<ul><li>The paper introduces ManGO, a diffusion-based framework for offline optimization of complex systems.</li><li>ManGO learns the design-score manifold to capture design-score interdependencies holistically, aiming for generalization beyond training data.</li><li>Unlike existing methods, ManGO unifies forward prediction and backward generation for better performance in various domains.</li><li>Extensive evaluations show that ManGO outperforms numerous optimization methods in synthetic tasks, robot control, material design, DNA sequence, and engineering optimization.</li></ul>

Learning Design-Score Manifold to Guide Diffusion Models for Offline Optimization

Discover more