LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kLLaDA MoE v2 Base–Accuracy: 83.93
code-generation-on-humanevalLLaDA MoE v2 Base–Pass@1: 50
code-generation-on-livecodebench-v6LLaDA MoE v2 Base–Pass@1: 31.86
code-generation-on-mbppLLaDA MoE v2 Instruct–Accuracy: 81.03
code-generation-on-multipl-eLLaDA MoE v2 Instruct–Accuracy: 67.52
math-word-problem-solving-on-mathLLaDA MoE v2 Instruct–Accuracy: 80.02
mathematical-reasoning-on-aime-2025LLaDA MoE v2 Instruct–Accuracy: 20.00
mathematical-reasoning-on-aime24LLaDA MoE v2 Instruct–Acc: 30.00
mathematical-reasoning-on-olympiadbenchLLaDA MoE v2 Instruct–Accuracy: 46.44
multi-task-language-understanding-on-mmlu-proLLaDA MoE v2 Base–Accuracy: 57.28
sentence-completion-on-hellaswagLLaDA MoE v2 Base–Accuracy: 77.19