NL2AGBench: leaderboard
NL2AGBench (2026) asks LLMs to translate 48 English Olympiad-style geometry problems into the domain-specific language of AlphaGeometry, zero-shot with the syntax definitions in the prompt. The score is the share of translations that AlphaGeometry executes without error, in percent; higher is better.
Metric: Executable translation rate (%; share of 48 Olympiad-style geometry problems from the JGEX collection with verified AlphaGeometry formalizations whose English statement the model translates, zero-shot with the AlphaGeometry syntax definitions and handwritten clause rules in the prompt, into a specification that AlphaGeometry executes without error). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 75 | #121 |
| 2 | Grok 3 | 52.08 | #600 |
| 3 | Claude Sonnet 4.6 | 37.5 | #158 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=nl2agbench · How It Works · Data refreshed daily, snapshot 2026-11-12.