Modular Pretraining Enables Access Control
Proposes GRAM, a pretraining method using gradient-routed modules so dual-use capabilities can be removed at inference, offering cheaper access control than data filtering and resisting finetuning recovery better than unlearning.