to list invited speakers or panelists and to explain how they would identify the archival status of submissions.^[48] Workshop papers should not be described as main-track acceptances unless the program explicitly gives them that status.

ICLR also runs socials, mentoring sessions, blog-post activities, and affinity events. The 2025 fact sheet listed events associated with Women in Machine Learning, LatinX in AI, Queer in AI, and Muslims in ML, along with a Tiny Papers activity integrated into workshops.^[26] These programs support participation and networking, but their names and number vary by edition.

Governance, conduct, and ethics

ICLR publishes separate pages for its standing officers and board and for each year's organizing committee. The board page associated with the 2026 conference is explicitly dated May 2025 through May 2026. It lists Yann LeCun as president, Katja Hofmann as secretary, Yan Liu as treasurer, and Kyunghun Cho, Chelsea Finn, Been Kim, Carl Vondrick, and Yisong Yue as board members.^[34] Because that stated term ended in May 2026, the roster should not be described as the current board after that date without a successor record.

The 2026 organizing committee was a separate, year-specific body. Carl Vondrick was general chair, Bharath Hariharan was senior program chair, and Colin Raffel, Lerrel Pinto, Diyi Yang, and Aleksandra Faust were program chairs. The committee also listed ethics review, workflow, workshop, mentoring, participation, blog-post, volunteer, and social roles.^[35] The published rosters establish this division of work but do not, by themselves, establish election methods, appointment rules, legal powers, or term limits.

The code of conduct covers behavior in conference spaces and provides mechanisms for reporting, enforcement, and appeal.^[32] The code of ethics applies to research and reviewing contributions and addresses honesty, attribution, conflicts of interest, confidentiality, discrimination, potential harms, and professional responsibility.^[33] Annual author and reviewer guides add operational rules, including anonymity, dual-submission restrictions, use of automated tools, review quality, and escalation to area chairs or ethics reviewers.^[14][15]

The 2026 review cycle

The 2026 cycle combined unprecedented scale with two documented review-integrity problems: concerns about the use of large language models in papers and reviews, and an OpenReview security incident during discussion.^[27][36]

Large-language-model use and reference checks

The 2026 author guide allowed general-purpose LLM assistance but required authors to disclose use when an LLM played a significant, contributor-like role in research ideation or writing. The reviewer guide required reviewers to disclose LLM use in their reviews through a field on the review form. Both guides held people responsible for content submitted under their names.^[14][15] The program chairs later described two enforcement workflows. They ran two automated detectors over reviews and notified area chairs when both systems flagged a review as entirely generated by a language model. Area chairs were asked to consider those flags as one signal of review quality, not as an automatic verdict.^[27]

For papers, organizers also searched for references to nonexistent publications or references containing serious bibliographic errors. Their automated system compared extracted references with bibliographic databases and web search, but the retrospective acknowledged a substantial false-positive rate. Flagged cases therefore received human review by an area chair and then by program chairs. The organizers said each paper with flagged references was checked by at least three people, and confirmed hallucinated references led to desk rejection with an appeal channel.^[27] The account does not provide a count that would support estimating how common the problem was.

OpenReview identity exposure

On November 27, 2025, ICLR was notified of an OpenReview API vulnerability that could expose otherwise anonymous authors, reviewers, and area chairs. According to ICLR's incident report, OpenReview fixed the bug about an hour after notification. The conference then learned that submission and identity data for more than 10,000 papers, described as 45 percent of the conference, were circulating online.^[36]

The next day, organizers found a malicious account posting public comments that identified reviewers on 600 papers. ICLR removed the comments, blocked the account through OpenReview, froze public comments and review editing, and notified authors, reviewers, and chairs.^[36] The incident report also said the organizers received reports of attempted collusion, harassment, intimidation, and offers of bribes. Those descriptions are findings reported by ICLR and should not be generalized beyond the documented cases.

To reduce the risk that exposed identities had influenced discussion, ICLR reverted review text and scores to their state at the beginning of the discussion period and reassigned every submission to a new area chair. The new chairs were instructed to evaluate the original reviews, author responses, and discussion and then write meta-reviews. Organizers extended the meta-review period and allowed chairs to consult in small groups on difficult cases.^[36] The later retrospective says the conference also investigated reported misconduct, banned the person associated with exploiting the API, and desk rejected or banned offending members of the conference community.^[27]

Despite the disruption, the selection process concluded with 76,139 reviews from 18,054 reviewers contributing to the final program.^[27] On December 17, after the incident response, ICLR publicly reaffirmed its support for OpenReview and said it had committed earlier in 2025 to double its regular financial contribution to the platform.^[37] The incident therefore changed the 2026 workflow without ending the conference's use of open review.

