Language Models Need Sleep

· HF Daily Papers ·

A sleep-like consolidation mechanism converts context into fast weights, reducing long-context burden while preserving inference latency.

Categories: Research

Excerpt

Sangyun Lee, Sean McLeish, Tom Goldstein, Giulia Fanti — 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.

Read at source: https://arxiv.org/abs/2605.26099

Discussions