AI news ·
AI solves six-year-old math research problem for the first time
An AI solved a six-year-old unsolved math problem from the FrontierMath benchmark, originally posed in a 2019 paper. The problem's author plans to publish the solution in a journal, with the AI users listed as co-authors.
AI System Solves Math Problem Researchers Couldn't Crack for Six Years
An AI system has solved a problem from FrontierMath: Open Problems, a benchmark of unsolved research questions that mathematicians have attempted without success. The problem originated from a 2019 paper by mathematicians Will Brian and Paul Larson. Multiple attempts to solve it over the past six years failed.
Researchers Kevin Barreto and Liam Price first extracted a solution using GPT-5.4 Pro. Geby Jaff subsequently elicited a solution as well. Brian categorized the problem as "Moderately Interesting" within the benchmark framework.
Several AI models have now demonstrated the ability to solve the problem, including GPT-5.4 (xhigh), Gemini 3.1 Pro, and Claude Opus 4.6 (max). Epoch AI, which operates the FrontierMath benchmark, replicated the results in its own testing environment.
Publication and Follow-Up Work
Brian plans to publish the solution in a specialist journal. Barreto and Price will be offered co-authorship on the resulting paper alongside Brian.
Brian assessed that the solution will likely generate new research questions. Any follow-on work sparked by the AI's approach may be included in the publication.
The full chat transcript from GPT-5.4 Pro's original solution is available on the FrontierMath website, along with solutions from other models.
For professionals working in research and mathematics, this development demonstrates how current Generative AI and LLM Courses can address open problems. Those interested in understanding how these systems approach research challenges may also benefit from AI Research Courses.