AI Came for Math. Now It’s Chasing the Rest of Science

Summary

According to MIT researchers, the accuracy of frontier AI models on theoretical problems in physics, economics, statistics, and related fields rose from approximately 2% to 35% between mid-2024 and mid-2026. This indicates that AI's impact in mathematics is poised to spill over into other theory-heavy scientific disciplines.

Why it matters

If AI becomes the default tool for solving theoretical problems in high-energy physics, cosmology, quantum mechanics, and economics, the entire scientific discovery process will change. It is not just about who finds the answers; the way ideas are generated, validated, and connected will also transform. An improvement of 17% in less than two years with no visible ceiling is compressing the adaptation timeline for institutions, funding structures, and graduate education processes in these fields.

Highlights

• Accuracy on theoretical problems jumped from 2% to 35% in less than two years, with no signs of slowing down
• Models were tested on problems published after their training cutoff dates, with internet access disabled to prevent search-based solutions
• OpenAI retracted some solutions post-publication, indicating a genuine validation gap

Critical take

The study is limited to purely theoretical, non-experimental problems and a single provider's (OpenAI) model family. Therefore, generalizing the findings to 'the rest of science' is difficult. Moreover, a 35% accuracy rate still implies a 65% error margin. The article itself acknowledges that models confidently present incorrect solutions. In fields where a single wrong derivation can mislead years of research, this poses a serious risk.