New Unlearning Method Reduces Knowledge Rebound

HaydnBelfield · x · 2026-07-10

The cited content discusses a new unlearning method called GRAM. While common methods only achieve partial forgetting and often see harmful knowledge return after fine-tuning, GRAM exhibits a much lower rebound rate, closely approaching the ideal effect of "data filtering." The poster emphasizes that if this technology can scale, it will significantly impact model deployment, access control, and safety verification.

Original post →

More from Safety

Safety channel →