{"version":"1.0","generated_at":"2026-10-29T15:52:25.618817","id":3652,"slug":"the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning","title":"睡眠框架：让大语言模型在持续学习中保留知识的新方法","summary":"Google Research 提出在持续学习中引入“睡眠阶段”，让 LLM 在活跃交互与离线记忆巩固之间交替，把脆弱的短期上下文固化为长期参数，以缓解灾难性遗忘和静态知识截止；原文称可连续获取新事实、保留最长 1000 万 token 的上下文知识并提升复杂推理。","abstract":"睡眠框架：让大语言模型在持续学习中保留知识的新方法 Google Research 提出在持续学习中引入“睡眠阶段”，让 LLM 在活跃交互与离线记忆巩固之间交替，把脆弱的短期上下文固化为长期参数，以缓解灾难性遗忘和静态知识截止；原文称可连续获取新事实、保留最长 1000 万 token 的上下文知识并提升复杂推理。 Google Research 提出在持续学习中引入“睡眠阶段”，让 LLM 在活跃交互与离线记忆巩固之间交替，把脆弱的短期上下文固化为长期参数，以缓解灾难性遗忘和静态知识截止；原文称可连续获取新事实、保留最长 1000 万 token 的上下文知识并提升复杂推理。 原贴提到：Ali Behrouz and Vahab Mirrokni, Google Research We introduce a sleep pha 来源：research.google Ali Behrouz 和 Vahab Mirrokni，Google Research：我们在持续学习中引入一个睡眠阶段，使大语言模型能够持续整合新事实并增强推理，同时缓解灾难性遗忘或静态知识截止。现代大语言模型通过预训练和上下文学习取得令人印象深刻的性能。然而，一旦部署，它们在很大程度上保持静态，以固定的知识截止日期运行，这使它们无法获得长期能力或适应新信息。虽然用新数据更新模型看起来是显而易见的解决方案，但持续的参更新常常触发灾难性遗忘，这是一个根本问题：学习新任务会严重降低在先前学习任务上的性能。或者，严格依赖上下文学习会把知识保留限制在即时上下文窗口内：一旦上下文会话结束，获得的短期信息就会被遗忘…","access_level":"member","access_label":"免费阅读","access_mode":"full","is_preview":false,"canonical_url":"https://opc.beizhux.com/content/3652/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning","html_url":"https://opc.beizhux.com/content/3652/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning","json_url":"https://opc.beizhux.com/content/3652/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning.json","published_at":"2026-10-28T01:35:00","updated_at":"2026-10-29T15:01:32","category":{"slug":"ai-business","name":"AI小生意项目库"},"source":{"site":"research.google","author":"Google Research Blog","url":"https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/"},"tags":["Google Research","MoE","大模型记忆","大语言模型","持续学习","灾难性遗忘","睡眠框架","记忆巩固"],"topics":[{"slug":"ai-side-business","name":"AI副业","url":"https://opc.beizhux.com/topics/ai-side-business","reason":"这篇内容命中「项目、小生意、变现」等主题信号。"},{"slug":"ai-tools","name":"AI工具","url":"https://opc.beizhux.com/topics/ai-tools","reason":"这篇内容命中「模型」等主题信号。"},{"slug":"ai-super-individual","name":"AI超级个体","url":"https://opc.beizhux.com/topics/ai-super-individual","reason":"这篇内容命中「学习」等主题信号。"}],"keywords":["AI小生意项目库","AI副业","AI工具","AI超级个体","工具","自动化","模型","Cursor","Claude","Agent","工作流","副业"],"questions":[{"question":"睡眠框架：让大语言模型在持续学习中保留知识的新方法主要讲什么？","answer":"Google Research 提出在持续学习中引入“睡眠阶段”，让 LLM 在活跃交互与离线记忆巩固之间交替，把脆弱的短期上下文固化为长期参数，以缓解灾难性遗忘和静态知识截止；原文称可连续获取新事实、保留最长 1000 万 token 的上下文知识并提升复杂推理。"},{"question":"这篇文章最值得关注的要点是什么？","answer":"Google Research 提出在持续学习中引入“睡眠阶段”，让 LLM 在活跃交互与离线记忆巩固之间交替，把脆弱的短期上下文固化为长期参数，以缓解灾难性遗忘和静态知识截止；原文称可连续获取新事实、保留最长 1000 万 token…；原贴提到：Ali Behrouz and Vahab Mirrokni, Google Research We introduce a sleep pha；来源：research.google"},{"question":"这篇文章和哪些AI专题相关？","answer":"它适合放在AI副业、AI工具、AI超级个体专题里阅读。 关联原因：这篇内容命中「项目、小生意、变现」等主题信号。；这篇内容命中「模型」等主题信号。；这篇内容命中「学习」等主题信号。"},{"question":"阅读这篇文章建议先理解哪些关键词？","answer":"建议先理解AI工具、工具、自动化、模型、Cursor这些关键词，再结合正文判断工具、机会或风险是否值得进入自己的工作流。"