Menu

ICML 2026

Announcing the ICML 2026 Awards

By ICML 2026 Program Chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, Martin Jaggi and ICML 2026 Position Paper Track Chairs Dale Schuurmans and Jerry Zhu.

We are pleased to announce the following award recipients, to be recognized at ICML 2026! More details on the selected papers and the selection process follow. 

ICML 2026 Outstanding Paper Award: 

The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Zanlin Ni, Shenzhi Wang, Yang Yue, Tianyu Yu, Weilin Zhao, Yeguo Hua, Tianyi Chen, Jun Song, Cheng Yu, Bo Zheng, Gao Huang

High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
Fan Chen, Sinho Chewi, Constantinos Daskalakis, Alexander Rakhlin

ICML 2026 Outstanding Position Paper Award: 

Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit
Sarah Ball, Phil Hackemann

ICML 2026 Outstanding Paper Honorable Mention: 

The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Mohammad Taufeeque, Stefan Heimersheim, Adam Gleave, Chris Cundy

Motion Attribution for Video Generation
Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, Laura Leal-Taixé, Olga Russakovsky, Sanja Fidler, Jonathan Lorraine

How much can language models memorize?
John Xavier Morris, Chawin Sitawarin, Narine Kokhlikyan, Chuan Guo, G. Edward Suh, Alexander M Rush, Kamalika Chaudhuri, Saeed Mahloujifar

A Random Matrix Perspective on the Consistency of Diffusion Models
Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan

To Grok Grokking: Provable Grokking in Ridge Regression
Mingyue Xu, Gal Vardi, Itay Safran

ICML 2026 Outstanding Position Paper Honorable Mention: 

Position: AI/ML Deepfake Research is Misaligned with AI Generated Non-Consensual Intimate Imagery (AIG-NCII)
Li Qiwei, Wells Lucas Santo, Sarita Schoenebeck, Eric Gilbert

ICML 2026 Test of Time Award: 

Asynchronous Methods for Deep Reinforcement Learning
Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, Koray Kavukcuoglu

ICML 2026 Outstanding Paper Award

The ICML 2026 program chairs (Alekh Agarwal, Miro Dudik, Martin Jaggi, and Sharon Li) selected 53 initial award candidates, based on a combination of signals from Reviewer scores and Area Chair nominations. Papers were explicitly included from across the eight top-level subject areas, to compensate for differences in standards across areas and ensure that each of the topic areas gets some representation in consideration. 

The initial set of 53 papers was reviewed by program chairs, who selected a short list of 22 papers. The key criteria were: strong-accept quality, perceived nontrivial “longevity potential,” interest beyond a niche subcommunity, and topic diversity.

The short list was then passed to the Outstanding Paper selection committee. This committee had 11 members, including Andreas Krause (chair), Yoav Artzi, Leon Bottou, Michael Bowling, Jordan Lee Boyd-Graber, Marco Cuturi, Aleksandra Faust, Claudio Gentile, Amir Globerson, Manik Varma, and Yu-Xiang Wang. 

Committee members received anonymized versions of the papers. Each paper was reviewed by two committee members (one of them marked as primary and one of them marked as secondary), and each committee member reviewed 3-5 papers (to help with calibration). Committee members were explicitly told that there is no set quota on the number of awards (so, in principle, all 22 shortlisted papers could be selected). A small set of conflicts of interests was marked (two of the shortlisted candidates had conflicts with two committee members each; no member was conflicted with more than one paper); conflicted committee members were asked to recuse themselves from decision making on the paper where they had a conflict.

The committee met and discussed every paper. For each paper, the primary reviewer provided a short summary and assessment, followed by the secondary reviewer. For the papers where at least one reviewer suggested a recognition, the committee discussed it in more detail with the entire committee. This way, the committee identified candidates for outstanding papers and honorable mentions. For the cases where the committee felt they needed more information, they reached out to additional experts offline to make the final decision. In the end, the committee chose to recognize two Outstanding Papers and five Honorable Mentions. 

Outstanding Papers

The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Zanlin Ni, Shenzhi Wang, Yang Yue, Tianyu Yu, Weilin Zhao, Yeguo Hua, Tianyi Chen, Jun Song, Cheng Yu, Bo Zheng, Gao Huang

With the context that diffusion large language models (dLLMs) are progressively cementing their status as viable contenders to the autoregressive paradigm for language generation, this paper offers a refreshingly counter-intuitive view that challenges one of the field’s dominant assumptions. While arbitrary-order (e.g., confidence-based) generation is often presented as the main feature of dLLMs that allows for a faster generation, the authors convincingly show that on general reasoning tasks dLLMs exploit this freedom to bypass exactly the high-uncertainty “forking” tokens that matter most, collapsing solution diversity. This is a non-obvious failure mode that was far from evident before this work. Based on this insight, the authors propose to revisit post-training for dLLMs using a simpler fixed left-to-right generation order for reinforcement learning (RL) rollouts (their JustGRPO recipe), while retaining parallel decoding at inference. That new approach highlights the still under-explored question of which sampling strategy should drive RL rollouts in dLLMs.

High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
Fan Chen, Sinho Chewi, Constantinos Daskalakis, Alexander Rakhlin

This paper settles a long-standing question in the theory of score-based sampling: whether ε-error can be achieved in polylog(1/ε) steps using only score (gradient) evaluations, rather than the poly(1/ε) steps that discretization-based samplers require. The key construction is first-order rejection sampling (FORS), a meta-algorithm that simulates rejection sampling from first-order queries alone, bypassing density evaluations that previous high-accuracy methods relied on. Under minimal assumptions on the data, this yields an O(d·polylog(1/ε)) sampler, an exponential improvement over prior score-based results, with refinements under non-uniform Lipschitz and intrinsic-dimension conditions, and as a byproduct, the first polylog(1/ε) gradient-only sampler for general log-concave distributions. For diffusion models specifically, the result shows that the number of denoising steps (score-function evaluations) needed to reach a target sample accuracy can in principle drop from polynomial to polylogarithmic in 1/ε, suggesting that highly accurate diffusion sampling can be obtained with a smaller number of function evaluations (NFEs). 

Outstanding Paper Honorable Mentions

The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Mohammad Taufeeque, Stefan Heimersheim, Adam Gleave, Chris Cundy

This paper provides a rigorous investigation into the risks of training large language models against white-box linear lie detectors within a realistic code optimization landscape that is prone to reward hacking. A major contribution of this work is its combination of empirical findings in a coding environment with theoretical proofs, successfully demonstrating that policy gradient methods do not generate direct optimization pressure toward activation manipulation. To understand how models bypass these detectors, the authors establish a helpful taxonomy of training outcomes—identifying blatant deception, obfuscated activations, and obfuscated policies. Crucially, the research provides useful engineering foundations for scalable oversight by showing that alignment failures can be mitigated through high KL regularization paired with strong detector penalties. While the framing of model deception and honesty can appear somewhat anthropomorphic, the paper ultimately tackles an important problem and provides highly useful insights for the field of AI alignment.

