Artificial intelligence may soon reshape theoretical science much as it has begun transforming mathematics. Wired.com reports on research led by MIT’s Neil Thompson showing rapid gains in frontier AI models across probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics.
Thompson, working with Gabriel Manso and Claudio Mayrink Verdun, tested several OpenAI models on theoretical problems drawn from research papers published between December 2025 and May 2026. The problems were selected so their solutions appeared after the models’ training cutoffs, reducing the possibility that AI systems were simply recalling known answers. Internet access was also blocked during testing.
The results suggest a sharp acceleration in AI’s problem-solving ability. Between mid-2024 and mid-2026, the best-performing models improved from solving roughly 2% of theoretical problems correctly to about 35%. The researchers expect AI capabilities in theory-heavy scientific fields to continue advancing rapidly.
The study focused on disciplines where problems can be addressed without physical experiments or computer simulations. Still, the findings suggest that disruption already visible in mathematics could spread into other scientific fields. Recent AI claims about difficult mathematical problems have generated excitement and concern about verification, research practices, and the changing role of human experts.
AI could accelerate theoretical discoveries that eventually lead to practical scientific advances. Research organizations and major AI companies are also exploring systems that combine theoretical reasoning with real-world experimentation.
Human scientists, however, remain essential. Researchers must generate worthwhile questions, connect findings, interpret results, and verify AI-generated solutions. Complex reasoning remains vulnerable to errors, even when a model presents its answer confidently. The emerging scientific workflow may therefore depend less on researchers solving every theoretical problem themselves and more on their ability to guide, evaluate, and validate increasingly capable AI systems before those results can influence scientific understanding.
