AIR: Inference-Time Refinement for Discrete-Diffusion Antibody Humanization

Antibody humanization replaces non-human framework residues with human-like ones while preserving the murine CDR loops that determine binding. Discrete-diffusion models such as HuDiff cast this as conditional sequence generation, but their one-pass sampling cannot navigate the trade-off between humanness and structural fidelity. We propose AIR, an inference-time refinement framework that lets a pre-trained discrete-diffusion model audit and revise its own predictions. Each cycle remasks residues flagged as low-quality by the model's own confidence, an external biological scorer, or both, and resamples them within a more complete sequence context. On the Humab25 benchmark, AIR traces a controllable trade-off between humanness and fidelity. Pairing external nativeness scoring with a low-rate self-consistency stage produces sequences that are as human-like as the experimentally validated humanizations and retain the same fraction of the original murine residues. AIR requires no retraining and operates as a drop-in wrapper for existing discrete-diffusion humanization models.