LLMs Get a Bedtime: Boosting Long-Context Reasoning via Sleep Consolidation

This research introduces a 'sleep' phase for LLMs to convert temporary context into permanent weights, overcoming the scaling limits of traditional Transformer attention. By shifting context management to offline periods, it enables superior performance on complex reasoning tasks without sacrificing real-time inference speed.