Modular Pretraining Method Controls AI Model Capabilities for Dual-Use Scenarios

AI developers face a significant challenge in managing the "dual-use" nature of powerful models, where beneficial capabilities could also be misused. Current solutions, like training separate models for different access levels, are prohibitively expensive. This new research introduces a novel pre-training technique called Gradient-Routed Auxiliary Modules (GRAM) to tackle this. GRAM works by integrating specialized modules into a neural network during pre-training. These modules are selectively updated to learn specific capabilities. During inference, certain modules can be "ablated" or removed, effectively disabling particular functionalities without retraining the entire model. This approach allows for fine-grained access control, ensuring that only trusted deployments can access sensitive or potentially harmful AI capabilities. Experiments across various domains, including virology and cybersecurity, demonstrate GRAM's effectiveness in disabling targeted capabilities while preserving others. It also shows better resistance to recovery under fine-tuning compared to post-hoc unlearning methods. A key finding is that the cost of GRAM training is independent of the number of capability profiles, offering significant cost reductions compared to traditional data filtering methods, especially as models scale.