Open review as research data

Public ICLR records have enabled research on peer review itself. A 2020 preprint by David Tran and colleagues analyzed ICLR submissions from 2017 through 2020 and reported associations among scores, decisions, institutional affiliation, author gender, and later citations.^[38] A separate study by Gang Wang and colleagues assembled 5,527 submissions and 16,853 reviews together with arXiv and citation data to examine the public double-blind process.^[39] These studies illustrate the kinds of questions possible with public records, but their statistical associations do not establish that author identity or any other single factor caused an individual decision.

Rita Gonzalez-Marquez and Dmitry Kobak later assembled abstracts and metadata for about 24,000 ICLR submissions from 2017 through 2024. Their 2024 workshop paper used the collection to study changes in research topics and to compare text representations.^[40] Such datasets are unusually detailed because OpenReview links manuscripts, reviews, rebuttals, and decisions, although researchers still need to account for withdrawals, desk rejections, missing fields, and policy changes across years.

Relationship to other machine-learning conferences

ICLR overlaps with NeurIPS and ICML in authors, reviewers, and subject matter, but its organizing theme is representation learning. Its current call reaches well beyond the narrow sense of learning an embedding: it includes reinforcement learning, generative modeling, optimization, theory, language, vision, robotics, biology, and questions about machine-learning systems and society.^[2] The difference is therefore one of history and emphasis rather than a rule that assigns every paper to only one community.

ICLR's clearest procedural distinction is its public review record. Review visibility, author discussion, and revision are built into the conference workflow, while reviewer identities remain anonymous.^[13][14] Those features have also made ICLR a recurring source of data for empirical work on scientific reviewing.^[38][39][40]

See also

• Representation learning
• Deep learning
• Machine learning
• NeurIPS
• ICML
• Yoshua Bengio
• Yann LeCun

