ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-28 17:02

谷歌LLM睡眠框架

4 steps1 pieces3 Sources

【背景】大语言模型部署后常因灾难性遗忘或上下文限制而难以持续学习。2026年10月,Google Research 提出“睡眠框架”,通过模拟人类睡眠机制,让 LLM 在主动交互(醒)与离线巩固(睡)间切换。该框架利用混合专家(MoE)参数扩展和广义知识蒸馏将短期记忆转化为长期参数,并通过“做梦”阶段生成合成样本自我提升。作者自测显示其在 BABILong 基准上实现 1000 万 token 近完美扩展,并在数学推理任务中超越 SFT 和 GRPO 基线。目前该成果主要基于作者自报数据,业界对其通用性及与其他方法(如 OPSD)的对比优势仍在验证中。

How it got here

  1. 2025-09-01

    相关论文版本在 OpenReview 公开可用(据 arXiv 页面注释)。

    arxiv.org ↗
  2. 2026-07-10

    论文《Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories》提交至 arXiv。

    alphaxiv.org ↗
  3. 2026-10-27

    Google Research 官方博客发布文章介绍“睡眠框架”,详述其通过周期性离线巩固解决灾难性遗忘的机制。

    research.google ↗
  4. 2026-10-28

    Disconfirm 收录并总结 Google Research 关于 LLM 睡眠框架的报道,强调其在持续翻译和长上下文中的表现。

    research.google ↗

What is still unsettled

  • 该框架在真实生产环境中的计算开销与延迟影响如何?
  • 除作者自报外,是否有独立复现结果支持其在更广泛任务上的泛化能力?
  • 与现有的在线持续学习方法(如 Titans, SEAL)相比,其核心优势是否仅体现在特定基准测试中?

Sources

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources
  • research.googleGoogle Research Blog↗
  • arxiv.orgarXiv Paper↗
  • alphasignal.aiAlphaSignal News↗

Coverage in this topic

  • Google Research's Sleep Framework: Enabling Continuous Learning for LLMs via Periodic Offline Consolidation Without Catastrophic Forgetting2026-10-28News

    Ali Behrouz and Vahab Mirrokni of Google Research propose a 'sleep framework' that allows large language models to cycle between active interaction and offline consolidation: expanding parameters via mixture-of-experts to write new knowledge, compressing short-term context into long-term parameters using generalized knowledge distillation and reinforcement learning-based imitation, and self-improving through synthetic practice samples generated by random expert activation. The authors' self-tests report near-zero forgetting in continual translation, near-perfect scaling up to 10 million tokens on the BABILong benchmark, and performance exceeding SFT and GRPO baselines on AIME and HMMT mathematical reasoning tasks.