AI models show rapid gains in solving theoretical problems across multiple scientific fields
Միայն փաստեր

AI models show rapid gains in solving theoretical problems across multiple scientific fields

Ամփոփում

A MIT study finds that advanced AI systems increased their success rate on theoretical questions in economics, physics and related disciplines from about 2% to 35% between mid-2024 and mid-2026, suggesting broader impacts beyond mathematics.

Researchers at the Massachusetts Institute of Technology reported that frontier AI models have markedly improved their ability to solve theoretical problems in fields such as probability, statistics, economic theory, high-energy physics, gravitation, cosmology and quantum physics. The study examined problems published in research papers between December 2025 and May 2026 and tested a series of OpenAI models, including versions of GPT-5, on solving them without internet access or reliance on post-cutoff training data. Results show that the proportion of correctly solved problems rose from roughly 2% to about 35% over a two-year span, with no indication of slowing progress.

"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," said Neil Thompson, a computer-science researcher at MIT.

The researchers emphasized that the test set was limited to problems that could be addressed without running simulations or experiments, yet they argue the trend mirrors the recent disruption AI has caused in mathematics. Earlier in the year, OpenAI announced that an unreleased model solved the Navier-Stokes problem, prompting debate within the mathematical community. Subsequent AI-generated solutions to longstanding math challenges have sparked discussions about the role of AI in research.

Thompson and his collaborators caution that while AI may automate certain aspects of theoretical work, human expertise will remain essential for generating ideas, verifying results and interpreting findings. "We are still going to need people to generate good ideas to connect the dots," said Gabriel Manso, a PhD student involved in the study. He added that AI can produce errors, underscoring the need for robust validation methods.

The findings suggest that advances in AI-driven theory could accelerate scientific progress, but also raise questions about how researchers will integrate and oversee AI contributions across disciplines.

Աղբյուր

WIRED

Առաջին անգամ հաղորդվել է այստեղ

Շարունակեք կարդալ — անվճար

Ստեղծեք անվճար հաշիվ՝ նորություններին հետևելու համար։ Քարտ չի պահանջվում։