}],"terms":[{"slug":"ai-tools-term","name":"AI工具","definition":"在AI觉醒星球里，「AI工具」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/ai-tools-term","path":"/glossary/ai-tools-term","json_url":"https://opc.beizhux.com/glossary/ai-tools-term.json"},{"slug":"tools","name":"工具","definition":"在AI觉醒星球里，「工具」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/tools","path":"/glossary/tools","json_url":"https://opc.beizhux.com/glossary/tools.json"},{"slug":"automation","name":"自动化","definition":"在AI觉醒星球里，「自动化」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/automation","path":"/glossary/automation","json_url":"https://opc.beizhux.com/glossary/automation.json"},{"slug":"ai-model","name":"模型","definition":"在AI觉醒星球里，「模型」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/ai-model","path":"/glossary/ai-model","json_url":"https://opc.beizhux.com/glossary/ai-model.json"},{"slug":"cursor","name":"Cursor","definition":"在AI觉醒星球里，「Cursor」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/cursor","path":"/glossary/cursor","json_url":"https://opc.beizhux.com/glossary/cursor.json"},{"slug":"claude","name":"Claude","definition":"在AI觉醒星球里，「Claude」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/claude","path":"/glossary/claude","json_url":"https://opc.beizhux.com/glossary/claude.json"},{"slug":"agent","name":"Agent","definition":"在AI觉醒星球里，「Agent」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/agent","path":"/glossary/agent","json_url":"https://opc.beizhux.com/glossary/agent.json"},{"slug":"workflow","name":"工作流","definition":"在AI觉醒星球里，「工作流」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/workflow","path":"/glossary/workflow","json_url":"https://opc.beizhux.com/glossary/workflow.json"}],"preview_hidden_sections":[],"related_items":[{"id":3666,"title":"Alphabet Q3 财报：年化收入达 5000 亿美元，Google Cloud 增长 103%","url":"https://opc.beizhux.com/content/3666/alphabet-q3-5000-google-cloud-103","summary":"Sundar Pichai 披露 Alphabet Q3 业绩：年化收入 run rate 达 5000 亿美元，总收入同比增长 25%；Google Cloud 收入增长 103%，第一方模型 API 每分钟处理近 32B tokens，环比增长 40%；Search 查询同比增速为五年来最高，并提到 Argon…","access_level":"member"},{"id":3668,"title":"Ed Zitron 长文解析 Anthropic S-1：称其把散户当作退出流动性","url":"https://opc.beizhux.com/content/3668/ed-zitron-anthropic-s-1","summary":"评论者 Ed Zitron 批评 Anthropic 的 S-1 招股书披露口径，认为其只呈现 Q2 盈利、隐去 Q3 现金状况与付款明细，是在基本面变化前误导投资者。这是一条观点类信号，核心争议在于披露窗口的选择本身是否构成信息操纵。","access_level":"member"},{"id":3669,"title":"Anthropic IPO 文件披露现金及短期投资达 1317 亿美元，超过 Amazon 和 Microsoft","url":"https://opc.beizhux.com/content/3669/anthropic-ipo-131-7b-amazon-microsoft","summary":"公开信号显示，Anthropic 的 IPO 文件列出其现金及短期投资为 1317 亿美元，规模超过 Amazon 与 Microsoft，在美国非银行公司中仅次于 Berkshire Hathaway 和 Alphabet。","access_level":"member"}],"usage_policy":{"summarizable":true,"preferred_url":"https://opc.beizhux.com/content/3652/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning","private_fields_excluded":true}}