Do OpenAI's math-proof agents call a Lean verifier at inference time?

yoavgo · x · 2026-09-27

Yoav Goldberg raises an open technical question: do the math-proof agents at OpenAI and similar labs run inside a harness that calls a Lean verifier every few steps, or is Lean used mainly in training with inference relying on learned behavior? He assumes the former, but it's not publicly known — with implications for whether formal verification serves as a runtime guarantee or just a training signal.

Original post →

More from coding & agent

coding & agent channel →