Back to Devoured - June 04, 2026
DEVOURED
Sleep for Continual Learning

Sleep for Continual Learning

AI researchllmmachine-learning arXiv
Google researchers propose a "Sleep" paradigm for large language models, allowing them to consolidate short-term memories into long-term knowledge and self-improve through "Dreaming."
What: Google researchers Ali Behrouz, Farnoosh Hashemi, and Vahab Mirrokni introduced a "Sleep" paradigm in an arXiv paper (2606.03979, submitted June 2, 2026) for LLMs to continually learn. It involves "Memory Consolidation" (Knowledge Seeding via distillation) and "Dreaming" (reinforcement learning to generate synthetic curricula).
Why it matters: This research explores bio-inspired mechanisms for AI, suggesting a path for LLMs to overcome catastrophic forgetting and achieve more robust, autonomous continual learning, potentially reducing the need for constant retraining on new data.
Deep dive
  • Researchers from Google, Ali Behrouz, Farnoosh Hashemi, and Vahab Mirrokni, published a paper titled "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories" on arXiv (2606.03979) on June 2, 2026.
  • The paper introduces a "Sleep" paradigm for Large Language Models (LLMs) to enable continual learning, drawing inspiration from human learning processes.
  • The "Sleep" process has two main stages: Memory Consolidation and Dreaming.
  • Memory Consolidation involves "Knowledge Seeding," an upward distillation process where memories from a smaller network (smaller-self) are transferred to a larger network.
  • This distillation uses a new Generalized Distillation process, combining on-policy distillation with Reinforcement Learning (RL)-based imitation learning.
  • Dreaming is a self-improvement phase where the model uses RL to generate a curriculum of synthetic data.
  • This synthetic data is used to rehearse new knowledge and refine existing capabilities without human supervision.
  • Experiments showed the importance of the sleep stage for long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks.
  • A version of this work has been publicly available since September 2025 on OpenReview.
Decoder
  • Distillation: In machine learning, a technique where a smaller, simpler "student" model is trained to mimic the behavior of a larger, more complex "teacher" model, often to compress knowledge or improve efficiency.
  • In-context learning: The ability of a large language model to learn new tasks or adapt to new information provided within the input prompt, without requiring retraining or fine-tuning of its parameters.
  • Catastrophic forgetting: A phenomenon in neural networks where learning new information causes the model to rapidly forget previously learned information.
  • Reinforcement Learning (RL): An area of machine learning where an agent learns to make decisions by performing actions in an environment and receiving rewards or penalties.
Original article

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.

Submission history

[v1] Tue, 2 Jun 2026 17:56:55 UTC (2,961 KB)

Access Paper:

Current browse context:

cs.LG

References & Citations

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Explore all 55 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi