MathAtlas - Statements: leaderboard
Autoformalization of graduate and research-level theorems, examples and exercises from textbooks and papers into Lean 4 theorem statements, proofs not required; MathAtlas (Patel et al., 2026). Percent of statements whose formalization compiles against Mathlib and is judged faithful by CriticLean-32B, with zero-shot prompting for general models and each fine-tuned autoformalizer on its own prompt; higher is better.
Metric: Correctness (%; share of all theorem, example and exercise statements whose Lean 4 formalization compiles against Mathlib (Lean v4.24.0) and is judged faithful to the informal text by CriticLean-32B; zero-shot prompt for the general models, each fine-tuned formalizer's own prompt). Source: arxiv.org. Saturation forecast: Around August 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 6.8 | #699 |
| 2 | GPT-OSS-20B | 4.2 | #1150 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mathatlas-statements · How It Works · Data refreshed daily, snapshot 2026-11-12.