The sleep framework: A new way for LLMs to retain knowledge in continual learning
Summary
The Sleep framework introduces a sleep phase in continual learning for LLMs, enabling continuous knowledge consolidation by alternating between active wake and offline sleep phases. It employs memory consolidation with a Mixture-of-Experts architecture, knowledge seeding via generalized knowledge distillation, and a dreaming stage with RL refinement to improve long-context retention and reasoning, aiming to mitigate catastrophic forgetting. The approach claims stable, autonomous self-improvement without altering the core model architecture.