LLMs keep learning after deployment: sleep-style offline consolidation writes new facts into weights without erasing old skills
Synopsis
Deployed large language models stay frozen at a fixed knowledge cutoff, continual parameter updates cause severe forgetting of earlier tasks, and in-context learning loses information once the session ends. Google Research's sleep framework alternates active interaction with offline consolidation, in which the model opens pre-allocated Mixture-of-Experts memory blocks and updates weights using self-generated practice examples. In multi-stage continual translation into unseen languages such as Manchu and Kalamang, performance dropped near zero, while in-context learning, Cartridges and supervised fine-tuning suffered severe forgetting.
Interpretation
A large language model is given a periodic learning lifecycle that adds an offline phase to active interaction, consolidating fragile short-term context into stable long-term parameters. Previously a deployed model stayed largely static with a fixed knowledge cutoff; direct continual parameter updates triggered catastrophic forgetting, while relying on in-context learning limited retention to the current context window. In multi-stage continual translation into unseen languages such as Manchu and Kalamang, the framework showed a near-zero performance drop, whereas in-context learning, Cartridges and supervised fine-tuning suffered severe catastrophic forgetting during sequential learning.
The offline phase runs in two stages, memory consolidation followed by dreaming-style self-improvement, entirely within a fixed model architecture and human-defined training rules. Consolidation treats knowledge seeding as a structured low-rank compression problem, selecting a compact set of directions that preserve both the model's on-policy behavior and the semantic reward, with parameter expansion supplying write capacity; dreaming randomly activates different experts to generate practice samples and filters them by gradient to keep the most informative ones. On the BABILong long-context benchmark, the method showed near-perfect scaling up to 10M tokens, outperforming strong baselines such as GPT-4 and Titans.
New facts are written in without degrading existing representations, and mathematical reasoning improves when consolidation is paired with imitation learning. On SQuAD and continued pre-training evaluation, the method reached 48.9% on single-task and 46.2% on CPT, outperforming the base model, supervised fine-tuning and SEAL. On AIME 24/25 and HMMT 25 mathematical reasoning benchmarks, the sleep-enhanced base model reached 79.2%, 69.0% and 46.1% accuracy, outperforming supervised fine-tuning, GRPO and OPSD.
Perspective
The framework operates within a fixed model architecture and human-defined training rules, and its offline phase depends on pre-allocated Mixture-of-Experts memory blocks and parameter expansion; demonstrated settings include continual translation into unseen languages, long-context understanding up to 10M tokens, fact incorporation on SQuAD and continued pre-training, and mathematical reasoning on AIME 24/25 and HMMT 25.
Terms such as sleep and dreaming are borrowed from neuroscience, yet the framework runs entirely within a fixed architecture and human-defined rules rather than simulating biological sleep; the advantages on each benchmark are evaluation results reported by the team.