BREEZE is a new zero-shot reinforcement learning framework that improves the stability, expressivity, and performance of pre-trained generalist policies, achieving state-of-the-art results on benchmark tasks.