RightSignal
1 of 6

Your RightSignal

AI

SO-Arena: testing how reliable fast-answer AI models really are

Lichen Li, Ziping Dong, Sicheng Feng, Qi Li, Xinchao Wang · Submitted Nov 11, 2026 · 137 upvotes yesterday
Preprint · not peer reviewedTrending

The authors built a benchmark for 'System One' models, which answer quickly without writing out long reasoning. It covers 11 domains and checks whether a model's choices stay stable when the options change, and whether it notices when no option is right.

Key finding: Across 21 models, the best overall performer ranked only seventh for reliability.

Why it matters: A high score does not mean a model's decisions can be trusted. Builders should test both.