The ICML 2026 Outstanding Paper Award has been officially announced, with two papers on diffusion models jointly topping the list, and many of the authors are Chinese.
The ICML 2026 awards have been announced!
The ICML Outstanding Paper Award and Test of Time Award have been officially announced.
A total of nine outstanding papers were shortlisted, including seven research papers and two position papers, with three grand prize winners and six honorable mentions; the ICML Test of Time Award went to the reinforcement learning field, further cementing DeepMind’s classic masterpiece as a landmark achievement.
Complete list of winners:
https://blog.icml.cc/2026/07/05/announcing-the-icml-2026-awards/
ICML, short for the International Conference on Machine Learning, along with NeurIPS and ICLR, is one of the three premier conferences in the AI field, with over ten thousand submissions annually and an acceptance rate of less than thirty percent.
ICML 2026 will be held from July 6 to 11, 2026, at the COEX Convention & Exhibition Center in Seoul, South Korea.
The Outstanding Paper Award is the Oscar of the machine learning field.
The value of this list lies not only in recognizing technical contributions but also in sending a directional signal to the entire field.
Diffusion models emerged as this year's biggest winners, with two related papers receiving Outstanding Paper awards:
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models. This groundbreaking work provides a deep analysis of the key mechanisms in diffusion large language models.
High-precision sampling for diffusion models and log-concave distributions: achieves a major breakthrough in algorithmic accuracy.
The Outstanding Paper Award in stance papers describes a peculiar phenomenon in the field of AI safety: the alignment community is inadvertently building a toolkit of moderation tools.
Five research papers received honorable mentions for the Outstanding Paper Award:
- Confusion map: Deceiving probes to map the emergence of honesty in RLVR
- Motion attribution in video generation
- How much content can a language model remember?
- Diffusion Model Consistency: A Random Matrix Perspective
- Understanding Grokking: Provably Grokking in Ridge Regression
A position paper received an honorable mention for the Outstanding Paper Award:
Position: AI/ML deepfake research opposes AI-generated non-consensual intimate images (AIG-NCII)
Finally, the Time-Tested Award goes to the absolute hit of the year:
Asynchronous methods for deep reinforcement learning
Congratulations to the above winners.
Diffusion models dominate outstanding papers; the double award reflects a new consensus.
The two award-winning papers for the Outstanding Paper Award both focus on diffusion models.
It is rare in ICML’s history for two papers on the same topic to win awards simultaneously. Behind this coincidence lies a collective judgment: diffusion models have entered a phase requiring correction and infrastructure reinforcement.
The first paper, by Huang Gao’s team from Tsinghua University and Zanlin Ni et al., has a title with real edge: “The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models.” Just from the title, it’s clear this is a direct challenge.
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
ICML: https://icml.cc/virtual/2026/oral/71086
Project homepage: https://nzl-thu.github.io/the-flexibility-trap/
First, let me explain the context.
Diffusion language models are one of the hottest research directions today. Unlike autoregressive models such as GPT and Claude, which generate text token by token from left to right, diffusion language models produce complete text gradually by denoising it from an initial pool of noise, much like painting.
Theoretically, this architecture has a major advantage: the order of generation can be arbitrary. You can write the middle section first, then the beginning; you can establish the conclusion first and then add supporting arguments—anything works.
Sounds great. But the paper by Ni et al. put a damper on things.
Through extensive experiments, they demonstrated that the so-called "arbitrary order generation" not only fails to deliver the expected benefits in actual training but also becomes a trap.
Flexibility itself comes at a cost. By supporting all possible generation orders, the model performs worse on each specific order.
The impact of this conclusion lies in the fact that it undermines the core selling point of diffusion language models.
Over the past two years, numerous papers have treated "arbitrary order" as the key argument for why diffusion LLMs outperform autoregressive LLMs, with many teams investing substantial computational resources in experiments based on this assumption. Now, ICML has officially confirmed: this argument does not hold up.
The second award-winning paper, by Fan Chen et al., focuses on sampling accuracy in diffusion models.
High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
ICML: https://icml.cc/virtual/2026/oral/71132
Preprint: https://arxiv.org/abs/2602.01338
They proposed a higher-precision sampling method for diffusion models and log-concave distributions.
It addresses the fundamental bottleneck in practical deployment of diffusion models: the theoretical upper limit on generation quality.
Two papers: one dismantled the core assumption, and the other raised the technological ceiling.
ICML rewards both breakthroughs and foundational work—the message is clear: diffusion models are moving from "proof of concept" into the "deep water," requiring not more gimmicks, but more thoughtful evaluation and solid infrastructure.
The most explosive award goes to the sharpest criticism.
Let’s return to the paper that silenced the room.
Sarah Ball and Phil Hackemann's "Position: The Alignment Community Is Unintentionally Building a Censorship Toolkit" won the Outstanding Position Paper Award.
Title: Position: The Alignment Community Is Unintentionally Building a Censor’s Toolkit
ICML: https://icml.cc/virtual/2026/oral/71119
Paper: https://openreview.net/pdf?id=dy2HwmOvFX
The ICML Position Paper Award is specifically given to papers that do not conduct experiments or run data, but instead raise fundamental questions about the direction of the field.
The core argument of this paper is blunt to the point of being jarring: researchers in the field of AI safety and alignment, who set out to make AI safer and more controllable, are developing technical tools—RLHF, Constitutional AI, value alignment frameworks—that are being systematically repurposed as infrastructure for content censorship.
