<ul><li>Generative models traditionally require large datasets for training but face challenges in fields with data scarcity, such as molecular modeling and physics-based inference.</li><li>Meta AI has introduced Adjoint Sampling, a new learning algorithm that enables training generative models using only scalar rewards, derived from complex energy functions.</li><li>Adjoint Sampling is based on stochastic optimal control (SOC) and reframes the training process as an optimization task over a controlled diffusion process, eliminating the need for explicit data.</li><li>The algorithm achieves state-of-the-art results in synthetic and real-world tasks, particularly excelling in molecular conformer generation, demonstrating its scalability and performance in diverse molecular structures.</li></ul>

Sampling Without Data is Now Scalable: Meta AI Releases Adjoint Sampling for Reward-Driven Generative Modeling

Discover more