APO: Unsupervised Atomic Policy Optimization for 3D Structure Prediction of Atomic Systems
APO claims unsupervised self-correction can beat supervised coordinate alignment for atomic 3D prediction.
The paper proposes Atomic Policy Optimization, an alignment method that does not require ground-truth reference structures. It uses group-relative policy optimization with rewards for latent structural consistency and thermodynamic stability. The authors report stronger match rates and structural fidelity than fully supervised baselines on crystal and antibody structure benchmarks. They also say APO straightens probability paths, improving inference efficiency. ArXiv · AI/CL/LG's note
The paper proposes Atomic Policy Optimization, an alignment method that does not require ground-truth reference structures. It uses group-relative policy optimization with rewards for latent structural consistency and thermodynamic stability. The authors report stronger match rates and structural fidelity than fully supervised baselines on crystal and antibody structure benchmarks. They also say APO straightens probability paths, improving inference efficiency. ArXiv · AI/CL/LG's note
score 5