Advocate for Complete Benchmarks for Formal Reasoning with Formal/Informal Statements and Formal/Informal Proofs

Roozbeh Yousefzadeh, Xuenan Cao

Abstract

Formal reasoning—the intersection of artificial intelligence and machine-verifiable mathematics—stands at a critical juncture. While large language models (LLMs) show promise in automated theorem proving, the field is currently hindered by "incomplete" benchmarks, unreliable evaluation metrics, and a culture of data secrecy. This paper argues for a new standard in formal reasoning research: the adoption of "Complete Benchmarks" containing formal and informal versions of both problem statements and proofs, coupled with an open-science approach modeled after the success of the computer vision community.

Introduction

The ability of AI to reason within formal systems like Lean is often considered a frontier for machine learning. Unlike natural language tasks, formal reasoning provides a unique advantage: the ability to verify correctness through a compiler. However, this potential is being undermined by systemic bottlenecks. High-quality data is scarce, and the benchmarks used to measure progress are often riddled with errors or lack the necessary components to truly test a model's reasoning capabilities. Furthermore, a growing trend toward keeping training data and code private—often justified by the fear of "data contamination"—is preventing the community from understanding model failures and building upon existing work.

The Crisis of Evaluation

A primary concern in current research is "evaluation inflation." As the complexity of formal reasoning tasks increases, researchers have turned to LLMs to judge the outputs of other LLMs. This study demonstrates that such automated pipelines are fundamentally unreliable.

In a reproduction and audit of the Herald model—a state-of-the-art autoformalization system—the authors found a significant discrepancy. While the automated LLM judge reported a success rate of 97% on the miniF2F benchmark, manual verification by Lean experts revealed a true accuracy of only 67%. LLMs often produce "plausible-looking" formal statements that are semantically incorrect; these nuances are frequently missed by LLM judges but are immediately flagged by a Lean compiler or a human expert.

The Anatomy of a Complete Benchmark

To address these issues, this paper proposes that the community move toward "Complete Benchmarks." A benchmark should not merely consist of a problem statement and a binary "correct/incorrect" label. Instead, it must include four essential components:

  1. Informal Problem Statements: Natural language descriptions of the mathematical problem.
  2. Formal Problem Statements: The problem translated into a machine-verifiable language (e.g., Lean).
  3. Informal Proofs: Human-readable, step-by-step mathematical reasoning.
  4. Formal Proofs: Machine-verifiable code that completes the proof.

Without ground-truth formal proofs provided by benchmark creators, errors can persist indefinitely. For instance, four years after the release of miniF2F, new unprovable theorems were still being discovered as recently as April 2025. Providing the full proof allows researchers to verify the benchmark's own integrity and provides a richer signal for training.

Beyond Binary Metrics: Proof Length and Complexity

Currently, success in formal reasoning is often reported as a binary accuracy metric. This paper argues that this is misleading. There is often a total lack of correlation between a model's output length and the actual mathematical requirements of a problem.

In one striking example (IMO 1985 P6), a model generated a 3-line proof for a problem that requires over 1,200 lines of manual Lean code. In such cases, the model has likely simplified the problem out of existence or exploited a triviality rather than solving the actual mathematical challenge. The authors suggest that researchers should report proof lengths and the use of automatic tactics (like linarith). If a model's proof is orders of magnitude shorter than a human-written formal proof, it warrants closer scrutiny.

The Transparency Tax and the ImageNet Parallel

The field currently suffers from a "Black Box" culture. To avoid contamination, many teams keep their formal proofs, fine-tuning scripts, and search algorithms (such as those used in BFS-Prover) private. This forces the community into "parallel redundant work," where researchers spend months attempting to replicate baselines rather than innovating.

The authors draw a historical parallel to ImageNet in the computer vision field. The rapid progress in vision was driven by open access to labels and data. While the risk of LLMs "memorizing" proofs is real, the authors argue that memorization is a necessary step toward reasoning. The benefits of transparency—allowing others to see why a model failed—far outweigh the risks of contamination.

Conclusion

Writing formal proofs is a form of intensive manual labor, akin to developing complex software. This effort should be credited, shared, and used to build a more robust foundation for the field. By adopting Complete Benchmarks and embracing open-science principles, the formal reasoning community can move past "plausible-looking" results toward verifiable, reproducible progress in AI mathematics.

Summary updated

BibTeX
@misc{yousefzadeh2025advocate,
  author = {Roozbeh Yousefzadeh and Xuenan Cao},
  title = {{Advocate for Complete Benchmarks for Formal Reasoning with Formal/Informal Statements and Formal/Informal Proofs}},
  year = {2025},
  eprint = {2507.04719},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.04719},
}
Figure