about training config

[zzwjames]

Hi,

Thanks for your great work and open-sourced code. I am wondering do you have any reasons for setting max_length=256 for GRPO? why dont we set larger output length?

[siyan-zhao]

Because we used 128 generation steps for a length of 256, the generation time is slow without kv caching. We find a generation length of 256 to be reasonable for training LLaDA at that moment, since most of our evaluation tasks also use generation lengths in the 128–512 range. However, to scale to harder tasks like AIME, longer reasoning lengths would be preferable.