HorizonMath: 100 Unsolved Math Problems Benchmark Where GPT-5.4 Pro Scores 7%

arXiv / Import AI 451·medium signal

Oxford, Harvard, Princeton, and Ellison Institute published HorizonMath (arXiv 2603.15617), a benchmark of 100 genuinely unsolved mathematical research problems across 8 domains with automatic verification computing ground-truth values to 100+ digit precision. GPT-5.4 Pro achieved 7% overall, with a 50% baseline only on the easiest calibration tier. Unlike FrontierMath's closed problems, HorizonMath focuses on problems that would constitute actual mathematical discoveries if solved, with contamination-proof automated evaluation.