Faults in Our Formal Benchmarking: Dataset Defects and Evaluation Failures in Lean Theorem Proving
arXiv:2606.29493v1 Announce Type: new Abstract: Benchmarks for LLM-assisted theorem proving in Lean are often treated as intrinsically reliable because every solved instance comes with a machine-checked proof. However, the kernel only checks that a proof establishes a \emph{formal} statement; it does not verify that the statement faithfully encodes the intended informal problem, nor that evaluation harnesses are robust to trivial or adversarial solutions. We audit five widely used Lean theorem-proving benchmarks and their forks, using corpus-scale static checkers to surface 4,833 findings, inc
The proliferation and increasing reliance on large language models in formal theorem proving necessitate rigorous evaluation of their underlying benchmarks, revealing current defects as the field matures.
Reliable benchmarks are critical for the advancement and trustworthy deployment of AI systems, particularly in sensitive areas like automated reasoning, impacting their eventual integration into high-stakes applications.
The understanding of the robustness and actual capabilities of current LLM-assisted theorem provers is now more nuanced, requiring a re-evaluation of progress and a focus on improved benchmarking methodologies.
- · AI safety researchers
- · Developers of robust benchmarking tools
- · Formal verification specialists
- · Developers relying solely on current benchmarks
- · Hyperscalers overstating LLM capabilities in reasoning
- · Academic groups with flawed evaluation methods
Immediate re-evaluation of established LLM performance claims in theorem proving.
Increased investment in developing more sophisticated and adversarial-resilient AI benchmarks across various domains, not just theorem proving.
A broader institutional skepticism towards AI performance metrics, leading to more rigorous, formal verification of AI claims before deployment.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI