The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kJustGRPO (LLaDA 8B Instruct)#72Accuracy: 89.8
code-generation-on-mbppJustGRPO (LLaDA 8B Instruct)#102Accuracy: 52.4
code-generation-on-humanevalJustGRPO (LLaDA 8B Instruct)#103Pass@1: 49.4
mathematical-reasoning-on-math-500JustGRPO (LLaDA 8B Instruct)#123Accuracy: 45.2