AI models jump from 2% to 35% accuracy on theoretical science problems, MIT study finds

The leap MIT measured

A study led by Neil Thompson, an MIT researcher who studies the progress of computing and artificial intelligence, measured how much OpenAI's models improved at solving theoretical problems outside of pure mathematics. Between mid-2024 and mid-2026, the accuracy rate on problems in probability, statistical theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics jumped from about 2% to about 35%, according to Wired's report.

Thompson presented the data in a talk at MIT and did not hold back on the warning tone.

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.

Neil Thompson, MIT researcher

The study was done in partnership with Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist, both from Thompson's lab at MIT.

How the test was designed to keep AI from cheating

The methodology is the most relevant point for anyone who works with models day to day: the researchers took problems published in scientific papers posted between December 2025 and May 2026, after the training cutoff of the models tested. In practice, this means no model could have seen the answer during training.

The models evaluated were o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5, all from OpenAI. Internet access was blocked during the tests, which prevented any model from simply looking up the solution published online. The original papers served as the ground truth to check whether the generated answer matched the actual solution.

The study's scope also matters: the researchers chose areas where the model could reach the answer through reasoning, without needing to run a simulation or physical experiment. It's a controlled universe, but it's exactly the kind of task that already shows up today in research pipelines with Python, notebooks, and agents that chain together calculation steps.

The crisis that had already erupted in pure mathematics

The context behind the study is the wave that hit pure mathematics. In September 2026, OpenAI announced that a yet-to-be-released model had solved a decades-old problem known as Navier-Stokes, about the equations that describe the behavior of fluids in motion. The episode sparked controversy after a mathematician at NYU, who was working with another academic affiliated with Anthropic on the same problem, accused OpenAI of moving in on their work as soon as it learned of the pair's progress.

Since then, OpenAI and other labs have announced hundreds of solutions to longstanding mathematical problems, which Wired describes as a kind of crisis in the field. Part of the mathematics community has called for limits on the use of AI in this kind of work, arguing that the process of reaching the answer is often as valuable as the answer itself.

One detail that went almost unnoticed: OpenAI itself withdrew several solutions from its set of results after publication, because they didn't hold up under closer checking.

Why this matters to programmers, not just theoretical physicists

Readers who only deal with a product stack might think gravitation and cosmology are far removed from their day to day. But the pattern described in the study is identical to what's already seen in code copilots and agents that write entire functions: the model produces a fluent answer, with the appearance of a correct solution, and the task left for the human is to audit it.

Manso, a co-author of the study, sums up the problem in a way that any dev who has reviewed an AI-generated pull request will recognize.

We are still going to need people to generate good ideas to connect the dots. We also need ways to verify what was done, because it's very easy for a model to tell you 'this is the right answer,' and when you look at the entire solution, you start to see maybe it is not.

Gabriel Manso, PhD student at MIT

For engineering and research teams here in Brazil who use Python as the standard language for scientific computing, in libraries like NumPy, SciPy and SymPy, the message is direct: generative AI tools can already propose proofs and calculations that used to require a senior theorist, but the verification layer (testing, peer review, independent reproduction) hasn't shrunk. It has just moved: from "checking whether the person calculated correctly" to "checking whether the model didn't invent a step."

What remains open

Thompson notes that theoretical advances sometimes drive practical progress in science, and that more AI use could mean more progress happening through this route. Several AI researchers have already founded labs dedicated to AI for science, and the major AI labs are expanding scientific research, including efforts to get models to run actual experiments, not just theoretical analysis.

The study by Thompson, Manso and Verdun covered a deliberately limited scope: areas where you can reach the answer just by reasoning, without simulation or a lab bench. It remains to be seen whether the same jump from 2% to 35% repeats in fields that depend on physical experiments, where the model can't just "think" its way to the answer.

For those building the tools that will run these checks, whether in a Jupyter notebook, a CI pipeline that runs formal proofs, or an agent that audits another agent's outputs, the next cycle of work isn't teaching the model to solve more problems. It's building the infrastructure that checks whether it solved them correctly.