kapynResearch

Locking Pretrained Weights via Deep Low-Rank Residual Distillation

Deep Low-Rank Residual Distillation is a technique for locking pretrained model weights against unauthorized fine-tuning. The method freezes core weight matrices while training a low-rank residual pathway that preserves general capability while restricting unauthorized domain adaptation. This approach offers creators a way to distribute open-weight models securely without sacrificing utility or downstream flexibility.

Apple ML Research·Aug 6, 2026

Opening Kapyn…