Live·Open questions in longevity research
All news
Scientific ComputingScience Research

OpenAI’s GPT-6.2 Astra AI model produced a computer-verified proof of a 2006 theorem on algebra growth in 94 minutes

19 November 2026· 261119002

OpenAI’s GPT-6.2 Astra AI model produced a computer-verified proof of a 2006 theorem on algebra growth in 94 minutes

On November 18, Vals AI, which evaluates AI models, published an analysis of Astra’s run on the Autoform-Hard benchmark. Experts had encoded research-level theorem statements in Lean, and the model had to supply proofs whose every step the program would accept. Astra was the only model tested to prove mathematician Agata Smoktunowicz’s theorem, with no human intervention and at a cost of 21,58 dollars at list prices. Vals knows of no other Lean proof of this theorem, and Astra’s approach diverges from the published proof after the first few steps.

An algebra is a collection of objects, such as polynomials, that can be added and multiplied. To measure its growth, one starts with a few elements and counts the linearly independent products formed from at most n factors. For polynomials in two variables, this count grows as n²; for polynomials in three variables, it grows as n³. The exponent is called the Gelfand-Kirillov dimension. For algebras in general, it can take any value from 2 upward. However, the mathematicians Artin and Stafford conjectured, and Smoktunowicz proved in 2005 to 2006, that algebras resembling polynomials cannot have a dimension between 2 and 3. Three properties define this resemblance: elements decompose by degree, degree zero contains finitely many linearly independent elements, and the product of nonzero elements is nonzero.

Around minute 47, Astra submitted a proof with a gap in the general case. The benchmark rejected it, and the model reported that it had not solved the problem. The benchmark rules prohibit giving up before spending 20 dollars. Astra had spent about 8, so the benchmark sent it back to work. It spent nine minutes reasoning without writing code, then began a new approach. The benchmark accepted a complete proof 39 minutes after the report, and an independent program checked the proof again.

The published proofs rely on three deep results that are absent from Mathlib, Lean’s mathematics library. Astra proceeded without them, using a proof by contradiction: suppose the dimension lies between 2 and 3. The number of linearly independent elements of degree n alone must then grow faster than linearly but slower than quadratically. In place of those results, it introduced the capacity of a set of fractions formed from elements of the same degree. This is the smallest increase in the number of linearly independent elements under multiplication by a fixed set, taken over all sets containing the original set.

Capacity is calculated at successively higher degrees. Each step up in degree adds at least as many linearly independent elements as the capacity. If capacity increased at every step, this would yield 1 + 2 + 3 + … elements, giving quadratic growth, whereas the assumption requires slower growth. Capacity must therefore stop increasing at some step, reaching a plateau. At the plateau, adjoining a copy of the set multiplied by any such fraction leaves capacity unchanged. It follows that the number of linearly independent elements of degree n grows at most linearly, contradicting the assumption.

After comparing Astra’s approach with three published proofs, none of which uses capacity or a plateau, Vals writes:

“This is a formal proof of Smoktunowicz’s theorem. It does not formalize her proof: the two coincide only in their first steps.”

In 1998, mathematician James Zhang used the increase under multiplication by a fixed set to define a numerical invariant of algebras. According to Vals, he proposed it as a route to proving this theorem. Vals is asking readers familiar with the literature to explain how Astra’s approach relates to existing work. In August, mathematicians found ideas from recent papers in two of Astra’s most prominent results written in Lean.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#gpt-6.2-astra#automated-theorem-proving#lean#smoktunowicz-theorem#algebra-growth#formal-verification