Motion Attribution for Video Generation
Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, Laura Leal-Taixé, Olga Russakovsky, Sanja Fidler, Jonathan Lorraine

Every machine learning practitioner knows that carefully choosing the training data can significantly improve model performance. To enhance motion quality in video generation, this paper introduces an attribution method that tracks the contribution of individual training examples to the relevant metrics on test cases. A clever numerical trick makes it work at scale. Thanks to this attribution information, superior performance is achieved using as little as one-tenth of the original training set. 

How much can language models memorize?
John Xavier Morris, Chawin Sitawarin, Narine Kokhlikyan, Chuan Guo, G. Edward Suh, Alexander M Rush, Kamalika Chaudhuri, Saeed Mahloujifar

There’s a big question about how much LLMs are regurgitating stored knowledge and how much they’re generalizing. This paper proposes that one could measure Kolmogorov memorization: how much of a distribution is explained by what is learned by a model. But the paper further breaks this down into intended memorization (generalization) and unintended memorization. The theoretical approach of this paper to distinguish unintended from intended memorization is a novel perspective to think about what LLMs are achieving, and this strong theoretical foundation is complemented by an empirical measurement of what happens in practice. Through this lens, the paper makes the bold claim that GPT-style language models can model 3.6 bits of a distribution per model parameter.

A Random Matrix Perspective on the Consistency of Diffusion Models
Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan

Recent empirical studies have shown that diffusion models exhibit a high level of across-run

reproducibility under deterministic sampling. Even when trained on separate data splits, different architectures will map a single noise seed to nearly identical image outputs. This paper demystifies this phenomenon by proving that diffusion consistency is fundamentally rooted in shared linear Gaussian statistics. In particular, by relying on a linear denoiser setup to track how finite-sample dataset fluctuations distort generated outputs, the authors show that this consistency is not due to complex, high-order deep learning dynamics or data memorization, but it has a linear origin: the shared empirical Gaussian statistics (mean and covariance) across data splits are already enough to predict much of the cross-split consistency. The paper is a tour de force that turns an uncanny empirical quirk into an elegant, mathematically principled baseline for generative reproducibility. By beautifully linking the spectral geometry of data to the stability of model outputs, the authors have provided a foundational cornerstone that will shape the future study of generalization in diffusion models.

To Grok Grokking: Provable Grokking in Ridge Regression
Mingyue Xu, Gal Vardi, Itay Safran

Grokking is the phenomenon in training models where they quickly overfit but generalize badly, and then over time get to a solution that generalizes well. This phenomenon is interesting because it suggests when the model learns the “true” underlying structure of complex problems. Recent works (e.g., Lyu et al., ICLR 2024) have provided an analysis of this phenomenon for deep models, but do not provide global convergence results. The current paper studies grokking for a purely linear model, demonstrating that a two phase behavior emerges in ridge regression. This is a useful finding because it shows that grokking can emerge in this simple setting, and this can be used as a “toy” model for grokking, similar to how deep linear networks have proven useful for analyzing non-linear networks.

ICML 2026 Outstanding Position Paper Award

Selection of the ICML 2026 Outstanding Position Papers was headed by the Track co-chairs, Dale Schuurmans and Jerry Zhu. Nominations were solicited from the Area Chairs for the Position Paper Track. The Track co-chairs then each independently read and assessed all nominated papers.  After their independent assessments, the ranking of nominated papers was found to be nearly identical, with two papers unanimously considered to be the most worthy of distinction. In their view, these two papers exemplify the ideals of the Position Track, from framing important but neglected issues, providing clear evidential support for the main claims, carefully considering alternative positions, and developing practical calls to action that should influence work in their respective areas. After discussion and consideration of all the factors outlined in the Call for Position Papers, one Outstanding Position Paper and one Honorable Mention were identified.

Outstanding Position Paper

Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit
Sarah Ball, Phil Hackemann

This paper challenges the comfortable assumption held by some of us that we are a force for good, asserting that even value alignment—the primary tool we build to ensure AI does no harm—can be misused. The paper supports this assertion with compelling, real-world evidence. Notably, the paper is not about any individual country or company; its scope generalizes to all authorities, present and future. Finally, it proposes several directions to mitigate alignment misuse, raising awareness of the dual-use nature of the technologies our community develops.

Outstanding Position Paper Honorable Mention

Position: AI/ML Deepfake Research is Misaligned with AI Generated Non-Consensual Intimate Imagery (AIG-NCII)
Li Qiwei, Wells Lucas Santo, Sarita Schoenebeck, Eric Gilbert

This paper addresses an important yet uncomfortable issue: identifying the misalignment between current deepfake research and the proliferation of AI-generated non-consensual intimate imagery (AIG-NCII). The paper clearly illustrates the gap between the current technical focus in the literature and the pervasive harm of AIG-NCII, demonstrating how the current neglect of victim impact is exacerbating rather than mitigating the negative impacts of AIG-NCII.  A cogent list of research imperatives is proposed to address the significant harm of AIG-NCII.  This paper needs to be read by everyone who conducts research in this space.

ICML 2026 Test of Time Award

The selection process for the ICML 2026 Test of Time award was orchestrated by Kilian Weinberger, program co-chair of ICML 2016. All papers accepted to ICML 2016 were considered. By identifying influential papers based on citations and scholarly reputation, a shortlist of eight candidates was constructed. Leading researchers in the respective sub-fields were consulted, to assess the long-term impact and suitability of each candidate publication. One paper emerged as a clear winner.

Asynchronous Methods for Deep Reinforcement Learning
Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, Koray Kavukcuoglu

This paper pioneered asynchronous reinforcement learning (RL), which has been a major contributing factor to the success of RL in LLM post-training and has reshaped the way RL is performed in practice. The insight that parallel actor-learners stabilize learning has since inspired numerous successors and has built a lasting legacy within the machine learning community.

Socials at ICML 2026

One of the best parts of any conference is meeting and engaging with other people, particularly those interested in the same things as you. Socials serve as a less hectic environment to do just that. There is no waitlist, signup, or separate admission: if you’re registered for ICML, you’re welcome to that day’s socials.

Our Social Chairs, Kevin Leyton-Brown and Chulhee Yun, have curated an excellent selection of 12 Socials, on a range of topics from open-source research to chess to networking at conferences. There will be four socials each day on Monday, Wednesday, and Thursday. To attend Monday Socials, you must be registered for Tutorials, and to attend Wednesday and Thursday Socials, you must be registered for the main conference. All socials will be from 7 PM to 9 PM. Grab an early dinner with your friends and bring them with you to a Social!

