Primary sourceshao__meng
View sourceEvent ·Multi-source
Google Research 发布 SLEEP 框架:让模型睡眠时把短期记忆写进参数
Google Research 发布 SLEEP 框架,让模型在清醒期处理任务、积累短期信息,在睡眠期暂停新输入并进行记忆巩固。巩固阶段采用「知识播种」设计,启用预留的低秩专家参数,通过蒸馏把短期记忆迁移进去,不直接改写承载旧知识的参数。随后的「做梦」阶段让模型生成合成练习并筛选样本强化所学。以 Qwen3-8B 为基座,SLEEP 在 AIME 2024 的 avg@16 得分为 79.2%,对照的在线策略自蒸馏为 76.6%,代价是计算开销约为普通监督微调的 4 倍。与 Anthropic Dreaming 整理外部笔记不同,SLEEP 把知识写进模型参数本身。
Current read
Google 让模型学完新东西后「睡一觉」把知识固化进参数,Qwen3-8B 上 AIME 涨到 79.2%,代价是 4 倍算力,和 Anthropic 那套路子不一样。
11 sources41° AI heatLast updated 2026/11/7 12:58:21
Evidence
11 sourcesRelated source 1METR
View sourceRelated source 2宝玉
View sourceRelated source 3Google Research
View sourceRelated source 4Gary Marcus
View sourceRelated source 5Andrew Curran
View sourceRelated source 6@koltregaskes
View sourceRelated source 7Thomas Wolf
View sourceRelated source 8rohanpaul_ai
View sourceRelated source 9Tech in Asia
View sourceRelated source 10VnExpress Số hóa
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.