References

1. ^International Conference on Learning Representations. "ICLR 2026." iclr.cc
2. ^International Conference on Learning Representations. "ICLR 2026 Call for Papers." iclr.cc/...CallForPapers
3. ^International Conference on Learning Representations. "International Conference on Learning Representations 2013." iclr.cc/...2013
4. ^International Conference on Learning Representations. "People - ICLR 2013." iclr.cc/...people
5. ^International Conference on Learning Representations. "Publication Model - ICLR 2013." iclr.cc/...publication-model
6. ^Larochelle, Hugo. "Welcome to ICLR 2016!" iclr.cc/...ch.php%3Fmedia%3Diclr2016%3Aopening.pdf
7. ^International Conference on Learning Representations. "Conference Proceedings - ICLR 2013." iclr.cc/...conference-proceedings
8. ^International Conference on Learning Representations. "Workshop Proceedings - ICLR 2013." iclr.cc/...workshop-proceedings
9. ^International Conference on Learning Representations. "ICLR 2015." iclr.cc/...2015
10. ^International Conference on Learning Representations. "Call for Papers (Main Track) - ICLR 2016." iclr.cc/...doku.php%3Fid%3Diclr2016%3Acfp
11. ^International Conference on Learning Representations. "ICLR 2017." iclr.cc/...doku.php%3Fid%3Diclr2017%3Amain
12. ^Ranzato, Marc'Aurelio. "Welcome to ICLR 2017." iclr.cc/...017%3Aranzato_introduction_iclr2017.pdf
13. ^International Conference on Learning Representations. "ICLR 2018 Call for Papers." iclr.cc/...CallForPapers
14. ^International Conference on Learning Representations. "ICLR 2026 Author Guide." iclr.cc/...AuthorGuide
15. ^International Conference on Learning Representations. "ICLR 2026 Reviewer Guide." iclr.cc/...ReviewerGuide
16. ^International Conference on Learning Representations. "ICLR 2026 Area Chair Guide." iclr.cc/...AreaChairGuide
17. ^International Conference on Learning Representations. "ICLR Proceedings." iclr.cc/...Proceedings
18. ^ICLR 2020 Program Chairs. "#OurHatata: The Reviewing Process and Research Shaping ICLR in 2020." December 20, 2019. iclr-conf.medium.com/...-iclr-in-2020-ea9e53eb4c46
19. ^International Conference on Learning Representations. "ICLR2020 as a Fully Virtual Conference." iclr.cc/...virtual
20. ^ICLR 2020 Organizing Committees. "Format for the ICLR2020 Virtual Conference." March 24, 2020. iclr-conf.medium.com/...al-conference-76716ddea640
21. ^Hofmann, Katja, Naila Murray, Alice Oh, and Ivan Titov. "The ICLR 2021 Review Process and Accepted Papers." January 28, 2021. iclr-conf.medium.com/...cepted-papers-7dc65002668e
22. ^International Conference on Learning Representations. "ICLR 2021 Fact Sheet." iclr.cc/...ICLR_2021_Fact_Sheet.pdf
23. ^International Conference on Learning Representations. "ICLR 2022 Press Release." iclr.cc/...ICLR_2022_Press_Release.pdf
24. ^International Conference on Learning Representations. "ICLR 2023 Press Release." iclr.cc/...ICLR_2023_Press_Release.pdf
25. ^International Conference on Learning Representations. "ICLR 2024 Press Release." media.iclr.cc/...ICLR2024_Press_Release.pdf
26. ^International Conference on Learning Representations. "ICLR 2025 Fact Sheet." media.iclr.cc/...ICLR2025_Fact_Sheet.pdf
27. ^ICLR 2026 Program Chairs. "A Retrospective on the ICLR 2026 Review Process." March 31, 2026. blog.iclr.cc/...ve-on-the-iclr-2026-review-process
28. ^International Conference on Learning Representations. "ICLR 2016." iclr.cc/...2016
29. ^International Conference on Learning Representations. "ICLR 2024 Test of Time Award." May 7, 2024. blog.iclr.cc/...iclr-2024-test-of-time-award
30. ^International Conference on Learning Representations. "Announcing the Test of Time Award Winners from ICLR 2015." April 14, 2025. blog.iclr.cc/...-time-award-winners-from-iclr-2015
31. ^International Conference on Learning Representations. "Announcing the Test of Time Awards from ICLR 2016." April 22, 2026. blog.iclr.cc/...test-of-time-awards-from-iclr-2016
32. ^International Conference on Learning Representations. "Code of Conduct." iclr.cc/...CodeOfConduct
33. ^International Conference on Learning Representations. "ICLR Code of Ethics." iclr.cc/...CodeOfEthics
34. ^International Conference on Learning Representations. "May 2025 - May 2026 ICLR Officers and Board Members." iclr.cc/...Board
35. ^International Conference on Learning Representations. "ICLR 2026 Organizing Committee." iclr.cc/...Committees
36. ^ICLR 2026 Program Chairs. "ICLR 2026 Response to Security Incident." December 3, 2025. blog.iclr.cc/...2026-response-to-security-incident
37. ^Yue, Yisong. "ICLR's Commitment to OpenReview." December 17, 2025. blog.iclr.cc/...iclrs-commitment-to-openreview
38. ^Tran, David, et al. "An Open Review of OpenReview: A Critical Analysis of the Machine Learning Conference Review Process." arXiv:2010.05137. arxiv.org/...2010.05137
39. ^Wang, Gang, Qi Peng, Yanfeng Zhang, and Mingyang Zhang. "What Have We Learned from OpenReview?" arXiv:2103.05885. arxiv.org/...2103.05885
40. ^Gonzalez-Marquez, Rita, and Dmitry Kobak. "Learning representations of learning representations." arXiv:2404.08403. arxiv.org/...2404.08403
41. ^Mikolov, Tomas, Kai Chen, Greg Corrado, and Jeffrey Dean. "Efficient Estimation of Word Representations in Vector Space." arXiv:1301.3781. arxiv.org/...1301.3781
42. ^Kingma, Diederik P., and Max Welling. "Auto-Encoding Variational Bayes." arXiv:1312.6114. arxiv.org/...1312.6114
43. ^Kingma, Diederik P., and Jimmy Ba. "Adam: A Method for Stochastic Optimization." arXiv:1412.6980. arxiv.org/...1412.6980
44. ^Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural Machine Translation by Jointly Learning to Align and Translate." arXiv:1409.0473. arxiv.org/...1409.0473
45. ^Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks." arXiv:1511.06434. arxiv.org/...1511.06434
46. ^Dosovitskiy, Alexey, et al. "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." ICLR 2021. openreview.net/forum
47. ^International Conference on Learning Representations. "ICLR 2026 Fact Sheet." media.iclr.cc/...ICLR2026_Fact_Sheet.pdf
48. ^International Conference on Learning Representations. "Guidance for ICLR 2026 Workshop Proposals." iclr.cc/...WorkshopGuide