The paper claims to evaluate six frontier LLMs prospectively before all 104 matches of the 2026 FIFA World Cup, archiving 4,494 scored predictions across seven markets, group winners, and outright picks. According to the abstract, average match-outcome accuracy was 63.9%, roughly matching a bookmaker-favorite baseline; models agreed more often than they were correct, while underpredicting draws and goals. The design would reduce answer memorization because predictions precede outcomes. However, the supplied metadata is dated August 2026, and the paper, archive, scoring code, and reported results cannot be independently verified from the provided material alone.
No heat snapshots are available in the last 24 hours.