MIT's Neil Thompson: AI theory-solving jumps from 2 percent to 35 percent

The study's finding that top AI models jumped from correctly solving about 2 percent of theoretical problems to about 35 percent in two years builds on a broader pattern: AI has already disrupted mathematics, with OpenAI last month announcing that an unreleased model solved the Navier-Stokes problem, a long-standing grand challenge. That episode proved controversial after a mathematician at NYU accused OpenAI of gazumping them, and OpenAI and others have since announced hundreds more solutions to longstanding math problems, triggering a crisis in the field.

The MIT study's methodology was deliberately strict. It took problems stated in research papers posted online between December 2025 and May 2026, and tested OpenAI models — o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5 — on them. Because solutions to these problems were found in papers published after the training cutoff for GPT-5.5 and preceding models, the models had to work out solutions for themselves rather than relying on answers fed in as training data. The researchers also blocked internet access. The study focused on fields where AI could figure things out without running simulations or experiments, so the results may not extend to areas requiring real-world data.

Thompson says we should expect the advances to have both positive and negative effects, noting that theoretical advances sometimes drive practical advances in science and suggesting greater use of AI may mean more progress happens that way. Some prominent AI researchers have already founded new labs focused on AI for science, and the big AI labs are all ramping up scientific research, exploring how AI could be empowered to conduct real experiments alongside theoretical analysis. AI may make mistakes: OpenAI, in fact, withdrew several solutions in its latest dataset post-publication, and Manso said experts will still be needed to generate good ideas and connect the dots, as well as verify what was done.