Monday

AI Co-scientists in the Research Loop: Share, Compare, Critique (Mon, ASEM BALLROOM 201-203)

AI co-scientists are beginning to reshape how researchers generate ideas, run experiments, write, review, and evaluate scientific work. This social invites attendees to share practices, compare workflows, and discuss emerging norms around disclosure, authorship, reproducibility, and the responsible use of AI agents in research.

The ICML Chess Club: Blitz, Bullet, and Over-the-Board Networking (Mon, ROOM 300)

Take a break from the main conference and connect with fellow ICML attendees over friendly over-the-board chess. Players of all skill levels are welcome for blitz, bullet, casual games, and relaxed networking.

India @ ICML: Numbers, Stories & What’s Next (Mon, ROOM E1 – E4 (3rd Floor))

This social brings together researchers based in India, the Indian diaspora, and anyone interested in India’s growing role in global AI research. The session will combine data-backed insights, community stories, and open discussion on how to strengthen the ecosystem going forward.

AI for Games (Mon, ROOM E5 – E6 (3rd Floor))

Games have long been a proving ground for AI, and today they are becoming a frontier for agents, world models, generative tools, and large-scale interactive systems. This social brings together researchers and practitioners from academia and industry to discuss where AI for games is heading and what open challenges remain.

Wednesday

How to Network at an AI Conference (Wed, ASEM BALLROOM 201-203)

Networking at a large conference can be intimidating, especially for early-career researchers and newcomers. This social offers a structured, low-pressure environment to practice meaningful academic networking, meet new collaborators, and make ICML feel a little less overwhelming.

Robot Learning Social (Wed, ROOM 300)

Robot learning is rapidly moving from academic labs toward real-world deployment, raising new questions about data, scale, evaluation, and the role of academia. This social gathers the robot learning community to discuss frontiers across reinforcement learning, computer vision, world models, foundation models, and industry-scale robotics.

Bridging Research and Open-Source: A Social for the ML Open-Source Community (Wed, ROOM E1 – E4 (3rd Floor))

Open-source software powers modern machine learning, but researchers and OSS maintainers often work in parallel rather than together. This social creates a space for maintainers, contributors, and academic users to meet, exchange challenges, and build stronger bridges between ML research and open-source infrastructure.

Data Foundations of AI Social (Wed, ROOM E5 – E6 (3rd Floor))

Data is at the center of modern AI, from training data science and curation to attribution, provenance, governance, and evaluation. This social brings together attendees working on different aspects of data foundations to identify shared problems, seed collaborations, and connect a growing community.

Thursday

AI Negotiation and Compensation Social (Thu, ASEM BALLROOM 201-203)

AI career opportunities now span startups, frontier labs, big tech, and global research organizations, each with different compensation structures and negotiation norms. This social offers practical, experience-driven discussion on evaluating offers, understanding leverage, and navigating career decisions in a fast-changing AI job market.

Muslims in ML (MusIML) Social @ ICML 2026: Connecting Global AI and the Muslim World (Thu, ROOM 300)

The MusIML Social is an open, respectful, and non-commercial gathering for the Muslim AI/ML community and all ICML attendees interested in deeper engagement with Muslim scholars, students, practitioners, and institutions worldwide. The event focuses on mentorship, academic exchange, early-career development, cross-regional collaboration, and long-term community building.

Collective, Decentralized Training as a Hedge Against AI Power Concentration (Thu, ROOM E1 – E4 (3rd Floor))

As frontier AI training becomes increasingly centralized, decentralized and collective training offer a compelling alternative technical and institutional direction. This social brings together researchers and practitioners working on distributed training, federated learning, systems, and AI governance to discuss what it would take to make decentralized frontier-scale AI viable.

Cooperative AI (Thu, ROOM E5 – E6 (3rd Floor))

As AI agents increasingly interact, negotiate, and make decisions in shared environments, cooperation becomes a central challenge for AI system design. This social gathers researchers interested in multi-agent systems, mechanism design, AI safety, collective decision-making, and the role of AI in helping humans cooperate on high-stakes problems.

ICML 2026 Registration Update: Capacity Limits and What Attendees Need to Know

ICML 2026 will be held at the Seoul COEX Convention & Exhibition Center this July, and registration has reached a scale that requires us to communicate clearly with the community about what to expect.

As of this week, we are very close to capacity for the event, and we are still weeks out from the start. Registration trends are faster than they have been in previous years, and we are taking steps now to manage that responsibly. 

Registration cap in place

After reviewing venue logistics, WiFi infrastructure, and the practical limits of the conference space, the organizing team has set a cap on total in-person registrations, marginally below the legal limit for the venue. This slight breathing room will enable a significantly more pleasant logistical experience for attendees.

“ICML has always been a conference built around the quality of scientific exchange,” said Kamalika Chaudhuri President of the ICML Board. “Setting a registration cap is not a decision we take lightly, but it reflects our commitment to making sure that when you come to Seoul, you can actually do the things you came to do: hear the talks, have the conversations, and engage with the work.”

General registration for the main conference is expected to sell out.

Within the total cap, registrations are being held in reserve for accepted paper authors who have not yet registered, sponsors with outstanding complimentary registrations, and workshop organizers with outstanding complimentary registrations. Authors in that group have already been contacted directly.

The number of tickets available for general public purchase is very limited. If you are planning to attend and have not yet registered, we encourage you not wait.

A few things to note:

  • Early registration ends today, May 24 (AoE). Registration may reach the cap prior to then.
  • In-person attendance is not required for publication. If you are not planning to come to Seoul, at least one virtual registration is required, though this is not subject to the registration cap.
  • If you require a visa, register before requesting your letter of invitation.
  • Cancellations are eligible for a full refund until June 15, 2026. No refunds will be issued after that date.

Workshops are not affected

As workshops are held on separate days from the main conference, they have their own independent cap on registrations. In particular, workshop registrations are not yet near capacity, and thus we are not applying a similar cap at this time. Prompt registration is still encouraged, and we will reevaluate if registration numbers approach capacity. 

A note on WiFi

Wireless connectivity at the venue has real limitations, and we want to be straightforward about this now rather than after you have arrived. We have invested in the strongest internet access the venue has to offer, but we are also pushing the limits of venue capacity. Thus, crowded areas, particularly the plenary hall, should be anticipated to have degraded or unreliable internet access. Keynotes will be streamed into multiple rooms to distribute attendance. The ICML mobile app stores the schedule locally and does not require a connection to function. We encourage all attendees to download it before traveling, and printing or saving your schedule locally is a reasonable precaution.

We will share more detail about WiFi conditions and logistics as the conference approaches. Further updates will be posted to the ICML 2026 website. If you have any questions or run into issues, please contact the team using the official ICML contact page.

