APO: Unsupervised Atomic Policy Optimization for 3D Structure Prediction of Atomic Systems

TL;DR - APO is an unsupervised alignment framework for predicting 3D atomic structures without ground-truth coordinates. It reportedly improves crystal and antibody structure prediction while reducing inference costs.

• Adapts group-relative policy optimization to 3D atomic environments.
• Combines rewards for dominant latent structural modes and thermodynamic stability.
• Selects physically plausible configurations from sampled groups, enabling self-correction.
• Reportedly surpasses supervised baselines in match rates and structural fidelity while straightening probability paths for faster inference.