LLMs Might Need "Sleep" and "Dreaming"
behrouz_ali · x · 2026-07-14
This post references research on continual learning and memory consolidation: 《LLMs Need Sleep and Dreaming!》.
核心想法
- Models shouldn't just "remember everything"; they need 记忆巩固 like sleep: retaining important information, discarding noise, and converting short-term memory into long-term knowledge through repeated "practice."
- The authors call this step a new form of knowledge distillation, Knowledge Seeding (KS), where a smaller model distills knowledge into a larger one.
实验结论
- This "sleep/dreaming" phase helps models achieve better performance on 持续学习 and reasoning tasks.
- It also 相对缓解灾难性遗忘.
The post emphasizes that when shifting from traditional ML to continual learning, even basic concepts like "training/testing time" need to be re-evaluated.
More from Research
- arXiv moderator's tip: accept the classifier's primary category or risk a hold — tdietterich · 2026-10-27
- arXiv has 4,000+ papers on hold amid submission surge; moderator shares how to avoid it — tdietterich · 2026-10-27
- CoreWeave, You.com and NVIDIA lift BrowseComp 22.9% in 3-week RL post-training sprint — RichardSocher · 2026-10-27
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27