MIT study: AI theory problem-solving jumps from 2% to 35% in two years
Researchers at MIT — Neil Thompson, Gabriel Manso and Claudio Mayrink Verdun — tested OpenAI's o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5 on theoretical problems in probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. According to the study, the best models' share of correct solutions rose from about 2% to about 35% between mid-2024 and mid-2026. The problems came from papers posted between December 2025 and May 2026 whose solutions appeared after the training cutoff of GPT-5.5 and earlier models, and internet access was blocked during testing, so the models had to derive answers themselves. Thompson says many theory questions could be "totally transformed" within a few years, while the authors caution that humans are still needed to generate ideas and verify solutions; OpenAI withdrew several solutions from its latest dataset after publication.
If the trend holds, AI could reshape theory-heavy scientific fields much as it is already disrupting mathematics, making verification and human judgment more important.