dailyai.report

Gradient Routed Auxiliary Modules (GRAM) isolate sensitive knowledge into specific, switchable components within a language model. Researchers can now toggle these modules to restrict or grant access based on user trust. This approach allows one model to mimic multiple versions with varying safety constraints. It provides a concrete mechanism for granular access control.