, , , ,

Google's Gemini 4 Multi-Agent System Proves a 2013 Math Conjecture in Lean 4

Close-up detail of a mathematician's desk with handwritten mathematical proofs on crumpled paper, chalk dust on the desk surface, a worn mec

Google says a multi-agent system built on its Gemini 4 family has fully proved the 2013 Courtade–Kumar conjecture, an open problem in information theory that had resisted a decade of attempts, and formalized the proof in roughly 50 million lines of Lean 4 code.

The result, described by Google Research as the first complete resolution of a longstanding mathematical conjecture by an AI system, is less about the specific theorem and more about what it portends. The benchmark was a collaboration across several specialized agents, each assigned a discrete sub-task of the proof, coordinated by a planning layer that broke a sprawling, opaque argument into verifiable steps. Google says the approach reflects a shift away from single monolithic models toward orchestrated systems that can sustain long chains of reasoning too error-prone for humans to check by hand.

The theorem, named for the information theorists Abbas El Gamal and Young-Han Kim, concerns the entropy of Boolean functions and has applications in coding theory. It was first posed publicly in 2013, and progress had been limited. Mathematicians not involved in the project cautioned that a formal proof in Lean is not the same as a proof understood by the community, but they acknowledged the verification is exact and machine-checkable.

For the AI industry, the advance lands at a moment when labs are competing on reasoning benchmarks and on the rigor of their outputs. Google's demonstration suggests the next frontier is not merely generating plausible text, but producing results that can be mechanically certified as correct. The practical implications extend beyond mathematics: the same verification discipline could apply to safety proofs for code, contracts, and autonomous systems, a frame the company is already pressing in its enterprise offerings.

Image source: i.ibb.co