TriGeoBench tests whether AI can reason from geometry figures, not just read them
A new benchmark pairs school-geometry questions with images and editable TikZ diagrams to probe what multimodal models understand about geometric structure.
When a diagram is part of the mathematics
A geometry question often depends on information that is hard to express in a sentence: which segments meet, what a label belongs to, or whether a line is an extension rather than a side. A preprint posted on October 20 introduces TriGeoBench, a benchmark for testing how multimodal AI systems handle those visual and structural parts of school mathematics.
Three representations of one problem
The dataset contains 7,000 question-and-answer pairs from Chinese middle- and high-school mathematics examinations, alongside 12,235 geometry figures. Each example combines a question in LaTeX, an ordinary RGB image, and TikZ source code for the figure. TikZ is a diagram-description language: unlike a flat image, it can preserve objects, labels, topology and stated relations in executable form.
The authors cover plane, solid, analytic, functional and vector geometry, with several response formats. They test models both on converting an image into TikZ and on solving the resulting mathematics. Their reported experiments find that fine-grained diagram structure and image-text alignment remain difficult. In an oracle setting, where models receive human-curated TikZ instead of an automatically generated conversion, the authors find that a more accurate structured representation can provide additional geometric cues. That oracle test is not an end-to-end system result.
A benchmark is a measurement tool, not a proof of understanding
TriGeoBench is useful because it separates two questions that are easy to blur together: can a model see a diagram accurately, and can it use its structure in a mathematical argument? The paper's results apply to its selected models, prompts, data and scoring rules. This is an unreviewed preprint, and MathsAI has not reproduced the evaluations or audited the dataset. For students and teachers, the practical lesson is familiar: a correct-looking sketch is not the same as a correctly interpreted geometric relationship.