Tsinghua researchers won the International Conference on Machine Learning's (ICML) top award for showing that arbitrary order generation, the feature diffusion language models use to outpace transformers, doubles as a reasoning shortcut.
The standout paper at this year's ICML is a diagnosis, not a new model or benchmark victory. A Tsinghua group led by Gao Huang won the Outstanding Paper Award for showing that diffusion language models use their signature feature, generating tokens in any order, to dodge the computations that make reasoning hard. The conference's awards announcement puts the result atop a record-setting program.
Diffusion language models, sometimes called dLLMs, generate text by denoising all tokens at once rather than going left-to-right the way transformers do. The pitch is straightforward: parallel generation is faster, and arbitrary-order generation, picking which token to write next on the fly, is supposed to be more flexible. "The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models," presented as an oral paper at ICML 2026, argues that flexibility has a price.
The authors show that on reasoning tasks, dLLMs systematically skip what they call "forking tokens," positions where the path forward branches and the correct continuation is genuinely uncertain. Those tokens are also the most expensive to compute: exactly where reasoning effort lives. Arbitrary-order generation lets the model defer or ignore them in favor of easier continuations, collapsing solution diversity on harder problems.
The proposed fix is a training recipe called JustGRPO. It keeps parallel decoding at inference but forces left-to-right generation during reinforcement-learning rollouts. On GSM8K with 512-sample inference, JustGRPO reaches 89.1% accuracy; on MATH-500, 45.1%. The contradiction inside the proposal confirms the diagnosis: dLLMs reason only when forced to compute in sequence, then go back to parallel generation to serve users. Details sit on the HuggingFace papers page.
That pair, Outstanding Paper #1 and its proposed mitigation, sat at the center of a program that has roughly doubled in a single year. ICML 2026, the 43rd International Conference on Machine Learning, opened on July 6 at Seoul's COEX with more than 11,000 attendees and 23,918 valid submissions, up from 12,107 in 2025. The acceptance rate has crept up too: 6,352 papers (26.6%) made the cut, with 536 Spotlight designations (2.2%) and 168 oral slots (0.7%).
The Outstanding Paper #2 sat in different territory. Fan Chen, Sinho Chewi, Constantinos Daskalakis, and Alexander Rakhlin resolved a long-standing theoretical question about score-based sampling. Their algorithm, FORS, achieves O(d·polylog(1/ε)) sampling, an exponential improvement over the previous poly(1/ε) bounds. Where Tsinghua's paper names a failure mode inside a hot architecture, FORS closes a math problem.
Across 398 reciprocal reviewers (those who also submitted papers), the program chairs recorded LLM-policy violations severe enough to trigger desk rejections. A March policy update preceded 497 papers (~2%) being desk-rejected for undisclosed LLM use. Chinese outlet Leiphone, reporting from Seoul, put the broader count at 506 reviewers and 795 violations across the full reviewer pool, not just submitters. ICML's 398 figure counts only reviewer-authors, so the two numbers describe different populations. The pattern underneath them is the same: the conference is now policing the very systems being published.
The Test of Time Award landed on a 2016 paper that defined asynchronous advantage actor-critic, A3C. Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu at DeepMind showed that a multi-thread agent running 16 CPU cores on copies of the same Atari environment could outperform a GPU-bound DQN on 57 of the benchmark's games, without a replay buffer. A3C's lasting contribution to deep RL was the recognition that gradient noise from asynchronous updates decorrelates updates the way experience replay does, a structural shortcut baked into the algorithm.
The awards were chosen by a committee of 11 chaired by Andreas Krause, with members Yoav Artzi, Léon Bottou, Michael Bowling, Jordan Lee Boyd-Graber, Marco Cuturi, Aleksandra Faust, Claudio Gentile, Amir Globerson, Manik Varma, and Yu-Xiang Wang. They winnowed 53 candidates from reviewer scores and area-chair nominations down to 22 shortlisted, then two Outstanding Papers and five Honorable Mentions.
What to watch next: the camera-ready version of "The Flexibility Trap," which replaces the arXiv v4 preprint and locks in the failure-mode definition; replication tests of JustGRPO on dLLM checkpoints from groups outside Tsinghua; and whether ICML 2026's remaining program treats the trap as a target or as a category.