Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

2026-06-02

AI Summary

This work addresses the challenge that current large language models struggle to convert short-term contextual knowledge into long-term memory and lack sustained learning capabilities. Inspired by human sleep mechanisms, the authors propose a novel “sleep” paradigm comprising two phases: memory consolidation and “dreaming.” In the consolidation phase, knowledge is distilled from a smaller model into a larger one through knowledge seeding; in the dreaming phase, synthetic data generated via reinforcement learning enables unsupervised self-improvement. This framework is the first to integrate sleep-inspired mechanisms into large models, combining policy distillation, generalized distillation, and curriculum learning to support self-modification and continual learning. Experiments demonstrate significant performance gains in long-horizon tasks, knowledge integration, and few-shot generalization, validating the efficacy of the proposed sleep mechanism.

Abstract

The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.

Problem

• continual learning
• memory consolidation
• language models
• long-term knowledge
• in-context learning

Innovation

• Sleep paradigm
• Memory Consolidation
• Knowledge Seeding
• Dreaming
• Continual Learning