LLM-SoccerArena: Benchmarking LLMs on Real-World Predictions in Sports

Researchers released LLM-SoccerArena, an open-source live benchmark for judging how well LLMs predict unresolved sports events before outcomes are known.

The platform records timestamped forecasts and compares performance across model version, information access, prompting strategy, and forecast horizon.

They tested seven models on all 104 matches and 15 questions for the 2026 FIFA World Cup, finding a small edge for models with web access.

The study positions sports as a practical setting for prospective real-world prediction benchmarks and suggests web access helps only modestly.