Language Models Need Sleep: Let AI Nap Before Continuing Inference

大模型也需要睡觉！让AI打个盹，醒来更聪明

Carnegie Mellon University and the University of Maryland propose a “sleep” mechanism for language models: when the context window is nearly full, the model stops accepting new tokens, runs multiple offline recursive forward passes to compress accumulated context into fast weights, clears the KV cache, and then resumes inference; tests cover cellular automata, multi-hop graph retrieval, and GSM-Infinite reasoning tasks.

Why it matters: HKR-H/K/R all pass: the sleep metaphor is clickable, and the mechanism is concrete. Score stays below 78 because the provided body lacks benchmark gains, code, or deployment evidence.