The Alibaba Tsinghua ICML 2026 (the International Conference on Machine Learning) Outstanding Paper shows diffusion models' flexible token generation skips the logical connectors the model hasn't resolved yet, decorating an answer it already
The ICML 2026 Outstanding Paper Award went to a joint Alibaba–Tsinghua team on July 5 for a finding that turns a celebrated feature of diffusion language models into a liability. Their "Flexibility Trap" paper argues that the arbitrary-order generation powering models like Google Gemini Diffusion and Renmin University of China's LLaDA systematically degrades logical reasoning, and that the highest-scoring fix is to take the freedom away.
Autoregressive models like GPT and Qwen generate one token at a time, strictly left-to-right, committing to each word before the next. Diffusion language models (dLLMs) generate tokens in arbitrary order, filling in high-confidence content first and leaving uncertain pieces for later. On paper, this looks like a reasoning advantage: the model can plan around its uncertainty, sketching the easy parts of an answer and returning to the hard logical connectors only after it knows the surrounding context.
The mechanism the authors call entropy degradation explains why arbitrary-order generation hurts reasoning. When a model is allowed to generate in any order, it tends to skip the highest-uncertainty logical connectors (words like "therefore" or "so" in English, "因此" or "所以" in Chinese) and fill them in after the surrounding context has already locked in the answer. A reasoning fork collapses into a fill-in-the-blank. The model is decorating an answer it has already produced, not reasoning across the connector. The full argument and ablations are in the paper's HTML version.
HumanEval shows the asymmetry the paper relies on: 21.3% of code-generation problems are solvable left-to-right but not in arbitrary order, while only 0.6% go the other way. Greater order freedom correlates with worse reasoning performance. Those numbers are reported in the paper and its project page and have not yet been independently replicated, so the cleanest read is the authors' measurement rather than an established benchmark fact.
JustGRPO is a stripped-down version of what came before. Earlier diffusion RL methods (d1, ESPO, SPG, GDPO) built elaborate attribution pipelines to credit or blame specific tokens under arbitrary-order generation. JustGRPO uses the standard GRPO reinforcement-learning algorithm but trains the model to generate strictly left-to-right during the RL phase. Inference speed is unchanged because the deployment-time sampling order is separate from the training order. The result: 89.1% accuracy on GSM8K, surpassing the specialized diffusion RL baselines without their engineering complexity. Code and pretrained models are public.
Google, Alibaba, and a wave of academic labs have been betting that diffusion architecture will out-reason GPT-style models by escaping the left-to-right bottleneck. The Alibaba–Tsinghua result argues the opposite: flexibility and reasoning quality can move in opposite directions, and the simplest move may be to give that freedom back. The authors are clear that this is a paper-level claim: replication, broader benchmark coverage, and behavior at larger scales are the obvious next questions.
The Chinese-language coverage in Lei Phone calls the result "大道至简" (the greatest truths are the simplest). That label fits JustGRPO's design choice: drop the architectural novelty that was supposed to differentiate diffusion reasoning, keep the part that works, and beat the field on the metric everyone cares about.
Two questions will determine whether the result holds up: does JustGRPO's left-to-right training regime transfer to other dLLM families, and does the entropy-degradation finding hold when the model is scaled past the paper's reported regimes? Until those are answered, the result stands as a paper-reported counter-narrative: architectural flexibility and reasoning quality can move in opposite directions, and the simplest fix is the most uncomfortable one.