ICML 2026: Best Papers, Awards, Dates & Research Themes

ICML 2026 is the Forty-Third International Conference on Machine Learning. The official conference runs July 6-11, 2026 at COEX in Seoul, South Korea: tutorials and expo talks on July 6, the main conference July 7-9, and workshops July 10-11.

For readers trying to catch up fast, the most important ICML 2026 signal is the awards list. The conference recognized two Outstanding Papers, one Outstanding Position Paper, five Outstanding Paper Honorable Mentions, one Outstanding Position Paper Honorable Mention, and a 2016 Test of Time winner.

ICML 2026 Quick Facts

Item	Detail
Official site	icml.cc/Conferences/2026
Dates	July 6-11, 2026
Location	Seoul, South Korea
Venue	COEX Convention & Exhibition Center
Main conference	July 7-9
Workshops	July 10-11
Official papers page	ICML 2026 papers
Awards page	ICML 2026 awards
Official awards announcement	ICML blog: awards

ICML 2026 Outstanding Papers

ICML does not label these as "best papers" on the official page; the conference calls them Outstanding Paper Awards. These are the two top research papers to start with.

Paper	Authors	Why it matters
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	Challenges the assumption that arbitrary token order is always the key advantage of diffusion language models. The result suggests that fixed left-to-right RL rollouts can improve reasoning while preserving parallel decoding at inference.
High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions	Fan Chen, Sinho Chewi, Constantinos Daskalakis, Alexander Rakhlin	Gives a theory result for score-based sampling: high-accuracy sampling can require polylogarithmic rather than polynomial dependence on the target error. If the theory carries into practice, it points toward more efficient diffusion sampling.

Outstanding Position Paper

Paper	Authors	Why it matters
Position: The Alignment Community is Unintentionally Building a Censor's Toolkit	Sarah Ball, Phil Hackemann	Argues that alignment techniques are dual-use: tools built to constrain harmful model behavior can also be repurposed for censorship. This is the sort of position paper people should read even if they disagree with it.

Honorable Mentions Worth Reading

Paper	Theme	Why it is useful
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes	Alignment, reward hacking	Studies whether training against deception probes produces honest behavior or just better evasion. Useful for anyone building agent evals.
Motion Attribution for Video Generation	Video generation, data attribution	Tracks which training examples affect video generation quality. Important for dataset pruning, rights analysis, and model improvement.
How much can language models memorize?	Memorization, generalization	Gives a sharper way to reason about what LLMs memorize versus what they generalize. Relevant to privacy, copyright, and evaluation.
A Random Matrix Perspective on the Consistency of Diffusion Models	Diffusion theory	Explains why diffusion models can produce similar outputs across runs using shared Gaussian statistics.
To Grok Grokking: Provable Grokking in Ridge Regression	Grokking, theory	Shows grokking-like behavior in a simple linear setting, giving researchers a cleaner toy model for delayed generalization.

The Outstanding Position Paper Honorable Mention is Position: AI/ML Deepfake Research is Misaligned with AI Generated Non-Consensual Intimate Imagery (AIG-NCII), a direct critique of how deepfake research often optimizes for technical detection tasks while underweighting victim harm.

Test of Time Award

ICML 2026 gave the Test of Time Award to Asynchronous Methods for Deep Reinforcement Learning, the 2016 paper by Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu.

The lasting idea is simple: parallel actor-learners stabilize deep reinforcement learning. That pattern still matters in 2026 because RL is central to LLM post-training, agent behavior shaping, and large-scale model improvement.

What ICML 2026 Says About AI Research

Diffusion is not just image generation anymore

Both Outstanding Papers are diffusion-adjacent. One looks at diffusion language models and reasoning; the other studies high-accuracy sampling theory. That combination is a strong signal: diffusion is now a live research frontier for language, reasoning, and sampling efficiency, not only media generation.

Agentic AI is crowded, but still important

The official workshop announcement says ICML 2026 accepted 44 workshops and 4 affinity workshops, and that variants of "agentic AI" appeared in at least 60 workshop submissions. That does not mean every agent paper is high quality. It means the community is trying to turn agents from product demos into measurable systems with failures, diagnostics, and guarantees.

Evaluation is shifting from benchmark scores to behavior under pressure

Several recognized papers are about what models do under stress: reward hacking, deception probes, memorization, grokking, and diffusion consistency. The practical takeaway is that AI evaluation is becoming less about a single leaderboard number and more about whether a model remains reliable when incentives, prompts, data, or training dynamics shift.

Public communication is becoming part of the research process

ICML continued its lay-summary policy for accepted papers in 2026. Authors are asked to explain the problem, solution, and impact in language an informed non-specialist can understand. This matters for AI companies too: if a paper cannot be explained clearly, it is harder for answer engines, journalists, operators, and policy teams to use it responsibly.

How to Track ICML 2026 Papers

Start with the official awards announcement.
Browse the official ICML 2026 papers page.
Check the ICML proceedings at PMLR once the proceedings volume is finalized.
Use paper titles, authors, and project pages to find code, arXiv versions, and follow-up commentary.
Prioritize papers with clear failure modes, benchmarks, or implementation details if you are building products rather than doing pure research.

FAQ

When is ICML 2026?

ICML 2026 runs July 6-11, 2026 in Seoul, South Korea. The main conference is July 7-9, with workshops on July 10-11.

What were the ICML 2026 best papers?

The official term is Outstanding Paper Award. ICML 2026 recognized two Outstanding Papers: "The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models" and "High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions."

What is the most practical ICML 2026 paper for AI builders?

For AI builders working on agents and model reliability, "The Obfuscation Atlas" is especially practical because it studies reward hacking, deception probes, and when training pressure creates honest behavior versus detector evasion.

What is the biggest research theme from ICML 2026?

The strongest theme is reliability under new generation and training regimes: diffusion language models, high-accuracy sampling, deception probes, memorization, and grokking all point toward understanding model behavior beyond surface benchmark performance.