Do Models Need Sleep? CMU Paper Lets LLMs Consolidate Memory During “Sleep”

模型也需要「睡觉」？CMU新论文让LLM在梦中「巩固记忆」

CMU and the University of Maryland propose Language Models Need Sleep: when each L-token context window fills, the model runs N offline recurrent forward passes and updates SSM fast weights before evicting the KV cache. On GSM-Infinite, Jet-Nemotron 2B with 6 sleep loops improves 6-step arithmetic accuracy from 0.742 to 0.812.

Why it matters: HKR-H/K/R all pass: the hook is strong, and the post gives a testable mechanism plus Jet-Nemotron 2B numbers. It is still a single early paper, not an industry-level release, so it stays just above the featured threshold.