Announcing the ICML 2026 Invited Talks

We are pleased to announce the invited speakers for this year’s conference. The lineup brings together six researchers whose work spans machine learning theory, AI safety and ethics, economics and policy, computational biology, natural language processing, and human–computer interaction.

Pascale Fung

Co-founder and Chief Research & Innovation Officer, AMI Labs; Chair Professor, Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology; Director, Centre for AI Research (CAiRE) 

pascalefung

Pascale Fung is renowned for her expertise in conversational AI, ethical AI, affective computing, statistical NLP, and human-machine interaction, particularly through the medium of spoken language. She influences policy through her roles with the United Nations Advisory Body on AI Governance and the Global Futures Council, a think tank for the World Economic Forum. She is a Fellow of AAAI, IEEE, ACL, ISCA, and the recipient of Outstanding Paper Awards from ACL and EACL.

Susan Athey

Economics of Technology Professor, Graduate School of Business, Stanford University 

susanathey

Susan Athey’s current research focuses on the economics of digitization and the intersection of causal inference and artificial intelligence. She has worked on several application areas, including timber auctions, internet search, online advertising, the news media, labor market transitions, health, and digital technology for social impact. She is a Fellow of the Econometric Society, the American Academy of Arts and Sciences, a Distinguished Fellow of the American Economic Association, and has received the John Bates Clark Medal and the R.K. Cho Economics Prize.

Sham M. Kakade

Rampell Family Professor of Computer Science and Statistics, Harvard University; Co-director, Kempner Institute 

shamkakade

Sham Kakade’s research advances both the theory and practice of AI and deep learning. His thesis established core statistical foundations of reinforcement learning, and his subsequent work developed the first provably efficient policy search methods and mathematical foundations for sequential decision making. His contributions to ML and AI include scalable training methods for large language models and foundation models, tensor methods for latent variable estimation, and foundational theory for deep learning (feature learning, memorization, emergence). He has received the ICML Test of Time Award and the INFORMS Revenue Management and Pricing Prize.

Aviv Regev

Executive Vice President and Head of Genentech Research and Early Development, Genentech

avivregev

Aviv Regev oversees all aspects of gRED’s drug discovery and drug development activities, including the integration of AI/ML into drug discovery in a “Lab in the Loop.” Her own research has pioneered experimental methods and computational algorithms to decipher intra- and intercellular circuits in cells and tissues in health and disease. She is a member of the National Academies of Sciences and Medicine, as well as the American Academy of Arts and Sciences and the Royal Society. Her numerous awards include the Keio Medical Science Prize, the L’Oréal-UNESCO For Women in Science International Award, and the Lurie Prize in Biomedical Sciences.

Verena Rieser

Research Scientist, Google DeepMind

verenarieser

Verena Rieser leads research into the responsible development and alignment of frontier AI models. Her work takes an interdisciplinary approach to develop conversational RL agents, advanced evaluation methodologies, and sociotechnical alignment. She is a Leverhulme Trust Senior Research Fellow, an honorary professor at Heriot-Watt University, and a two-time finalist leader of the Amazon Alexa Prize.

Arvind Narayanan

Professor of Computer Science, Princeton University; Director, Center for Information Technology Policy

arvindnarayanan

Arvind Narayanan’s research studies the societal impact of digital technologies, especially AI. Some interests include the science of AI agent evaluation, the impact of AI on institutions and professions, algorithmic amplification on social media, and fairness and ethics in computing. He has co-authored the popular book AI Snake Oil, the essay AI as Normal Technology, and a widely-read newsletter by the same name. He was one of TIME’s inaugural list of 100 most influential people in AI, is a recipient of the Presidential Early Career Award for Scientists and Engineers (PECASE), and a two-time awardee of the Caspar Bowden PET Award. 

We look forward to the perspectives and insights these speakers will bring to ICML 2026. Further details on talk topics and scheduling will be announced in the coming weeks.

ICML 2026 Lay Summaries

We are continuing the lay summary policy for ICML 2026. As we did last year, all authors of accepted papers will be asked to write a lay summary — a general audience summary — about their work. Below is the guidance we shared with authors in 2025. This advice remains fully applicable for 2026, and we encourage all authors to review it carefully as you prepare your submissions.

Written with the help of Julien Besset, Science Communication Advisor at Mila

How to write a lay summary of your abstract

Starting in 2025, all authors of accepted papers at ICML will be asked to write a lay summary, or general audience summary, about their work. This is different from a technical abstract: it is meant to be understandable by an informed layperson — including people outside of your field of expertise. This blog post provides guidelines and examples to help ICML authors prepare their lay summary.

Why making science more accessible matters

As machine learning continues to reshape many aspects of society, researchers have a responsibility to proactively make their knowledge more accessible. Lay summaries are an important step in that direction, and they are already required by several scientific journals.

Lack of public trust and understanding in science can foster misinformation, including about your own work if it is misinterpreted by the public or by the media.

As a researcher, you have the power to bridge this gap by granting access to trustworthy information about how machine learning works.

This encourages a healthy debate, allowing society to better use machine learning tools and make informed decisions about this transformative technology.

How to write a lay summary

Writing a lay summary may require you to get out of your comfort zone to make your research more accessible. The good news is that it is surprisingly easy to learn.

We suggest using this 3-step formula as a guide: (1) Problem, (2) Solution, (3) Impact. Keep it to a maximum of 10 sentences / 200 words for ICML.

  1. What problem prompted you to start your research?
  2. How did you tackle the problem?
  3. Why does your research matter?

Don’t be so general that your lay summary could be any other paper at ICML.

Fictional example in 50 words:

(1) Methane emissions are one of the biggest contributors to climate change, but tracking them is challenging. (2) We built AI-based satellite imaging systems to monitor methane emissions in remote places. (3) This will help to quickly identify new methane sources and provide insightful data to mitigate the effects of climate change.

See two examples from ICML 2024 below.

5 tips to become a better communicator

  • Know your audience. Don’t expect everyone to know as much about your research as you do. You have spent years developing expertise in a very niche field, but most people have not: try to put yourself in their shoes. Write your lay summary so that a science journalist could understand it.
  • Avoid jargon. Jargon sums up complicated ideas in a few words to speed up communication among experts in the same field. But this can lead to miscommunication if your audience does not know what these words mean. Be creative to avoid jargon words: use metaphors whenever you can; if you can’t, explain the word before using it.
  • Keep it concise. Keep only the most insightful elements, and don’t try to say everything: you cannot sum up years of hard work in a few sentences. Instead, focus on the core aspects of your research that highlight its purpose and impact. The idea of the lay summary is to get other people interested in your work and to increase its reach.
  • Tell a story. Think of your lay summary like a movie trailer for your paper. You don’t want to say everything (once again, you can’t), but you have to provide enough material so that people feel they need to know more. Make your writing engaging. Help your audience understand why your work matters.
  • Read it out loud. This is a great way to cut superfluous words. If it is awkward to pronounce, or you feel that you need to stop to catch your breath, the sentence is too long. Remove words that are unnecessary to understanding what you mean, starting with adjectives. You will be surprised how little words it takes to effectively communicate your ideas.

