Google Research proposes a sleep phase so LLMs keep learning

An AI product that reads a customer's own data cannot keep it in the context window forever. The common fix is a database and a retrieval step at request time. Google Research has proposed a different one. The model learns while it is running. It then learns again, offline, from what it saw. The team calls the offline step a sleep phase.

The post has four authors. Ali Behrouz works at Google Research and Cornell. Vahab Mirrokni is at Google Research. Farnoosh Hashemi is at Cornell. Adel Javanmard works at Google and the University of Southern California. The research paper is on arXiv.

The two phases

Google Research describes two states for the model. Awake, the model takes an input and stores short-term memory. Asleep, it does two things. It moves fragile short-term memory into stable long-term parameters. That step uses a mixture-of-experts layer that can grow its own capacity. It then runs what the authors call dreaming. Dreaming is an offline self-training loop over synthetic data, with reinforcement learning on top.

Writing in Google Research's own words, the method is "a sleep phase in continual learning". The team says it lets models "continuously integrate new facts and enhance reasoning while mitigating catastrophic forgetting."

The numbers

Google Research tests the sleep method against several baselines. Those include in-context learning, supervised fine-tuning, GPT-4, Titans, SEAL and on-policy self-distillation. The team reports 48.9 percent on SQuAD. That score is on a single factual-knowledge task. On the BABILong long-context test, the model scales to 10 million tokens. The authors say that outperforms strong baselines including GPT-4 and Titans. On AIME 24 and 25 maths problems, it scores 79.2 percent. On a continual translation test, Google Research reports near-zero drop in earlier stages after the model has moved on.

Each of these is the authors' own number on the authors' own test. Google Research has not published open weights. No outside team has reproduced the results.

Why it is a different kind of claim

A retrieval system looks up facts and puts them in the prompt. The weights do not change. The sleep framework works the other way. The weights do change, during a time the model is not serving traffic. The changes come from the inputs the model saw during the day. If this works at a product's scale, the model carries the week's work inside it. A search index on the side is no longer the only place a user's history can live.

The paper is a research proposal, not a product. The sleep step costs compute during the offline window. The mixture-of-experts layer adds parameters over time. A team that wants to use this today does not have the model, and the paper gives no price. The question for a buyer is simpler. For an assistant that must remember what a user said six months ago, is a longer context window enough? Or does the model itself need to change?

Source

The sleep framework: A new way for LLMs to retain knowledge in continual learning, Google Research. Paper: arxiv.org/abs/2606.03979.