There's a newer version of this model that takes a really cool RL based approach: https://arxiv.org/pdf/2408.08152