NL2AGBench: Benchmarking LLM Auto-Formalization for AlphaGeometry
The Natural Language to AlphaGeometry Benchmark (NL2AGBench), which evaluates LLMs in translating English geometry problems into AlphaGeometry-compatible formal representations and introduces an error taxonomy distinguishing syntax and logic errors, which yield measurable improvements across multiple model families.