Examples from ICML 2024:

Example 1:

DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition (Test of Time Award)
by Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, Trevor Darrell

Original abstract:

We evaluate whether features extracted from the activation of a deep convolutional network trained in a fully supervised fashion on a large, fixed set of object recognition tasks can be re-purposed to novel generic tasks. Our generic tasks may differ significantly from the originally trained tasks and there may be insufficient labeled or unlabeled data to conventionally train or adapt a deep architecture to the new tasks. We investigate and visualize the semantic clustering of deep convolutional features with respect to a variety of such tasks, including scene recognition, domain adaptation, and fine-grained recognition challenges. We compare the efficacy of relying on various network levels to define a fixed feature, and report novel results that significantly outperform the state-of-the-art on several important vision challenges. We are releasing DeCAF, an open-source implementation of these deep convolutional activation features, along with all associated network parameters to enable vision researchers to be able to conduct experimentation with deep representations across a range of visual concept learning paradigms.

Lay summary in ~150 words:

We teach computers to analyze images by training them on vast collections of labeled object pictures. We wondered if this learned ability could then be applied to different kinds of image tasks — like being able to differentiate a room from a forest — even with very few new examples.

We took the learned components from a powerful image-analyzing program and used it for tasks like understanding scenes or telling apart different types of birds. We looked at how the program’s internal understanding of images — technically called pre-trained visual features — grouped together for these new tasks. Surprisingly, we found that these representations worked really well, even better than existing methods in some cases.

To help other researchers explore this idea, we have released a free and easy-to-use tool called DeCAF, along with the program’s settings. This lets anyone take these pre-trained visual features and try them out on their own image-related problems.

Example 2:

Information Complexity of Stochastic Convex Optimization: Applications to Generalization, Memorization, and Tracing (winner of one of the 10 best paper awards)

by Idan Attias, Gintare Karolina Dziugaite, Mahdi Haghifam, Roi Livni, Daniel M. Roy

Original abstract:

In this work, we investigate the interplay between memorization and learning in the context of stochastic convex optimization (SCO). We define memorization via the information a learning algorithm reveals about its training data points. We then quantify this information using the framework of conditional mutual information (CMI) proposed by Steinke and Zakynthinou (2020). Our main result is a precise characterization of the tradeoff between the accuracy of a learning algorithm and its CMI, answering an open question posed by Livni (2023). We show that, in the L2 Lipschitz–bounded setting and under strong convexity, every learner with an excess error ϵ has CMI bounded below by Ω(1/ϵ2) and Ω(1/ϵ), respectively. We further demonstrate the essential role of memorization in learning problems in SCO by designing an adversary capable of accurately identifying a significant fraction of the training samples in specific SCO problems. Finally, we enumerate several implications of our results, such as a limitation of generalization bounds based on CMI and the incompressibility of samples in SCO problems.

Lay summary in ~170 words:

How much information do good learning algorithms have to reveal about the data they have been trained on? We wanted to answer this question by using a mathematical tool called conditional mutual information (CMI) to measure the “memorization” of the training data.

Our paper presents the surprising result that for some learning problems, some memorization is necessary to get good performance. This is surprising, as machine learning scientists traditionally thought that learning methods should avoid memorizing the training data to genuinely “understand” how to solve the task — akin to a student who memorizes homework solutions instead of grasping the underlying concepts. Our second result is that, contrary to what was previously believed, CMI cannot always be used to explain why a learning algorithm succeeds in certain scenarios.

Our findings have implications for how we assess the ability of learning algorithms to make predictions on previously unseen data and suggest that there are limits to how much we can compress the data used for training.

Additional resources

Announcing the ICML 2026 Workshops and Affinity Workshops

By ICML 2026 Workshop Chairs Gergely Neu and Courtney Paquette. 

We are thrilled to announce the accepted workshops at ICML 2026. This year’s cohort reflects the remarkable breadth and energy of the machine learning community, spanning foundational theory, applications, and the pressing societal questions our field continues to grapple with. Whether you are a seasoned researcher, a newcomer to the conference, or somewhere in between, we hope you will find workshops that challenge and inspire you. ICML 2026 will feature 44 workshops (which will run during the post-conference workshop days, Friday July 10 and Saturday July 11),  as well as 4 affinity workshops (which will run concurrently to the main conference, from Tuesday July 7 to Thursday July 9). 

In this post, we’ll first give some insight into the workshop selection process, and then share the full list of accepted workshops and affinity workshops. 

Selection Process

The process was especially competitive this year, due to an unprecedented volume of submissions. Up from about 150 proposals for ICML 2025, we received 247 proposals (not counting a handful of desk rejected and withdrawn submissions). Nearly all submissions were of very high quality, which made it very challenging to select the 44 workshops that we will be able to host. Our selection procedure was driven by the following principles.

Our objective throughout the selection process was to make sure that we end up with a program that covers a diverse range of topics that excites the ICML community. We had outstanding material to work with, which forced us to make some hard decisions to make sure that our goals are met within the strict space constraints imposed by the venue. To deal with this challenge, we designed a two-phase process. In the first phase, we arrived at a shortlist of outstanding proposals. In the second phase, this was refined down to the final list. For consistency, the selection process was conducted entirely by the two workshop chairs, with no external reviewers. As a main principle, each proposal was evaluated according to the criteria laid out in the Call for Workshops, but the massive volume of excellent submissions required us to additionally make the final decisions based on soft criteria and subjective comparisons of proposals.

In the first round, each of the submissions was read by one of the two workshop chairs to verify they meet the standards we expect at ICML. Within this round, we had to eliminate several workshops that had any sort of shortcoming: e.g., not having a (nearly) fully confirmed lineup of speakers or a feasible schedule, not having a well designed organization committee with clearly specified roles, not having a clear explanation of the relevance of the proposed speakers to the topic of the workshop, or not having sufficient detail on the background of the organizers to help us verify that they can indeed run their proposed event successfully. We also ruled out workshops with topics that are not relevant for the ICML audience (or whose relevance was not argued clearly enough in the proposal). After this initial round, we ended up with about 70 proposals. We are confident that all these 70 proposals would have resulted in great ICML workshops.

