Implicit generation and generalization methods for energy-based models

In three linesOpenAI reports progress on stable and scalable training of energy-based models (EBMs) with improved sample quality and generalization. EBM generation refines answers iteratively, producing samples competitive with GANs at low temperatures while maintaining mode coverage guarantees of likelihood-based models.

Summary generated by Claude — human-verified