Formal theorem proving with language models depends on frequent calls to a verifier, such as Lean. According to a new arXiv abstract, these checks are often slow or time out, leaving GPUs idle and potentially denying rewards to proofs that are correct but take too long to verify.

The paper, titled "Learning Lean Feedback for Efficient Proof-Candidate Evaluation on GPUs," describes training neural evaluators to address this bottleneck. The abstract states that the authors train these evaluators to... but the text cuts off, so the specific architecture, training method, and reported results are not available from the source.

As the only source is a truncated abstract, this summary is limited to the problem statement and the general approach. No experimental outcomes or comparisons are provided in the available text.