Event ·Single source

Google Research 发布 SLEEP:用「睡眠巩固」周期减轻大模型灾难性遗忘

Google Research 发布 SLEEP 方法,让 LLM 在「清醒交互」和「睡眠巩固」之间周期切换。睡眠阶段动态扩容 MoE 记忆块,用广义知识蒸馏(GKD)把短期上下文信息写入长时参数,再用合成数据做「做梦」式自我改进。持续翻译评测中,SLEEP 对满语、Kalamang 等未见语言近乎零掉点,遗忘程度低于 ICL、Cartridges 和 SFT 基线。长上下文方面在 BABILong 上称可扩展到约 1000 万 token,数学推理在 AIME 24 上达到约 79.2%,优于 SFT 和 GRPO。

Current read

Google Research 这个 SLEEP 挺有意思,让模型像睡觉一样离线巩固记忆,学新语言不掉老本事,BABILong 上号称能撑到 1000 万 token。

2 sources72° AI heatLast updated 2026/10/29 02:34:55

Evidence

2 sources
Related source 1Google Research
View source

Conflict 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.

Google Research Releases SLEEP Framework for Offline Memory Consolidation in LLMs · 事件专题 | AITOP