Language Models Might Need "Sleep"
ai · x · 2026-07-14
This paper proposes that language models do not need to be "frozen" after training. Instead, they can periodically enter an offline "sleep" phase to consolidate fragile context memory into long-term parameters while generating synthetic data to review knowledge and continue improving capabilities.
The proof-of-concept in the paper shows that this approach outperforms SFT and GRPO on multiple math benchmarks; it achieves 80% on few-shot abstract reasoning, higher than SEAL's 72.5%; and it approaches a perfect score on the BABILong sequence test with up to 10 million tokens. Based on this, the authors envision a different AI lifecycle: models can continuously learn, sleep, integrate experiences, and "wake up" with stronger capabilities.
More from Research
- Paper: data schedules and learning rate schedules should be designed together — alifcoder · 2026-10-27
- Open Models Match GPT-6 on MedMCQA but Trail 17 Points on Hard Clinical Cases — keunwoochoi · 2026-10-27
- Health-Focused Benchmark Reranks LLMs: Muse Spark Open Jumps from #9 to #1 — keunwoochoi · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- Goodfire commits $10M to a 'Human Genome Project for AI' to reverse-engineer frontier models — CatAstro_Piyush · 2026-10-27
- π0.5 study: raising demos from 10 to 438 lifts Place Ball in Bowl success from 7% to 83% — stingning · 2026-10-27