GPT Solves Years-Old Gradient Descent Math Proof Baffling Researchers

konstmish · x · 2026-08-03

ML researcher Konstantin Mishchenko shared how an LLM helped crack a theoretical math problem that had bothered him since 2019. In their paper on adaptive gradient descent, the authors couldn't determine if a constant factor in their algorithm could be tightened to exactly 1.

After prompting GPT, the model concluded that the method would not converge even with a constant factor of 0.98 and constructed a PEP-style (Performance Estimation Problem) proof. This highlights the potential of frontier LLMs in handling complex, highly specialized mathematical reasoning.

Related event: Researcher Uses GPT to Solve Long-Standing Gradient Descent Proof(2 posts)→

Original post →

More from Fun

Fun channel →