A new study presented at MIT last week shows that leading AI models improved their ability to correctly answer theoretical problems in several scientific disciplines from roughly 2 percent in mid-2024 to approximately 35 percent by mid-2026. Researcher Neil Thompson, who studies computing and AI progress, conducted the work alongside Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist. The findings suggest that fields including probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics may soon face the same upheaval currently shaking mathematics.
The researchers tested six OpenAI models—o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—against problems drawn from research papers posted online between December 2025 and May 2026. Importantly, answers to these problems appeared in papers published after the training cutoff for GPT-5.5 and earlier models, meaning the AI systems had to determine solutions independently rather than recall answers from training data. The team also blocked internet access to prevent the models from searching for solutions online. The study's scope was deliberately limited to disciplines where AI could reason through problems without running simulations or conducting experiments.
"Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI's ability to do them," Thompson says. The work arrives as mathematicians grapple with a crisis triggered by OpenAI's September announcement that an unreleased model had solved the Navier-Stokes problem, a longstanding grand challenge involving equations that describe fluid motion behavior. OpenAI and competitors have since announced hundreds more solutions to persistent math problems, prompting some mathematicians to call for restrictions on AI use and warn that the problem-solving process itself holds value beyond the final answer.
Thompson notes that theoretical breakthroughs sometimes drive practical scientific progress, suggesting that expanded AI use might accelerate this pathway. Prominent AI researchers have already launched new labs dedicated to AI for science, while major AI companies are scaling up scientific research efforts that explore how AI could perform actual experiments alongside theoretical work. The researchers caution, however, that experts will still be essential for evaluating solutions and ensuring their validity. "We are still going to need people to generate good ideas to connect the dots," Manso says, adding that for highly complex problems, AI might make errors—OpenAI itself retracted several solutions from its latest dataset after publication.
The report emphasizes that advances in these areas will carry both beneficial and harmful effects, much like those currently rippling through mathematics. According to the researchers, verification mechanisms will prove critical because models can easily claim an answer is correct even when the complete solution reveals flaws. The outlook suggests that while AI may eliminate the need for theorists to solve certain key problems, human judgment remains indispensable for generating strong ideas and connecting insights across disciplines. The scientific method's traditional division of labor appears poised for fundamental restructuring as machines claim territory once reserved exclusively for human reasoning. Organizations that treat AI as a wholesale replacement for domain expertise rather than a tool requiring expert oversight may find themselves amplifying errors at the same accelerated pace they hoped to achieve breakthroughs.