OpenAI's most powerful o3 model has been exposed for cheating, gaining privileged access to the FrontierMath test question bank in advance.

Golden Finance reported that a contractor named ‘Meemi’ from Epoch AI revealed on the LessWrong forum that OpenAI not only provided financial support for the FrontierMath Benchmark test, but also obtained privileged access to the test question bank. This may be an important reason for the significant improvement in the performance of o3 in a short period of time. This indicates that o3 has tremendous access to advanced mathematical reasoning, as claimed by Carina Hong, a mathematics PhD student at Stanford University. Under the arrangement of Epoch AI, OpenAI has privileged access to FrontierMath. Despite the progress, it faced a reputation reversal after the contractor’s disclosure. Faced with controversy, Tamay Besiroglu, the Vice Director and one of the co-founders of Epoch AI, quickly admitted to this on the X platform. FrontierMath is a heavyweight benchmark for advanced mathematical reasoning abilities. It was jointly developed by Epoch AI and more than 60 top mathematicians, including multiple Fields Medal winners and senior problem setters for the International Mathematical Olympiad.