Archive

Bebop's TV loss pushes MTP acceptance to 95%, delivering 1.8x RL speedup

The work removes the rollout stage as the key bottleneck in RL training pipelines by showing that a pre-RL MTP training recipe with TV loss and rejection sampling sustains high acceptance rates throughout RL without costly online updates, delivering up to 1.8x end-to-end acceleration.