GPT-5.6 Sol Refutes 30-Year-Old Unsolved Statistical Conjecture in 90 Minutes

A statistics professor at the University of Pennsylvania used OpenAI's latest model "GPT-5.6 Sol Pro" to refute a long-standing unsolved conjecture in statistics in approximately 90 minutes. The conjecture pertains to the "Benjamini-Hochberg procedure," a technique for multiple hypothesis testing that has remained unresolved in the mathematics and statistics community for nearly 30 years.

The Benjamini-Hochberg procedure is a statistical method used to suppress errors when simultaneously testing multiple hypotheses, and it is widely employed in scientific research. The conjecture that was refuted this time concerned a central property of this method. Despite decades of efforts by numerous researchers, it had remained an "unsolved problem" without proof or refutation.

Notably, the previous version, "GPT-5.5," failed to produce an answer despite spending more than 20 hours on the same problem. GPT-5.6 Sol Pro overcame this constraint in just 90 minutes. However, the derived solution was not "an entirely new mathematical discovery" but rather a combination of existing methods.

What this result demonstrates is that a change in model generation can significantly alter the quality of reasoning. The contrast between a problem remaining unsolved after 20 hours of effort and being resolved in 90 minutes is not merely a difference in speed; it suggests that the approach to the problem itself may have fundamentally changed. At the same time, the fact that the solution amounts to a "combination of existing methods" raises important questions.

That question is: "Can artificial intelligence truly generate new knowledge, or is it merely reconfiguring information it has already learned?" In this case, the result is reported to fall within the scope of existing knowledge. While it has value in that the AI discovered a combination that human researchers overlooked, whether this constitutes "creation" that truly transcends human thinking remains open to debate.

On the other hand, the fact that AI contributed to solving a problem that remained unsolved for decades reaffirms the potential of AI applications in mathematics, statistics, and fundamental science. One more case has been added to the question of how far AI can function in domains requiring specialized reasoning, beyond simple document creation or information retrieval. Going forward, the research community will likely focus on whether this kind of application can be reproducibly applied to other unsolved problems in a manner that transcends the realm of "combining existing knowledge."