Best-of- ∞ \infty – Asymptotic Performance of Test-Time Compute

Junpei Komiyama, Daisuke Oba, Masafumi Oyamada · 2025

Read Best-of- ∞ \infty – Asymptotic Performance of Test-Time Compute by Junpei Komiyama, Daisuke Oba, Masafumi Oyamada (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-21091.txt.

Description

Asking a language model the same question many times and taking the majority answer works well, but infinite tries are impossible. Junpei Komiyama, Daisuke Oba, and Masafumi Oyamada analyze that limit in a 2025 paper set to appear at ICLR 2026. They call the limiting scheme best-of-infinity.

The authors characterize how majority-vote performance behaves as the number of samples grows without bound. They then propose adaptivity: choose how many samples to draw based on how much the answers agree. Agreement signals that further sampling is wasted. They also extend the framework to weighted ensembles of several models, computing the best weights as a mixed-integer linear program. Experiments show such mixtures can beat every individual model inside them.

Catalog record

CREATORJunpei Komiyama, Daisuke Oba, Masafumi Oyamada
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-21091

Browse the full library · About the Tower of Babel