Skip to content
AI.info

The Pulse

Vals AI says GPT-6.2 Astra produced Lean proof of algebra gap theorem

GPT-6.2 Astra produced a Lean proof of a 2006 theorem about the growth of certain algebras, according to Vals AI. The benchmark operator says the proof took a different route from the published argument and passed its automated checks.

Vals AI says GPT-6.2 Astra produced Lean proof of algebra gap theorem

Vals AI says GPT-6.2 Astra produced a machine-checkable proof of a 2006 theorem that rules out a range of growth rates for certain algebras. The benchmark operator says the 94-minute proof took a different route from the published argument and passed the checks for its Autoform-Hard benchmark.

The theorem, proved by Agata Smoktunowicz, concerns a measure called Gelfand–Kirillov dimension. Vals says it had not previously encountered a formal proof of the result.

The mathematical problem

The theorem concerns finitely generated algebras over a field that are graded by the natural numbers, have no zero divisors and have a finite-dimensional degree-zero part. For those algebras, the dimension is either at most 2 or at least 3, with no value strictly between them.

Gelfand–Kirillov dimension describes how an algebra grows as products of its generators are formed. Vals says Astra argued by contradiction: it assumed a dimension in the forbidden range, then showed that the resulting growth would have to be slower than that assumption allowed.

A benchmark run resumed

The proof was the distinguishing result among the leading models on Autoform-Hard, according to Vals. The company says GPT-6.2 Sol and Claude Fable 5.2 solved the same 48 of the benchmark’s 56 research-level formalization tasks, while Astra also solved the Smoktunowicz task.

Astra worked without human intervention in a sandbox with Lean, the Mathlib library and a shell, Vals says. The benchmark imposed a three-hour limit and a $200 budget; the company reports that the run cost $21.58 at Astra’s published price and produced 1,794 lines of code over 146 agent turns.

Around 47 minutes into the run, the model submitted a partial proof that covered the easier commutative case but left a gap in the general case. The grader rejected it, and Astra reported the task unsolved. The benchmark sent it back to work because it had spent about $8 and the rules did not allow a model to give up before spending $20.

Vals says the model then spent nine minutes thinking without writing code before starting a new approach. The grader accepted the full proof 39 minutes after the unsolved report.

Checking the proof

Vals says the benchmark rebuilds the theorem from a trusted task copy, checks that its statement matches the original and permits only Lean’s standard axioms. An independent kernel then replays the proof.

The company says its manual review found that the proof built without gaps or warnings and used only those standard axioms. It also checked that the statement matched the theorem’s hypotheses, including the finite-dimensional degree-zero part, and that the proof did not introduce new axioms or redefine benchmark objects.

A different route

The two proofs share their opening steps, but Astra’s later argument differs from Smoktunowicz’s published proof, according to Vals. The company says the published argument draws on three results not present in Mathlib, none of which Astra used.

Instead, Astra defined a quantity it called capacity to measure growth when a space is multiplied by a fixed generating piece. Vals traces related ideas to earlier research but says the minimization and plateau argument used here are not present in the cited work.

Vals published an annotated version of the proof with commentary on the code, the mathematical argument and the run. Its account of the model’s thinking is reconstructed from the run log; quoted reasoning in the annotation comes from provider-generated summaries rather than raw reasoning.

Source

  1. Vals AI

Related stories

Explore

More articles