In an ideal world, the process could have stopped there, but unfortunately we still had to make significant cuts to arrive at the final list. This is where we took a closer look at the topics of the workshops, and clustered together proposals on similar topics to get a better side-by-side comparison within each area. We converged on the final decisions after reading all of the surviving proposals one more time, followed by several video calls where we discussed the merits of each potential workshop. Inevitably, we had to make some subjective decisions at this stage, reflecting our personal taste as scientists. We also had to consider some more complex soft criteria, such as whether to admit second (or third, etc.) installments of previous ICML workshops run by one group of organizers or to give a new group a chance to run a workshop on the same topic instead (answers varied on a case-by-case basis).

During the review process, we had the privilege of getting a first-hand account of the topics that get the AI/ML community excited in 2026. It is no surprise that a significant part of the submissions were concerned with aspects of generative AI and large language models. Perhaps more surprisingly, some specific topics within this area were especially popular, most notably with some variation of “agentic AI” appearing in the title of no less than 60 submissions! Many of these proposals had significant overlap not only in their titles, self-described scopes and objectives, but also their invited speakers. Despite the massive interest, we could only accept a handful of these workshops to leave space for other topics as well. And indeed, besides a handful of exceedingly popular areas, the distribution of the topics observed in the submissions was extremely heavy-tailed: these ranged from philosophy, ethics, and governance through statistical hypothesis testing and game theory, all the way to wireless networking and finance.

The diversity of the originally received proposals is clearly reflected in the final program, which stands as a testament to the lively creativity of our field today. We hope that each member of the ICML audience will find inspiration, insight, or possibly other unexpected values within the program.

The selection of affinity workshops was independent of the workshop selection process, and done in consultation with the Inclusion & Accessibility chairs. Seven affinity workshop proposals were submitted, and four were selected for inclusion in the program. The proposals not selected were incomplete in some way, e.g., having unconfirmed speakers (or no speakers at all), or not having enough content to fill an entire workshop. 

Workshops

Links to individual workshops are provided where available, several workshop submission servers can be found on OpenReview. Details may be updated, please email Gautam Kamath if there is information that is incorrect or missing. 

Affinity Workshops

Announcing the ICML 2026 Tutorials

By ICML 2026 Tutorial Chairs Claire Vernade and Adam White.

Tutorials are a critical part of the conference experience. They provide graduate students the opportunity to learn in depth from topic experts. They provide practitioners and theoreticians the ability to learn foundational background and the dark-arts of how to get things working in practice. With our growing community, tutorials can often provide a much needed smaller community to connect with.

With this rapid growth comes challenges. A huge diversity of topics. A large number of proposed tutorial submissions. At the same time we want to channel the community’s inputs, while relying on non-artificial evaluation of submitted proposals. In this post, we outline the review process we used for this year’s tutorials and formally announce the selected tutorials.

This year’s process

This year we used a three-pronged approach. We wanted some invited tutorials, suggestions from the community, and a rigorous review process. In late December we reached out and confirmed three invited tutorial presenters: 

  • Is numerical optimization theory irrelevant to machine learning practice in 2026?

After we confirmed these speakers, we reached out to the ICML community to solicit nominations for both topic suggestions and members of the community to deliver tutorials. We received 118 suggestions from a health diversity of folks from undergraduates all the way to industry professionals:

From these initial seeds, we created a call for proposals. We distilled the community suggestions into six suggested topics (as well as “Other,” of course):

  • Beyond-Transformer sequence models (state-space, S4, etc.)
  • Deep Learning and Deep RL Theory and Applications
  • Diffusion models – quantitative & theoretical understanding
  • LLM post-training and test-time training
  • Safety, machine unlearning, watermarking, and fingerprinting
  • Theorem proving and Lean

We also reach out to 14 individuals, personally inviting them to submit a proposal, as they were nominated by the community in phase two. 

The call resulted in 52 submissions! All proposals were evaluated in the same way, even if they were invited. Both chairs independently reviewed and read every single proposal and made a short list for each of the 7 topic areas (including “Other”). In the final calibration, both chairs agreed almost unanimously on all topic areas. The main criteria used were: 1) quality of proposal, 2) established topic expertise by presenters, and 3) teaching experience. The latter did not exclude industry researchers but required folks to explain their teaching experience. 

In the end, we selected an additional seven tutorials for ICML 2026 (making a total of ten) and we are thrilled to announce them below. All tutorials will take place on the first day of the conference, Monday July 6. 

Retrospective on PAT x ICML 2026 AI Paper Assistant Program

By Google Researchers Rajesh Jayaram and Vincent Cohen-Addad, and ICML 2026 Program Chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, Martin Jaggi.

Following the success of the STOC Experimental AI Feedback program, we launched a similar experiment using Google Research’s Paper Assistant Tool (PAT) before the ICML 2026 deadline. Our goal was to provide authors with actionable, substantive feedback prior to the review process, allowing them to improve their submissions before the deadline. The PAT feedback was totally separate from the peer review process, and was not visible to reviewers, area chairs, program chairs, or anyone outside of the authors. Now that the program has concluded and the feedback from our author survey is in, we want to share some of the results.

The program ran from January 14th to January 26th, providing feedback for approximately 4,500 papers. Papers sent to the system received feedback within ~30 minutes on average, with some edge cases (primarily due to large PDFs causing failures in the AI models). While we initially set conservative eligibility criteria to ensure a smooth rollout, we were able  to quickly relax these requirements. Ultimately we offered a voucher to any author with an OpenReview account older than one month. We believe this afforded a maximal and equitable offering to the community, without allowing for misuse of the tool. We were quite pleased that we could make the program as accessible as possible. 

Highlights

After concluding the program, we released a feedback survey to participating authors. We received 869 responses over the week it was open, and the results demonstrate a high level of satisfaction with the tool. In particular, the tool was able to frequently identify issues, and make suggestions which allowed the authors to make significant changes to their papers before the submission deadline. 

  • High Retention & Satisfaction: 92.1% of respondents stated they would use the tool again. Furthermore, 73.3% rated the feedback as “Very” or “Mostly” helpful. Only 1.6% found the tool to not be useful at all. 
  • Delivering Deep Feedback: The most encouraging results concern the depth of the feedback. The tool successfully identified suggestions beyond surface-level edits and typos:
    • 35.4% of authors of papers containing theory reported the tool identified significant theory gaps that took more than an hour to fix.
    • 31% of authors of papers with experimental results said the feedback prompted them to run new experiments.
  • Clarity and Education: 87.3% of authors felt the tool improved the clarity and readability of their papers. Additionally, 84.5% saw clear educational value in the tool.

