Google details AI ‘sleep’ framework, reporting better retention through offline learning

Google Research is outlining a different way for AI to keep learning: stop taking new inputs periodically and consolidate what it has encountered. In an October 27 research overview, the team describes a “sleep” framework that transfers temporary information into lasting model weights. The authors report better knowledge retention and reasoning in their experiments.

The overview, by Ali Behrouz and Vahab Mirrokni, explains existing research rather than introducing a newly published paper. The underlying paper was submitted to arXiv on June 2 and revised July 10; its authors note that a version had been publicly available on OpenReview since September 2025.

Keeping context versus changing the model

The researchers target two limits of conventional learning. Information supplied in a model’s context can guide an immediate answer, but that alone does not make it lasting knowledge. Updating the model’s weights—the numerical values learned during training—can make information durable, but can also damage earlier capabilities. That problem is called catastrophic forgetting.

Sleep separates those jobs into stages. During the active, or “wake,” phase, the model processes inputs and holds recent information in temporary memory. The model pauses processing new inputs during sleep and directs computation toward consolidation and practice. The aim is to carry useful information into later sessions without overwriting what the model already learned.

Reserved capacity, then synthetic practice

Consolidation unlocks pre-allocated memory blocks in a mixture-of-experts model, an architecture that routes work through selected parameter groups called experts. Additional capacity gives incoming knowledge somewhere to go without disrupting older parameters. A process called knowledge seeding combines distillation—learning from a teacher model’s outputs—with reward-based imitation to transfer information into those longer-term weights.

The teacher supplies token-level feedback on sequences the student generates. An imitation reward also evaluates how closely completions match the teacher’s meaning and text. Once consolidation finishes, temporary parameters are reset, freeing that capacity for the next active phase. The system therefore does not simply retain every short-term trace unchanged.

A second stage, “dreaming,” creates practice material. Random expert activation encourages combinations of knowledge outside the model’s usual routing. The framework generates synthetic examples, filters them using a gradient-based measure of their training value, and fine-tunes on the strongest samples. It then evaluates whether that practice improves performance on downstream tasks.

The biological language has a firm boundary. Google says the architecture remains fixed and the training rules are human-defined. “Self-improvement” here means an automated post-training cycle that changes numerical weights, not a model rewriting its core code or inventing a new design.

Retention results, not a bigger prompt

In sequential translation evaluations involving unfamiliar languages including Manchu and Kalamang, the authors report a near-zero performance drop. Their comparison methods, including context-only learning and supervised fine-tuning, suffered severe forgetting. This tests a different question from whether a model can learn one new task: whether it can keep earlier learning as more tasks arrive.

On BABILong, a long-context benchmark, Google reports near-perfect scaling up to 10 million tokens, the text units models process, beating GPT-4 and Titans baselines in its comparison. The mechanism matters: the researchers describe compressing long-context information into model weights through sleep-time consolidation. That is a retention result, not a claim of conventional attention over a single 10-million-token prompt.

The overview also reports factual-incorporation scores of 48.9% in a single-task evaluation and 46.2% in a continued-pretraining evaluation. On mathematics, the sleep-enhanced model scored 79.2% on AIME 2024, 69.0% on AIME 2025 and 46.1% on HMMT 2025. The authors report improvements over the fine-tuning and reinforcement-learning baselines they tested.