Recent months have seen several announcements in which the Large Language Models (LLM) behind modern Artificial Intelligence (AI) technologies have solved long-standing problems in the field. Among those previously unsolved problems are a handful originally conceived by Paul Erdős, who generated a catalog of 1,135 challenges for mathematicians to take on. Several AI developers target these problems as means of measuring the progress they're making in developing their systems.
But are today's AI systems really capable of solving these unsolved problems? LLMs have been described as an advanced form of an autocomplete function, or even as a "glorified autocorrect" program the AI systems use to statistically predict what response should follow the prompts it has been given based on the reams of data on which it has been trained.
In the case of the Erdős problems that have been solved by AI, there could be something to that argument. The solved problems have been described as being relatively low-hanging fruit, whose unsolved status may have more to with their relative obscurity. Being similar to other Erdős problems that have been solved, which would be part of the training library used by the LLMs, that similarity could be enough to solve them.
The alternative hypothesis is the math-LLMs are genuinely capable of coming up with original solutions for these problems. But how can we tell which hypothesis is closer to the truth?
A group of eleven mathematicians has proposed a interesting experiment to find out. They're tapping their currently unpublished research to remove the possibility that the LLM is effectively rehashing mathematical solution processes to which they have previously been exposed. Here's the abstract for their preprint paper, which they uploaded on 6 February 2026:
To assess the ability of current AI systems to correctly answer research-level mathematics questions, we share a set of ten math questions which have arisen naturally in the research process of the authors. The questions had not been shared publicly until now; the answers are known to the authors of the questions but will remain encrypted for a short time.
"Short time" was one week. The ten questions were unencrypted on 13 February 2026.
That action started a clock for challenging today's math-AI systems to see if they're genuinely capable of autonomously solving mathematical research questions. At this writing, we don't know what results, if any, have been put forward and whether they stands up to scrutiny.
Regardless of how it goes, it's a genuinely exciting research effort for which we're looking forward to learning the outcome.
Image credit: Artificial Intelligence - hand touches icon with symbol of network and brain by Gerd Altmann on PublicDomainPictures.com. Creative Commons Creative Commons - CC0 Public Domain.