Those who think they're building a secure lock are actually using blueprints that can also be used to build a prison.
This judgment is not without basis. Over the past year, controversies surrounding AI content moderation have continued to escalate. From Claude’s refusal-to-answer strategy to ChatGPT’s content filtering mechanisms, “over-alignment” has become a frequently used term among users’ complaints.
Every few weeks, people post screenshots on social media showing that AI refuses to answer, citing "safety" reasons, even when the query is a normal academic discussion or creative request.
Ball and Hackemann elevated this user-level frustration to the academic level: it is a structural risk inherent in the research paradigm itself.
The fact that ICML awarded this paper the Best Position Paper is itself a statement. The top conference is telling the entire alignment community: You need to pause and consider who is using your tools and how.
By the way, the honorable mentions for outstanding position papers are equally sharp.
The paper by Li Qiwei and others points out a significant disconnect between Deepfake research in the AI/ML field and AI-generated non-consensual intimate imagery.
Researchers are busy detecting deepfake videos of political figures, but have overlooked the most harmful misuse scenarios affecting ordinary people.
Honorable Mention Snapshot
Five honorable mentions of outstanding papers cover nearly all popular directions, each carving a breakthrough in its respective field.
Mohammad Taufeeque and others used "deception probes" to map the emergence of honesty in RLVR training.
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
ICML: https://icml.cc/virtual/2026/oral/71065
Preprint: https://arxiv.org/abs/2602.15515
In simple terms: At which layer did the model learn to lie?
This question is more valuable than the answer itself. If we can precisely identify the layer in the model where honesty emerges, future alignment efforts won’t need to rely on trial-and-error adjustments.
Xindi Wu et al. perform motion attribution in video generation.
Motion Attribution for Video Generation
ICML: https://icml.cc/virtual/2026/oral/71049
Preprint: https://arxiv.org/abs/2601.08828
In the video, when an object moves, did the model truly understand the laws of motion, or is it merely replicating pixel-level patterns? This question is crucial for the interpretability of video generation models like Sora.
John Xavier Morris and others questioned how much content large language models can actually remember, pinpointing the technical root of privacy and copyright controversies.
How much can language models memorize?
ICML: https://icml.cc/virtual/2026/oral/71168
Preprint: https://arxiv.org/abs/2505.24832
The model remembers your data—is that learning or plagiarism? The answer to this question may be more important than any copyright lawsuit.
Binxu Wang and others have re-examined the consistency of diffusion models from the perspective of random matrix theory.
A Random Matrix Perspective on the Consistency of Diffusion Models
ICML: https://icml.cc/virtual/2026/oral/71191
Preprint: https://arxiv.org/abs/2602.02908
After being trained on different, non-overlapping subsets of data, diffusion models often produce remarkably similar outputs when given the same noise seed. This consistency does not arise from the model memorizing identical data, but rather stems from deeper underlying reasons.
This consistency stems from a simple linear effect: the Gaussian statistics shared across different data splits are already sufficient to predict most of the content in generated images.
The most impressive work is by Mingyue Xu and others.
Title: To Grok Grokking: Provable Grokking in Ridge Regression
ICML: https://icml.cc/virtual/2026/oral/71134
Preprint: https://arxiv.org/abs/2601.19791
They provided a rigorous mathematical proof of the "aha" phenomenon on the classic ridge regression model.
Insight refers to a model suddenly acquiring generalization ability after its training loss has long since converged—like a student who has memorized formulas for six months and one morning wakes up truly understanding them.
This phenomenon has been observed many times in deep learning, but this is the first rigorous proof established on simple models.
Ten years after that DeepMind paper, it has finally received the Time Test Award.
The Time Test Award was given to "Asynchronous Methods for Deep Reinforcement Learning" by Volodymyr Mnih, David Silver, and other members of the DeepMind team.
Title: Asynchronous Methods for Deep Reinforcement Learning
Publication: https://proceedings.mlr.press/v48/mniha16.html
The A3C algorithm (Asynchronous Advantage Actor-Critic) proposed in this paper was a benchmark in the field of reinforcement learning when it was published in 2016.
The core idea is simple: instead of training with one massive process, run many small processes in parallel to explore different strategies and asynchronously aggregate gradients.
Simple, elegant, effective. This philosophy of “the greatest complexity lies in simplicity” appears even clearer today than it did a decade ago.
Ten years later, this idea has become embedded in the core of nearly all modern RL systems.
From AlphaGo to RLHF, from game AI to robotics control, A3C’s DNA is everywhere.
The absolute bestseller of its time, now a truly deserving classic masterpiece!
What signals does ICML 2026 convey?
When we lay out this year’s award winners, three key trends emerge.
First, diffusion models are currently the most densely researched area in machine learning, with outstanding papers and multiple honorable mentions, far outpacing other directions in visibility. Diffusion models have officially entered the competition for the architecture of the next generation of language models.
Second, AI safety research is undergoing internal scrutiny. The best paper award directly addresses how alignment community tools have been co-opted, while honorable mentions question blind spots in deepfake research. Academia is beginning to seriously confront a critical question: Where should the line be drawn between safety tools and censorship tools?
These signals, when overlaid, point to one conclusion: AI research is shifting from "rapid expansion" to "deep cleanup."
The list of winners for ICML 2026 is the first audit report of this cleanup.
Reference materials:
https://blog.icml.cc/2026/07/05/announcing-the-icml-2026-awards/
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by David.