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 a single model to mimic multiple specialized versions. It provides a technical path for granular access control in frontier models.