GPT-5.6 Solves Statistical Conjecture

soumitrashukla9 · x · 2026-07-15

This post claims that GPT-5.6 has solved a long-standing conjecture in the author's field of statistics. The author notes that it was one of the most interesting open problems they had seen, and they would have been thrilled to see a human peer construct this counterexample.

The quoted text provides background: this relates to FDR (False Discovery Rate) control in multiple hypothesis testing. Benjamini and Hochberg introduced the classic BH procedure in 1995, proving it controls FDR under independence; however, real-world test data often exhibits correlation, making this a significant ongoing challenge.

The focus here isn't the statistical method itself, but rather the fact that a large language model delivered a substantively valuable result on a complex mathematical/statistical problem.

Related event: GPT-5.6 Reportedly Found a BH Counterexample(8 posts)→

Original post →

More from Models

Models channel →