From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs

Benchmark Model Rank Results
code-generation-on-mbppNBDiff-7B (Instruct)#14Accuracy: 87.6
math-word-problem-solving-on-mathNBDiff-7B (Instruct)#14Accuracy: 84.0
code-generation-on-humanevalNBDiff-7B (Instruct)#33Pass@1: 89
arithmetic-reasoning-on-gsm8kNBDiff-7B (Instruct)#62Accuracy: 91.0
multi-task-language-understanding-on-mmlu-proNBDiff-7B (Instruct)#68Accuracy: 71.9