正文 · AI 翻译
Thanks to IT Home reader 咩咩洋 for the tip!
IT Home reported on October 30 that the Google Research team published a blog post on October 27 introducing a continual learning research project called the “Sleep Framework,” which aims to address the difficulty large language models face in absorbing new information once their knowledge becomes fixed after release.
Google Research made the study public on October 27. The related paper is titled “Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories.”
Google noted that, under normal circumstances, today’s mainstream large models enter a relatively static deployment state after pretraining, and their knowledge has a “cutoff date.” If they rely on in-context learning, new information is lost when the session ends. Repeatedly fine-tuning them directly on new data can cause “catastrophic forgetting,” with their existing capabilities declining markedly after they learn new knowledge.
Google’s approach draws on memory consolidation during human sleep, dividing model operation into a “wake phase” and a “sleep phase.” While awake, the model receives and processes input, temporarily storing information in short-term memory. In the sleep phase, it pauses processing new input and instead consolidates memories and improves itself.
Specifically, the “memory consolidation” step uses parameter expansion, mixture-of-experts memory blocks, and knowledge distillation to write fragile short-term contextual information into long-term parameters while avoiding overwriting existing knowledge. In the “dreaming” step, the model generates and selects synthetic training examples, then practices repeatedly through reinforcement learning to improve its reasoning and generalization abilities.
Experiments showed that the framework outperformed several baseline methods on tasks including continual translation, long-context understanding, factual knowledge acquisition, and mathematical reasoning: the model could handle contexts of up to about 10 million tokens and maintained a near-zero performance decline across multiple stages of continual learning involving previously unseen languages. It also achieved higher accuracy on mathematical benchmarks such as AIME 24/25.
IT Home provides reference links
The sleep framework: A new way for LLMs to retain knowledge in continual learning
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories
Advertising notice: External links in this article (including but not limited to hyperlinks, QR codes, and access codes) are intended to provide more information and save readers time in selecting sources. The results are for reference only. All IT Home articles containing external links include this notice.