Shanmu Jin, a postdoctoral researcher and neurosurgery resident at Peking Union Medical College Hospital in Beijing, has proved Crouzeix's Conjecture, a problem in matrix analysis that had remained open for more than two decades. The decisive argument did not come out of a mathematics department. It emerged from a 16-hour autonomous run of GPT-5.6 Sol inside ChatGPT Work mode, guided by a carefully engineered prompt that Jin designed himself.

An Unconventional Route to a Classic Problem

Jin's path to the conjecture was anything but linear. He studied geology as an undergraduate, later earned a medical degree, and built his mathematical knowledge through self-directed study — an effort initially motivated by research on transcranial ultrasound.

Crouzeix's Conjecture, formulated in 2004 by the French numerical analyst Michel Crouzeix, concerns the relationship between a matrix's spectral norm and the maximum of a polynomial over its numerical range — a compact convex region of the complex plane, also known as the field of values, that captures more of a matrix's behavior than its eigenvalues alone. The conjecture asserts that the constant 2 suffices in that bound, and 2 is known to be the smallest constant that could; Crouzeix's original paper established a finite constant of 11.08, and the sharpest published bound — 1 + √2, roughly 2.414, from Crouzeix and Palencia in 2017 — still sat above the target. Its resolution has direct applications in the analysis of matrix functions, iterative solvers such as GMRES — a standard workhorse for the large sparse linear systems that arise across scientific computing and engineering simulation — and rational Krylov methods.

Inside the Autonomous Session

Rather than prompting the model interactively, Jin ran an extended autonomous session using a prompt architecture adapted from one OpenAI had developed for an earlier mathematical result, the Cycle Double Cover conjecture. The setup denied the model internet access. It deployed competing subagents tasked with independently developing and stress-testing different proof strategies, with explicit instructions against premature convergence and a requirement to falsify candidate arguments only through concrete counterexamples. The model was directed to continue until a complete proof survived internal adversarial review.

The resulting preprint, posted on July 27, 2026, includes the full prompt, successive draft manuscripts, a Lean formalization — a machine-checkable rendering of the proof in the same proof-assistant technology mathematicians including Terence Tao have used to certify complex arguments — and an axiom audit.

Peer Confirmation and the Question of Who Does Mathematics Now

The proof has been reviewed and confirmed by Cornell mathematician Alex Townsend, numerical analyst Anne Greenbaum, and Michel Crouzeix himself — the conjecture's original author. Formal peer review is ongoing.

Eight days after the initial preprint appeared, mathematicians Emiel Lorist and Felix Schwenninger posted an independent five-page proof using a different method, combining classical double-layer potential theory with a perturbation lemma for 2-dilations. Jin welcomed the parallel result as a complement to his own work.

The episode makes visible a structural shift in mathematical research that is difficult to ignore. When a problem open since 2004 is resolved by a self-taught clinician using a consumer AI tool, the barriers to producing new results come down substantially — while the expertise required to verify those results becomes correspondingly more concentrated and more valuable. The result also extends a documented pattern of AI systems advancing into mathematics, following Google DeepMind's AlphaProof and AlphaGeometry 2, which reached silver-medal standard at the 2024 International Mathematical Olympiad.

For the AI field, the case offers a concrete demonstration of what frontier language models can accomplish when given well-structured autonomous tasks and sufficient compute time. Whether the mathematical community has the infrastructure to absorb the volume of results that may follow remains an open question.

Source: Metaverse Post