Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs

The paper introduces RILKE (Representation Intervention for Lifelong KnowledgE Control), a method designed to enable efficient knowledge updates in large language models (LLMs) without retraining. RILKE employs representation-space interventions to achieve fine-grained control over complex knowledge while keeping base weights frozen, utilizing paraphrase-robust and edit-localized modules to minimize interference during updates. Tested on LLaMA and Qwen models, RILKE demonstrates high edit success and paraphrase generalization across large-scale benchmarks, offering a practical solution for practitioners needing to manage evolving knowledge in LLMs.