Author Quotes on PAT:

The PAT feedback was invaluable in improving my paper’s technical rigor before submission. It helped me identify and fix critical issues including:
– Mathematical contradictions in objective functions
– Inconsistencies in formal problem definitions
– Unsupported claims in experimental sections
– Missing theoretical justifications
The structured, segment-by-segment analysis made it possible to address issues systematically under tight deadline pressure.
— Apoorv Varshney, Department of Mathematics and Computing, Dr. B. R. Ambedkar National Institute of Technology Jalandhar
Compared to most LLMs, PAT’s performance is already quite excellent. PAT discovered that one of my theorem proofs was not rigorous and provided some possible solutions.
— Peihan Wu, Shanghai Normal University
The system was very helpful and also rather “objective”. I look forward to seeing the “human” revisions and comparing them.
— Paolo Stefano Giudici, Professor of Statistics, University of Pavia
[PAT] gives a comprehensive review, allowing authors to address issues that could potentially be questioned by reviewers.
— Enver Menadjiev, M.S.+Ph.D. student, Statistical AI Lab (SAILab) at Kim Jaechul Graduate School of AI at KAIST
Outside feedback is so important when you’re about to submit. I [iterated] for two rounds with [competitor] before, which I paid for. PAT had deeper feedback.
— Domenic Rosati, Dalhousie University

How it worked

The pipeline used for this experiment was scaffolding built on top of state-of-the-art Gemini-based models. To handle the complexity of technical papers, PAT segments the document into logical categorical sections, giving separate feedback for each section, with a high level summary at the beginning. 

Looking Ahead

The PAT x ICML 2026 experiment shows that AI feedback can drive improvements in scientific work, enabling authors to fix theoretical gaps and conduct new experiments in advance of an expert-led peer review process. 

With the continued exponential growth in submissions to machine learning conferences,  AI driven feedback has the benefit both of improving the quality of authors’ papers, and therefore the likelihood of acceptance and reproducibility, as well as easing the increasing burden on human reviewers. 

The PAT pipeline is still an experimental program in development. We plan to incorporate learnings from the ICML collaboration as we prepare for the next iteration of the Paper Assistant Tool. 

Thank you to the thousands of authors who participated in the program, provided feedback, and helped shape the future of AI-assisted research. We look forward to our continued collaboration to make our conference even better!

On Violations of LLM Review Policies

By ICML 2026 Program Chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, Martin Jaggi, Scientific Integrity Chair Nihar B. Shah, and Communications Chairs Katherine Gorman and Gautam Kamath.

AI has increasingly become a valuable part of researchers’ workflows. Unfortunately, AI has the potential to hurt the integrity of peer review if improperly used. Conferences must adapt, creating rules and policies to handle the new normal, and taking disciplinary action against those who break the rules and violate the trust that we all place in the review process. 

ICML is actively working to adapt. This year, we desk-rejected 497 papers (~2% of all submissions), corresponding to submissions of the 398 reciprocal reviewers who violated the rules regarding LLM usage that they had previously explicitly agreed to. 

ICML 2026 has two policies regarding LLM use in reviewing:

  • Policy A (Conservative): No LLM use allowed.
  • Policy B (Permissive): LLMs allowed to help understand the paper and related works, and polish reviews.

This two-policy framework was formed based on community preferences and feedback — indeed, the community is divided on the best way to use LLMs in peer review, with issues such as author consent colliding with preferred reviewer workflows. Further details on the policy are available here. 

After a selection process, in which reviewers got to choose which policy they would like to operate under, they were assigned to either Policy A or Policy B. In the end, based on author demands and reviewer signups, the only reviewers who were assigned to Policy A (no LLMs) were those who explicitly selected “Policy A” or “I am okay with either [Policy] A or B.” To be clear, no reviewer who strongly preferred Policy B was assigned to Policy A. 

795 reviews (~1% of all reviews) written by 506 unique reviewers who were assigned Policy A (no LLMs) were detected to have used LLMs in their review. Again, recall that these are reviewers who explicitly agreed to not use LLMs in their reviews. The method used is described below, and generic AI-text detectors were not used. Every flagged instance was manually verified by a human, in order to avoid false positives. 

If the designated Reciprocal Reviewer for a submission produced such a review, their submission was rejected. In total, this resulted in 497 rejections, corresponding to submissions of 398 reciprocal reviewers. All Policy A (no LLMs) reviews that were detected to be LLM generated were removed from the system. If more than half of the reviews submitted by a Policy A reviewer were detected to be LLM generated, then all of their reviews were deleted, and the reviewer themselves was removed from the reviewer pool. 51 Policy A reviewers were detected to have used LLMs in more than half of their reviews, which is about 10% of the total of 506 detected reviewers. 

To be clear, we are not making a judgment call about the quality of flagged reviews or the reviewers’ intentions. This is simply a statement that the reviewer used an LLM at some point when composing the review, which is unfortunately a violation of the policy they agreed to abide by.

We regret the disruption this will cause in the peer review process. We have been in direct  communication with SACs and ACs impacted, and offered support where we can. The reviews that violated policy have been removed, and ACs may need to find new reviewers. Some submissions which already received a full set of reviews have been desk rejected. And some reviewers whose submissions have been desk rejected may become unresponsive. 

Technical Approach 

At a high level, the LLM detection involved watermarking submission PDFs with hidden LLM instructions, which would subtly influence any review produced via an LLM. Note that this is not a difficult measure to circumvent, particularly if it is known publicly (which was the case for almost the entire review period). Indeed, it may only catch some of the most egregious and careless uses of LLMs in reviewing, where the reviewer is inputting the PDF to the LLM and then directly copy-pasting the output from the LLM. Action was only taken for reviews that were explicitly written by reviewers who agreed to not use LLMs (Policy A). Despite all these caveats, 795 reviews (~1% of all reviews) were found to be violations of the policy. 

The method used was based on recent work by Rao, Kumar, Lakkaraju, and Shah. First, we created a dictionary of 170,000 phrases. For each paper, we sampled two phrases randomly from this dictionary. The probability with which a given pair of phrases is picked is thus smaller than one in ten billion. We watermarked the PDF of each paper submitted with instructions, visible only to an LLM, instructing it to include the two selected phrases in the review. (A human reading the PDF would not directly see this watermark.) 

This technique will not always be successful in detecting LLM generated reviews. The reviewer could discover the watermark and remove it or work around it. The review text could be modified. The LLM may simply ignore the hidden instructions. In experiments shortly before the submission deadline, frontier LLMs generally (though not always) followed the injected instructions. Success rates were over 80% for most models, potentially depending on the method chosen by some LLMs to read the PDF. 

Reviewers may be understandably concerned that their reviews could be wrongly flagged. As described in the paper that outlines the approach we used, false positive rates are strongly controlled. Every flagged instance was also manually inspected by a human to ensure that the review wasn’t simply mentioning that the watermark was present. 

Reviewers can fall short of our expectations in many ways, with or without AI involved. This initiative focused only on one particular action (breaking previously agreed-upon rules for LLM usage) and still identified it for ~1% of all reviews. 

We hope that by taking strong action against violations of agreed-upon policy we will remind the community that as our field changes rapidly the thing we must protect most actively is our trust in each other. If we cannot adapt our systems in a setting based in trust, we will find that they soon become outdated and meaningless.

