Explorative modeling: Train on the best of K guesses

Explorative Modeling (XM) is a new generative modeling paradigm that improves existing models by training on multiple candidate outputs and selecting the best one, rather than averaging all possibilities into a blurred prediction. This approach achieves significant efficiency gains across images, video, and language—reaching 6.2× sample efficiency and 4.1× FLOP efficiency—while matching diffusion models on control tasks with 256× less inference compute. The method addresses a fundamental problem in generative modeling where direct prediction collapses to the average of all valid outputs, which typically looks nothing like real data.