Study Finds AI Solves 35% of Theory Problems, Up From 2% in 2 Years
Updated
Updated · WIRED · Oct 14
Study Finds AI Solves 35% of Theory Problems, Up From 2% in 2 Years
2 articles · Updated · WIRED · Oct 14
Summary
35% of theoretical science problems were solved by the best AI models by mid-2026, up from about 2% in mid-2024, according to a Neil Thompson-led study.
The researchers tested OpenAI models from o1 through GPT-5.5 on problems drawn from papers posted between December 2025 and May 2026, then checked answers against later-published solutions.
Internet access was blocked and the benchmark used problems whose solutions fell after the models' training cutoffs, aiming to show the systems reasoned them out rather than retrieved memorized answers.
The study focused on theory-heavy fields including probability, statistics, economics, high-energy physics, gravitation, cosmology and quantum physics, suggesting math's recent AI disruption could spread across adjacent disciplines.
Researchers said experts will still be needed to generate ideas and verify outputs, especially after OpenAI recently withdrew several claimed math solutions over accuracy concerns.