ABOPD: Antibody CDR Design via On-Policy Distillation

ABOPD is an antibody CDR design framework based on on-policy distillation. Instead of training only on perturbed native structures, it supervises states visited during the model’s own denoising trajectories using privileged native geometry. The method targets error accumulation and backbone drift in flexible antibody loops, especially CDR-H3. On the RAbD CDR-H3 generation benchmark, ABOPD reduces RMSD from 2.37 Å to 1.95 Å, an improvement of 0.42 Å, and outperforms supervised fine-tuning and offline distillation controls according to the supplied abstract.