Diffusion Language Models (dLLMs) have long been celebrated for their free-form nature, allowing tokens to be generated in any order rather than the strict left-to-right sequence used by traditional models. While this flexibility sounds like a superpower, researchers from Tsinghua-LeapLab have identified what they call a Flexibility Trap. The study reveals that when dLLMs are given total freedom, they often skip over difficult, high-uncertainty tokens that are essential for logical exploration. This premature collapse of the solution space means the model takes the path of least resistance, ultimately failing at complex math or coding tasks. To bridge this gap, the team introduced JustGRPO. This method forgoes the chaotic arbitrary order and instead applies Group Relative Policy Optimization (GRPO)—a technique that refines the model's logic by comparing multiple generated answers. The result is a minimalist yet powerful framework that achieves 89.1% accuracy on the GSM8K math benchmark. Most impressively, this boost in reasoning doesn't sacrifice performance; the model fully retains its parallel decoding ability, allowing it to generate text much faster than traditional step-by-step (autoregressive) models. This finding challenges the industry's reliance on complex trajectories, proving that sometimes, less flexibility leads to smarter AI.

Diffusion Language Models (dLLMs) have long been celebrated for their free-form nature, allowing tokens to be generated in any order rather than the strict left-to-right sequence used by traditional models. While this flexibility sounds like a superpower, researchers from Tsinghua-LeapLab have identified what they call a Flexibility Trap. The study reveals that when dLLMs are given total freedom, they often skip over difficult, high-uncertainty tokens that are essential for logical exploration. This premature collapse of the solution space means the model takes the path of least resistance, ultimately failing at complex math or coding tasks. To bridge this gap, the team introduced JustGRPO. This method forgoes the chaotic arbitrary order and instead applies Group Relative Policy Optimization (GRPO)—a technique that refines the model's logic by comparing multiple generated answers. The result is a minimalist yet powerful framework that achieves 89.1% accuracy on the GSM8K math benchmark. Most impressively, this boost in reasoning doesn't sacrifice performance; the model fully retains its parallel decoding ability, allowing it to generate text much faster than traditional step-by-step (autoregressive) models. This finding challenges the industry's reliance on complex trajectories, proving that sometimes, less flexibility leads to smarter AI.