New 'Sleep-time Compute' method cuts inference needs by 5x, paper shows

tomssilver · x · 2026-08-24

The paper "Sleep-time Compute" proposes letting models pre-compute useful quantities offline before queries arrive. Experiments show this approach can reduce test-time compute needs by 5x for the same accuracy on reasoning tasks. Scaling sleep-time compute further boosts accuracy by up to 18% on AIME. Additionally, amortizing compute across related queries for the same context cuts average cost per query by 2.5x.

Original post →

More from Research

Research channel →