Artificial intelligence is showing growing ability to solve research problems beyond mathematics. A study led by Neil Thompson found that leading models became substantially better at answering theoretical questions in several fields, raising the prospect of changes to how scientists work.
A rapid change in theoretical problem-solving
The study examined probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. Between mid-2024 and mid-2026, the best models went from correctly solving about 2 percent of the theoretical problems tested to about 35 percent.
Thompson, who studies how advances in computing and artificial intelligence affect society and innovation, presented the findings in a talk at MIT. He said that within just a few years, many theory questions could be transformed by AI’s ability to tackle them.
The work was conducted with Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist. The team selected problems stated in research papers posted online between December 2025 and May 2026, then evaluated several OpenAI models: o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5.
How the researchers tested the models
The researchers designed the test to check whether models could work through problems themselves. Solutions appeared in papers published after the training cutoff for GPT-5.5 and preceding models, so those answers were not available as training material.
They also blocked internet access during testing. That prevented the models from searching online for a solution. The papers containing the solutions served as ground truth against which the researchers could assess the models’ answers.
The study had a defined boundary: it focused on areas where models could reason about problems without running simulations or conducting experiments. Its results therefore speak to theoretical work, rather than showing that AI can already carry out scientific research in every form.
Math offers a preview of the stakes
The findings arrive amid a period of disruption in mathematics. OpenAI recently announced that an unreleased model had solved Navier-Stokes, a longstanding challenge concerning equations that describe fluid motion. The announcement became controversial after a mathematician at NYU, who had been working with another academic associated with Anthropic, accused OpenAI of gazumping them after learning of their progress.
OpenAI and others have since announced hundreds more solutions to longstanding math problems. Some mathematicians have called for limits on AI use, arguing that the process of solving a problem can be as valuable as reaching the final answer.
That concern could matter beyond mathematics if AI takes on more theoretical questions in other sciences. The study suggests that similar shifts may reach fields that rely heavily on mathematics and theory. Thompson also sees a possible upside: theoretical advances can lead to practical advances in science, and AI could help more progress happen through that route.
Human judgment remains part of the work
AI labs are increasing their scientific research, and some prominent AI researchers have founded labs focused on AI for science. These efforts look at ways to use AI in real experiments as well as theoretical analysis.
Even if models can solve important problems, scientists will still be needed to judge what the answers mean and whether they hold up. Manso said people are needed to generate ideas and connect findings. He also cautioned that AI can make mistakes on complex problems.
Verification is a practical challenge. OpenAI withdrew several solutions in its latest dataset after publication, illustrating why a model’s confident answer cannot be treated as proof on its own. Researchers will need ways to examine the reasoning and confirm that the solution is sound.
The study points to a changing research landscape, not a finished handoff from scientists to machines. AI may broaden the range of problems researchers can explore, while human expertise remains essential for choosing worthwhile questions, interpreting results, and checking the work.