AdapterSwap introduces a novel approach to continuously train language models with strong guarantees for data access control and removal.

• Selective Knowledge Access: Enables user-based access controls to specific knowledge subsets
• Guaranteed Forgetting: Ensures complete removal of information when required
• Modular Architecture: Uses lightweight adapters instead of full model retraining
• Practical Implementation: Achieves these goals with minimal performance impact

For security teams, AdapterSwap offers a critical solution to data governance challenges in LLM deployments, addressing privacy regulations and access management requirements while maintaining model utility.

AdapterSwap: Continuous Training of LLMs with Data Removal and Access-Control Guarantees