July 2026
An adversarial tournament design for efficiently probing the frontier of AI forecasting
Abstract
We show that existing forecasting tournaments are inefficient probes of the capability frontier: most questions are either trivially easy or effectively intractable, with only a small fraction in the regime where leading systems disagree. We propose an adversarial design that admits more discriminating question formats and makes question-writing itself competitive, rewarding writers in proportion to the disagreement their questions induce among the strongest forecasters. We instantiate this as a live tournament with a $25,000 prize pool and report early findings on the questions that survive adversarial selection.
Authors
Ben Day, Scott Jeen, Simion-Vlad Bogolin, Maximilian Anthony Hugh Clark, Toby Shevlane
Venue
ICML 2026 Workshop on Forecasting as a New Frontier of Intelligence