The International Conference on Machine Learning (ICML) has recognized two papers as ‘outstanding’ at ICML 2026 in Seoul for their research contributions to diffusion models while a 10-year-old paper on Deep Reinforcement Learning has been chosen for “Test of Time’ award.
Under the outstanding papers category the first paper selected was- “The flexibility trap: Rethinking the value of arbitrary order in diffusion language models” authored by Zanlin Ni, Shenzhi Wang, Yang Yue, Tianyu Yu, Weilin Zhao, Yeguo Hua, Tianyi Chen, Jun Song, Cheng Yu, Bo Zheng, Gao Huang.
According to the study diffusion large language models (dLLMs) can generate text in any order, a feature believed to improve speed and performance, while the authors discovered that this flexibility can actually hurt reasoning. To address this the researchers proposed JustGRPO, a training method that uses a simple left-to-right generation order for reinforcement learning while still allowing fast parallel decoding during inference.
The second paper recognized by the committee was “High-accuracy sampling for diffusion models and log-concave distributions” authored by Fan Chen, Sinho Chewi, Constantinos Daskalakis, Alexander Rakhlin.
The paper proves that diffusion models can generate highly accurate samples much more efficiently than previously thought, which can make them faster and more practical for real-world applications. The researchers have developed a new algorithm called first-order rejection sampling (FORS) that simulates rejection sampling from first-order queries alone, bypassing density evaluations that previous high-accuracy methods relied on.
The ICML outstanding paper selection committee comprised of 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. The shortlisting of candidates was done by program chairs Alekh Agarwal, Miro Dudik, Martin Jaggi, and Sharon Li.
Another paper Asynchronous Methods for Deep Reinforcement Learning (RL) published in 2016 by Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, Koray Kavukcuoglu during 33rd ICML was chosen for “Test of Time” award by the committee. The paper paved the way for asynchronous RL becoming a major factor in RL for LLM post-training and reshaped how RL is done today in practice.
The committee also acknowledged five papers in the category of outstanding paper honorable mentions.
The first is “Obfuscation atlas: Mapping where honesty emerges in RLVR with deception probes” authored by 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. White-box lie detectors can inspect the model’s internal activations, basically its thought process to identify deception.
The second paper titled “Motion Attribution for Video Generation” is a contribution by Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, Laura Leal-Taixé, Olga Russakovsky, Sanja Fidler, Jonathan Lorraine. According to the paper every machine learning practitioner knows that carefully choosing the training data can significantly improve model performance.
In order to enhance motion quality in video generation, this paper has introduced an attribution method that tracks the contribution of individual training examples to the relevant metrics on test cases.
The third paper titled “How much can language models memorize” by John Xavier Morris, Chawin Sitawarin, Narine Kokhlikyan, Chuan Guo, G. Edward Suh, Alexander M. Rush, Kamalika Chaudhuri, Saeed Mahloujifar, distinguishes between intended memorization- learning useful patterns that enable generalization and unintended memorization where models simply store and reproduce specific data from their training sets. This provides a new way to understand whether LLMs are actually learning or merely recalling information.
The fourth paper earned an honorable mention at ICML 2026 is “A Random Matrix Perspective on the Consistency of Diffusion Models” for uncovering why diffusion models often generate nearly identical images from the same random seed, even when trained on different datasets or using different model architectures. This paper is authored by Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan
Lastly the fifth paper is “To Grok Grokking: Provable Grokking in Ridge Regression” introduced by authors Mingyue Xu, Gal Vardi, Itay Safran. This paper mentioned that grokking is not unique to complex neural networks, it can happen even in very simple mathematical models.
Meanwhile the selection of the ICML 2026 outstanding position papers and outstanding position paper honorable mention was headed by the track co-chairs, Dale Schuurmans and Jerry Zhu.
The paper titled- “Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit” authored by Sarah Ball, Phil Hackemann has been recognised with outstanding position papers.
This paper challenges the assumption that the primary tool built to ensure AI does no harm can be misused. It supports this assertion with compelling, real-world evidence. According to the researchers the paper is not about any individual country or company; its scope generalizes to all authorities, present and future.
The second paper titled “Position: “AI/ML Deepfake Research is Misaligned with AI Generated Non-Consensual Intimate Imagery (AIG-NCII)” which has been rewarded under the category of outstanding position paper honorable mention has addressed an important issue which is identifying the misalignment between current deepfake research and the proliferation of AI-generated non-consensual intimate imagery (AIG-NCII). The paper is authored by Li Qiwei, Wells Lucas Santo, Sarita Schoenebeck, Eric Gilbert.
This year’s ICML conference has received a total of 23,918 paper submissions, more than double of ICML 2025. Out of these 6,352 papers were accepted, resulting in an overall acceptance rate of 26.6%.
