Trading inference-time compute for adversarial robustness.

Wojciech Zaremba Evgenia Nitishinskaya Boaz Barak Stephanie Lin Sam Toyer Yaodong Yu Rachel Dias Eric Wallace Kai Xiao Johannes Heidecke Amelia Glaese · 2025

Read Trading inference-time compute for adversarial robustness. by Wojciech Zaremba Evgenia Nitishinskaya Boaz Barak Stephanie Lin Sam Toyer Yaodong Yu Rachel Dias Eric Wallace Kai Xiao Johannes Heidecke Amelia Glaese (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-2501-18841.txt.

Description

A word tool that thinks out loud at test time can buy grit with sheer thought span, or so a 2025 set of tests asks. Wojciech Zaremba and co-authors set two think tools, a full and a small, to a spread of foe tricks. No foe school was used: the tools just thought long or short.

Cross most tricks, long thought cut the trick's win rate, in many cases to nil as the span grew. New tricks built for think tools were tried as well. Some cases bucked the drift, and there long thought did not help, or hurt. The team read those holes for their cause, and pitch think span as a cheap grit dial with known gaps.

Catalog record

CREATORWojciech Zaremba Evgenia Nitishinskaya Boaz Barak Stephanie Lin Sam Toyer Yaodong Yu Rachel Dias Eric Wallace Kai Xiao Johannes Heidecke Amelia Glaese
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-18841

Browse the full library · About the Tower of Babel