Post-Training — Research in post-training techniques for music generation, including preference alignment methods (such as DPO, RLHF, or KTO), reward model design and training, and reinforcement learning to improve output quality, controllability, and human preference adherence. All of our products will put artists and songwriters first, through these four principles: Partnerships with record labels, distributors, and music publishers : We’ll develop new products for artists and fans through upfront agreements, not by asking for forgiveness later.