Update (3/26):

    • We have calculated the probability of incorrectly flagging even a single submitted review, where the reviewer was required to follow Policy A. This probability is 0.0001 (the family-wise error rate).

    • We want to express our gratitude to those of you who have reached out, acknowledging the policy violation and accepting responsibility. We recognize that this was not an easy situation, and your willingness to accept the outcome with integrity reflects the professionalism that makes our community strong.

    • The original version of this post stated that 497 papers, co-authored by 506 reciprocal reviewers, were desk rejected. These 497 papers actually correspond to submissions of 398 reciprocal reviewers, the remaining 108 flagged reviewers were not reciprocal reviewers of active submissions. The text above has been edited to reflect this. 

ICML Experimental Program using Google’s Paper Assistant Tool (PAT)

By Google Researchers Rajesh Jayaram and Vincent Cohen-Addad, and ICML 2026 Program Chairs Alekh Agarwal, Miroslav Dudik, Sharon Li, Martin Jaggi. 

ICML is excited to announce that we will be providing an experimental tool, in partnership with Google, called the Paper Assistant Tool (PAT), to support authors in improving their submissions to ICML 2026.

This program offers authors a unique opportunity to receive automated, private, and actionable feedback on their manuscripts before the final deadline. In a pilot at the Annual ACM Symposium on Theory of Computing (see blog post), 94% of participants found the pre-submission feedback generated by an AI assistant to be helpful, and 85% reported that the feedback resulted in improved clarity of their paper. The feedback the authors will receive from PAT through the ICML experiment will not be used in the review process. Reviewers, area chairs, and program committee members will not have access to the PAT feedback.

What is PAT?

PAT is a specialized Gemini-based tool powered by Google’s Gemini models, utilizing a reasoning-focused pipeline. It is similar to those that have achieved high-level performance in mathematical competitions.

While the program at STOC focused heavily on theoretical correctness, the ICML experiment  has been specifically tuned to address the needs of the machine learning community. The model is designed to help authors identify issues that human reviewers might flag, including (but not limited to) experimental and methodological rigor, narrative clarity in English, and technical correctness.

The goal is to help authors improve the quality of their papers, not to replace human peer review. By fixing clarity issues and potential technical gaps before the work is submitted to the conference, we hope to give authors actionable feedback before their paper enters the review process.

Logistics and Eligibility

The program is entirely optional and operates completely outside the official review process. The program will run for a limited window between January 14 and January 22, 2026.  This deadline may be extended subject to the determination of the program organizers. Eligible submissions uploaded within the window will typically receive the feedback within 24 hours, but there might be additional delays if the request volume is too high. 

To manage resources fairly, each eligible author is granted one virtual “voucher” to have a single paper run through the AI feedback system. We define an eligible author as someone who has authored at least one published paper at one of the following venues *hosted on OpenReview*: ICLR, NeurIPS, ICML, CVPR, or AISTATS.

To redeem this voucher, authors will select a button “Ready for LLM Feedback” on their paper submission form to flag the manuscript for AI review. This feature becomes active only after a PDF has been successfully uploaded to the OpenReview server. The author who checks the button redeems their voucher for the paper, and the version of the document at the time of the edit is then sent to the pipeline for automated feedback. If an ineligible author or an author who has already used their voucher attempts to select the “Ready for LLM Feedback” button, an error message will appear and the paper will not be sent out for review. 

The technical staff will be able to provide limited, best-effort support on the program, such as answering questions or checking for failed paper delivery. Please direct such questions, as well as any feedback you may have on the program, to icml-paperassistant@google.com.  

Privacy and Data Safety

We recognize the sensitivity of unpublished research. Trust is the cornerstone of this experiment, and we have implemented strict protocols to ensure author safety:

  1. Strict Separation from Peer Review: The AI Feedback is entirely independent of the official ICML review process. It is visible only to the authors. Reviewers, Area Chairs, and Program Chairs will have no access to this feedback.
  2. Stateless Inference (No Training): Submissions will not be used to train, fine-tune, or improve Google’s models. The model operates in a stateless “inference-only” mode; it processes the text to generate feedback and retains no memory of the specific content for future learning.
  3. Data Destruction: To minimize data exposure, Google will employ a strict deletion policy. All PDFs and feedback submitted to Google are stored in a restricted access environment and are scheduled for permanent deletion shortly after the feedback is delivered and the program is completed.
  4. Restricted Access: Google access to the data (submission PDFs and generated feedback) during the retention window is strictly limited to the primary program organizers and a minimal number of essential technical staff (listed at the bottom of this post) required for system maintenance. Google organizers and staff will only inspect the data in the event of technical difficulties with a submission (and if necessary to resolve those difficulties). 

Transparency: The AI feedback will be clearly flagged with a disclaimer stating it is AI-generated. It is framed strictly as “feedback” rather than a “review” to manage author expectations. It will be posted as a comment on the submission page that is only visible to authors.

Caveats and Disclaimers

Like all LLMs,  the models used by the PAT pipeline are not infallible. Authors should treat the generated feedback with the same critical eye they would apply to a human review.

  • The model may occasionally flag correct statements as errors or miss actual flaws. It is the author’s responsibility to verify the validity of the feedback.
  • Note that the model may make suggestions for the paper that you disagree with. This is not necessarily a bug, and may even be expected. Think of these suggestions as potential points of criticism from a reviewer. By considering why you disagree with the suggestion, you may be able to add justification to the paper which preempts these reviewers’ comments, thereby increasing the strength of the paper.

Outcomes

Our primary objective is to empower authors to elevate the quality of their submissions by acting as a high-precision filter. This tool is designed to help authors catch nuanced errors—in proofs, experimental setups, or reasoning—that human reviewers might miss or lack the time to detail.

To achieve this, the program utilizes models that are more fit for purpose than standard generalist LLMs. We are leveraging a reasoning-focused pipeline related to the Gemini models that achieved Gold-medal performance at the International Mathematical Olympiad (IMO). By designing a pipeline specialized for ML conference structures, we can offer distinct advantages over generic tools. In the post-mortem for the STOC experiment, authors consistently confirmed that these confidential models provided quality meaningfully superior to public alternatives.

We believe this tool represents a unique opportunity to pioneer the responsible use of AI to elevate the standard of our own scientific submissions. We look forward to seeing the results of this experiment, which will be shared with the broader community via the ICML blog once the experiment is over.

Google Program Organizers: Rajesh Jayaram Vincent Cohen-Addad 

Google Sponsors: Vahab Mirrokni

Google Technical Staff: Rajesh Jayaram Vincent Cohen-Addad Drew Tyler Jieming Mao Jon Schneider

Contact (for feedback and best-effort assistance): icml-paperassistant@google.com 

Load More