ICML 2026 Award Papers Announced: Diffusion Models Sweep Outstanding Papers, DeepMind Classic Wins Test of Time Award

On July 5, 2026, the International Conference on Machine Learning (ICML 2026) announced its best paper awards in Seoul, South Korea. A total of 10 awards were presented, including 2 Outstanding Paper Awards, 1 Outstanding Position Paper Award, 5 Outstanding Paper Honorable Mentions, 1 Outstanding Position Paper Honorable Mention, and 1 Test of Time Award. ICML, together with NeurIPS and ICLR, is one of the top three AI conferences. This year marks the 43rd edition, held from July 6 to 11 at the COEX Convention & Exhibition Center in Seoul. Submissions exceeded 10,000, with an acceptance rate below 30%.

Outstanding Paper Awards: Diffusion Models Dominate

Both Outstanding Papers focus on diffusion models, signaling a shift from "proof of concept" to "deep waters" in the field.

• "The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models"

  • Authors: Tsinghua University Huang Gao's team, Alibaba Zanlin Ni, et al.
  • Key finding: The arbitrary-order generation capability of diffusion large language models (dLLMs) actually limits reasoning potential in general reasoning tasks like math and programming. Models tend to bypass high-uncertainty tokens, causing premature contraction of the solution space. The researchers propose JustGRPO, which abandons arbitrary order and uses standard GRPO, achieving 89.1% accuracy on GSM8K while retaining parallel decoding capability.

• "High-accuracy sampling for diffusion models and log-concave distributions"

  • Authors: MIT, Yale University Fan Chen, et al.
  • Contribution: Proposes a class of algorithms for sampling from diffusion models. Given a score estimate with accuracy ε, the algorithm achieves error ε in Õ(d log(1/ε)) steps, an exponential improvement over previous results. Complexity depends on the intrinsic data dimension d, which can be further reduced under non-uniform conditions.

Outstanding Position Paper Award: Dual-Use Risks of Alignment Techniques

• "Position: The Alignment Community is Unintentionally Building a Censor's Toolkit"
  • Authors: LMU Munich Sarah Ball, independent researcher Phil Hackemann.
  • Thesis: Modern AI alignment methods (e.g., RLHF, Constitutional AI), while intended to prevent harmful outputs, are dual-use technologies easily exploited by malicious actors for censorship and manipulation. The authors call on the community to acknowledge this risk and propose mitigation strategies.

Test of Time Award: DeepMind Classic Reinforcement Learning Algorithm

• "Asynchronous Methods for Deep Reinforcement Learning" (A3C algorithm)
  • Authors: DeepMind team (Volodymyr Mnih et al.), published at ICML 2016.
  • Significance: Proposed an asynchronous deep reinforcement learning framework that significantly improved training efficiency, profoundly influencing subsequent algorithms such as A2C and IMPALA.

Outstanding Paper Honorable Mentions (5 papers)

• "The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes" (FAR.AI): Studies where honesty emerges during RLVR training, identifying two deception strategies: activation obfuscation and policy obfuscation.
• "Motion Attribution for Video Generation" (NVIDIA, Princeton University, MIT): Proposes the Motive framework, the first to attribute motion in video generation, guiding data filtering to improve temporal consistency.
• "How much can language models memorize?" (Meta FAIR, Google DeepMind, Cornell University, NVIDIA): Proposes a new method to distinguish memorization from generalization, quantifying language model capacity.
• "Diffusion Model Alignment via Stochastic Matrix Perspective" (specific institutions not detailed): Studies diffusion model alignment from a stochastic matrix perspective.
• "Understanding Grokking: Provable Grokking in Ridge Regression" (specific institutions not detailed): Provides a provable explanation of the "grokking" phenomenon in ridge regression.

Outstanding Position Paper Honorable Mention (1 paper)

• "Position: AI/ML Deepfake Research is at Odds with AIG-NCII" (Li Qiwei et al.): Points out the disconnect between deepfake research and the misuse scenario of AI-generated non-consensual intimate images (AIG-NCII).

Selection Process

The program committee selected 53 candidate papers from all submissions, ensuring balanced coverage across eight primary topic areas. After further review, the pool was narrowed to 22 papers, and the final winners were chosen by an 11-member Outstanding Paper Selection Committee through anonymous cross-review.