An MIT study found that leading AI models’ success rate on theoretical problems rose from about 2% in mid-2024 to 35% in mid-2026. The models were tested on recent questions in probability, statistics, economics, high-energy physics, gravitation, cosmology and quantum physics.