GoKawiilMIT study finds AI models solving far more theoretical science problems than in 2024
Neil Thompson's MIT lab tested OpenAI models including o1 through GPT-5.5 on unpublished theoretical problems drawn from recent papers in fields like probability, statistics theory, economic theory, high-energy theory, cosmology and quantum physics. Success rates rose from about 2 percent in mid-2024 to roughly 35 percent by mid-2026, with internet access blocked to ensure models reasoned rather than searched for answers. The researchers, including PhD student Gabriel Manso and scientist Claudio Mayrink Verdun, say the trend shows no sign of plateauing.
GoKawiil's interpretation of the reporting above, not reported fact.
The findings suggest theoretical work across several scientific disciplines could become increasingly automatable, following a pattern already seen in mathematics. Thompson argues this could transform how researchers approach theory-heavy problems within just a few years, though the study's scope was limited to fields solvable without simulation or experimentation, so broader scientific impact remains uncertain.
- AI models' success rate on theoretical science problems jumped from 2% to 35% in two years, per MIT research
- Tests used unpublished problems from recent papers with internet access blocked to prevent lookup
- Fields like probability, economic theory, cosmology and quantum physics could follow math's path toward AI automation
Source: wired.com — Will Knight, 2026-10-14
Published there as: “AI Came for Math. Now It’s Chasing the Rest of Science”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.