To improve the geometric reasoning capabilities of multimodal large-scale language models (MLLMs), acquiring large-scale, high-quality inference data is crucial. To overcome the limitations of existing data generation methods, we propose NeSyGeo, a novel neural-symbolic framework. NeSyGeo uses a domain-specific language that comprehensively represents all elements of planar geometry, synthesizes symbolic sequences to map them to visual and textual representations, and generates inference paths through backward search and forward validation. Based on this framework, we build the NeSyGeo CoT and NeSyGeo-Caption datasets, each containing 100,000 samples, and release NeSyGeo-Test, a new benchmark for evaluating the geometric reasoning capabilities of MLLMs. Experimental results demonstrate that the proposed method significantly improves the performance of several MLLMs, particularly with a small number of samples and a small number of training epochs.
