Stabilizing Reinforcement Learning for Diffusion Language Models

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kStableDRL (SDAR-8B, static sampling)#46Accuracy: 92.4
mathematical-reasoning-on-math-500StableDRL (SDAR-8B, static sampling)#94Accuracy: 79.2
mathematical-reasoning-on-aime24StableDRL (SDAR-8B, static sampling)#126Acc: 16.7