Study finds AI models solve about 35 percent of tested theory problems

A study led by MIT researcher Neil Thompson found that the best tested AI models improved from solving about 2 percent of theoretical problems to about 35 percent between mid-2024 and mid-2026, WIRED reported. The fields included probability, economic theory and quantum physics. Researchers tested OpenAI models against problems whose published solutions came after the models' training cutoffs, and blocked internet access. The evaluation covered fields requiring neither simulations nor experiments. Researchers cautioned that experts remain necessary to interpret and verify solutions.