AI models solved 35% of selected theory problems, up from 2% in two years

The best AI systems in an MIT study solved about 35% of selected theoretical research problems by mid-2026, compared with roughly 2% two years earlier. The tests covered probability, statistics, economics, high-energy theory, gravitation, cosmology and quantum physics.

Researchers Neil Thompson, Gabriel Manso and Claudio Mayrink Verdun drew problems from papers posted between December 2025 and May 2026. The published solutions came after the training cutoff of the tested models, and internet access was disabled. Those choices were intended to reduce the chance that a model could reproduce a memorized or retrieved answer.

The scope was deliberately limited to areas where a model could reason without running a physical experiment or simulation. That makes the result evidence of progress on theory-heavy tasks, not a measure of scientific work as a whole. Experts still have to decide which questions matter, connect a result with a broader field and verify long derivations.

Verification is a concrete concern. OpenAI has withdrawn several claimed mathematical solutions after publication, and a plausible final answer can hide errors in intermediate steps. The study suggests that AI may take over more routine theoretical calculation, but it does not show that models can independently choose valuable research directions or validate their own work.