GeoSym127K: A Scalable Framework for Multimodal Geometric Reasoning

A group of researchers has launched GeoSym127K, a comprehensive dataset designed to improve multimodal geometric reasoning. This project tackles challenges seen in Large Multimodal Models (LMMs), like visual hallucinations and a lack of Chain-of-Thought data. The GeoSym Engine, which is an automated neuro-symbolic system, uses type-conditional grammar and an analytic SymGT Solver to generate accurate symbolic ground truths, along with a rendering pipeline for creating precise diagrams. The dataset itself includes 51,000 high-resolution images, 127,000 questions with symbolic ground truths, and 55,000 answer-verified CoT QA pairs, sorted by difficulty. Additionally, GeoSym-Bench offers a specially curated set of 511 complex samples for detailed evaluation, showing that GeoSym boosts performance in diagram and multi-step geometry tasks during supervised fine-tuning tests.