新智元 (微信公众号)T2·Models
72 pts
A NeurIPS 2026 paper introduces D-DOIT, which frames reward adaptation as a distribution transport problem and derives a Doob h-transform formulation for masked discrete diffusion, enabling a frozen model to target arbitrary black-box rewards without retraining or reward gradients. On HepG2 enhancer design, high-budget D-DOIT-Q reaches a predicted activity of 6.19, surpassing the fine-tuned DRAKES (5.61), with results also shown on protein inverse folding.