8 Lean Formalizations in OpenAI's Repo 'Broken' — Wrong Proofs Pass Verification
AlexTensor · x · 2026-10-09
@Timeroot flagged that 8 Lean formalizations in OpenAI's repo are "broken" — trivially provable with wrong proofs. As @KyleCranmer clarifies, the claim isn't that the proofs are wrong, but that verification is hackable: definitions were whitelisted and could be overwritten, letting flawed proofs still pass formal verification. A caution for LLM-driven formal math workflows.
Related event: Eight OpenAI Lean Formalizations Found Trivially Provable with False Proofs(3 posts)→
More from Research
- SubDGuide: modeler-inspired agentic workflow for mesh-to-SubD reconstruction — ssh4net · 2026-10-10
- Simons Institute launches Fall '27 semester program on diffusions and flows, postdoc fellowships open — giannis_daras · 2026-10-10
- MIT paper: AGI's binding constraint is human verification bandwidth, not intelligence — Afinetheorem · 2026-10-10
- Ex-OpenAI Operator Researcher Previews Real-World RL Gains on High-Precision Tasks — ZhaoMandi · 2026-10-10
- Stanford's Mulligan targets failure states for on-robot learning, beats uniform collection across 2,550 blind evals — ZhaoMandi · 2026-10-10
- RCT: AI as tutor raises test scores that persist; letting AI write for you fades within a week — emollick · 2026-10-10