MIT 研究：前沿模型在多个理论科学领域的解题率两年内从 2% 升至 35%

Last week, I attended a talk at MIT by Neil Thompson, who studies progress in computing and artificial intelligence. Over the course of just half an hour, Thompson made a convincing case that several scientific fields are about to be upended by AI.

As mathematicians reel from the way AI is automating their field, Thompson’s latest study suggests that probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics may be next, thanks to the rapidly evolving capabilities of frontier models. Between mid-2024 and mid-2026, the study shows, the best AI models went from being able to correctly solve about 2 percent of theoretical problems in these fields to about 35 percent of such problems. And there are no signs that this progress will slow down.

“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,” says Thompson.

Broadly, Thompson’s lab digs into how advances in computer science impact society and affect innovation. This study—which Thompson conducted alongside Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist—took problems stated in research papers posted online between December 2025 and May 2026, and then tested whether different OpenAI models—o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—could solve them.

Crucially, solutions to these problems were found in papers published after the training cutoff for GPT-5.5 and preceding models. This means that the models they tested had to work out solutions for themselves, rather than relying on answers fed in as training data. (Such papers served as ground truth for testing the models’ solving skills.) The researchers also blocked internet access, which prevented the models they tested from looking up solutions online.

The areas covered by the study were limited—the trio focused on fields in which AI models could figure things out without needing to run simulations or experiments. But the work suggests that the kinds of shock waves currently being felt in math are likely to ripple out to other math- and theory-heavy areas of scientific inquiry.

Last month, OpenAI rocked the math world by announcing that an unreleased model had solved a long-standing, grand challenge in mathematics known as Navier-Stokes, which concerns equations that explain the behavior of fluid motion. The episode proved extra 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, triggering something of a crisis in the field. Some mathematicians have called for limits on the use of AI, warning that the process of solving problems is often as useful as having the final answer.

As with math, Thompson says we should expect advances in other areas to have both positive and negative effects. He notes that theoretical advances sometimes drive practical advances in science and suggests that greater use of AI may mean that more progress happens this way.

Some prominent AI researchers have already founded new labs focused on AI for science, and the big AI labs are all ramping up scientific research. These efforts explore how AI could be empowered to conduct real experiments alongside theoretical analysis. AI may mean that theorists are no longer required to solve key problems, but the researchers say that doesn’t mean that experts won’t be needed to make sense of solutions and to make sure they’re sound.

“We are still going to need people to generate good ideas to connect the dots,” says Manso. He adds that for very complex problems, it’s possible that AI might make mistakes. (OpenAI, in fact, withdrew several solutions in its latest dataset post-publication.) “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.”

Introducing the app: More ways to make the most of WIRED; Meta failed to catch hundreds of AI child abuse ads; Big Story: The search for Silicon Valley’s most powerful woman; Hackers got inside a Flock camera—its data shows how the system really works; In your inbox: Go inside the world of digital security with Kernel Panic