AI success on selected science theory problems rose from 2% to 35%, MIT study finds

AI models are solving a growing share of selected scientific theory problems without looking up the answers. A study led by MIT researcher Neil Thompson puts the best models’ success rate at about 35% in mid-2026, up from about 2% in mid-2024, according to WIRED’s October 14 report. The gains span six fields.

The finding concerns a specific kind of scientific work: questions that models could tackle without running simulations or conducting experiments. It does not measure the whole research process. Thompson and his collaborators instead examined whether successive OpenAI models could work out solutions to problems drawn from recent research papers.

Recent papers, with the answers kept out of reach

Thompson conducted the study with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun. They selected problems stated in papers posted online between December 2025 and May 2026. Published solutions then served as the reference answers against which the models’ work could be judged.

The timing of those solutions was central to the design. They appeared in papers published after the tested models’ training cutoffs, reducing the risk that a model was simply reproducing an answer it had already encountered during training. The researchers also blocked internet access, preventing an online search for the published solution.

The model lineup comprised o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5. That gives the study a comparison across successive OpenAI systems, rather than a contest among different companies’ models. Its headline result describes how the best models’ performance changed between mid-2024 and mid-2026.

Six fields, one deliberate boundary

The researchers’ selection reaches beyond mathematics into several theory-heavy areas of science. The six fields were:

• Probability
• Statistics theory
• Economic theory
• High-energy theory
• Gravitation and cosmology
• Quantum physics

That boundary matters when interpreting the result. The study tests whether models can solve selected problems within those fields, not whether they can independently carry out every kind of research those disciplines require. Work dependent on a simulation or experiment falls outside the task selection described by the researchers.

Thompson nevertheless expects substantial changes to theoretical research. “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,” he told WIRED. That is his forecast for the field, rather than an outcome measured by the study.

Solving a problem leaves other work to do

Thompson points to a possible route from these gains to practical science: theoretical advances sometimes drive practical advances. He suggests that greater use of AI could make that route more common. His expectation is not that every generated solution immediately becomes a useful scientific application.

The researchers also distinguish solving problems from deciding what their solutions mean. They say experts will still be needed to interpret results and establish whether they are sound. “We are still going to need people to generate good ideas to connect the dots,” Manso told WIRED.

Manso warns that complex problems can still produce mistakes, even when a model presents its answer confidently. His concern is about the full argument, not just the stated conclusion: a solution that initially looks right may reveal errors under scrutiny. Faster answer generation therefore leaves a separate job of checking the reasoning.