AI Models Rapidly Improving at Solving Theoretical Physics and Mathematics Problems
OpenAI's latest models have jumped from solving 2% to 35% of theory problems in physics, statistics, and other fields between mid-2024 and mid-2026.
What happened
Researchers at MIT, led by Neil Thompson, tested OpenAI models (o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5) against theoretical problems from research papers published between December 2025 and May 2026—after the models' training cutoffs. The models were blocked from internet access and had to solve problems without relying on training data. Success rates surged from approximately 2% to 35% across fields including probability, statistics theory, economic theory, high-energy physics, gravitation, cosmology, and quantum physics. Thompson states there are no signs this progress will slow, with theoretical breakthroughs expected within years. The study follows OpenAI's announcement that an unreleased model solved the Navier-Stokes equations, a long-standing grand challenge in mathematics, which triggered controversy and a broader crisis in the math community.
Context
AI's proven ability to solve theoretical problems at scale raises questions about the future role of human theorists and the value of the problem-solving process itself. Some mathematicians have called for limits on AI use, arguing that working through problems teaches as much as the answer itself. However, researchers note that human expertise remains essential: to generate good ideas, verify solutions for correctness (OpenAI has already withdrawn several incorrect solutions), and connect theoretical advances to practical applications. The results suggest that shock waves currently felt in mathematics will likely extend to other theory-heavy scientific fields. Major AI labs are already establishing new research initiatives focused on AI for science, exploring how AI could conduct experiments alongside theoretical work, potentially transforming how scientific progress is made.