Scientists developed and applied a set of fifty unpublished research problems related to the field of theoretical physics that have known solutions. The goal of the test was to determine the ability of artificial intelligence agents to solve these problems independently, without human intervention or assistance.
During the testing, it was found that Codex, a GPT-6 Astra model, was the best agent. This agent demonstrated high performance, correctly solving 49 out of the presented problems.
The results of this study have been published as a preprint, which can be found on the arXiv.org platform.