Multimodal large language models excel in vision‑language tasks, yet their geometry‑centric reasoning remains poorly understood. Existing mathematical benchmarks mostly rely on rasterized figures, cover limited geometric domains, and use uniform answer formats, making it hard to tell whether models reason over geometric structures or exploit superficial visual patterns. To address this, we introduce TriGeoBench—a geometry‑focused multimodal benchmark built from Chinese middle‑ and high‑school mathematics examinations. TriGeoBench comprises 7,000 question‑answer pairs and 12,235 geometric figures accompanied by TikZ code, spanning plane geometry, solid geometry, analytic geometry, functions, and vector geometry with diverse answer formats. Each instance is a {question‑answer, image, TikZ} triplet: the Q&A is stored in LaTeX, images are RGB bitmaps, and TikZ preserves geometric topology, symbolic annotations, and relational structure. We evaluate representative open‑source and closed‑source MLLMs via image‑to‑TikZ conversion and multimodal reasoning probes. Results reveal persistent challenges in fine‑grained structural understanding and image‑text alignment. Oracle experiments using human‑curated reference TikZ show that accurate executable representations provide complementary geometric cues, though they do not assess a full end‑to‑end conversion pipeline. TriGeoBench thus offers a testbed for structure‑aware mathematical reasoning.
Review