Language Models Need Sleep
This work addresses the context length scaling problem in large language models for long-horizon reasoning tasks, offering a biologically-inspired solution that improves performance on tasks requiring deeper reasoning.
The paper proposes a sleep-like consolidation mechanism for transformer-based models that periodically converts recent context into persistent fast weights, improving performance on long-horizon tasks. On synthetic tasks and math reasoning, models with sleep outperform regular transformers and SSM-attention hybrids, with performance gains scaling with sleep duration.
Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast weights before clearing its key-value cache. During sleep, the model performs $N$ offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. During inference, this shifts extra computation to sleep while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration $N$ for our models improves performance, with the largest gains on examples that require deeper reasoning.