| task_categories: | |
| - image-text-to-text | |
| language: | |
| - en | |
| tags: | |
| - geometry | |
| - multimodal | |
| - geometry-problem-solving | |
| # GeoFocus | |
| [Paper](https://huggingface.co/papers/2602.08524) | [Code](https://github.com/dle666/GeoFocus) | |
| GeoFocus is a novel framework for Multimodal Geometry Problem-Solving (MGPS). It addresses the challenges of recognizing global shapes and intricate local geometric relationships through two core components: | |
| 1. **Critical Local Perceptor**: Automatically identifies and emphasizes critical local structures (e.g., angles, parallel lines, comparative distances) through thirteen theory-based perception templates, boosting local feature coverage. | |
| 2. **VertexLang**: A compact topology formal language that encodes global figures using vertex coordinates and connectivity relations, reducing training time while improving topology recognition accuracy. | |
| ## Dataset Description | |
| The GeoFocus project involves several data splits used for training and evaluation: | |
| - **Global_Perceptor_Data**: Training data focused on global figure recognition using the VertexLang encoding. | |
| - **Local_Perceptor_Data**: Training data featuring fine-grained visual attribute annotations for critical local structures. | |
| - **Geo_test**: Evaluation datasets covering benchmarks such as Geo3K, GeoQA, and FormalGeo7K. | |
| The models trained on this data, GeoFocus-3B and GeoFocus-7B, demonstrate superior performance and robustness in geometry reasoning tasks. | |
| ## Citation | |
| If you use this work or dataset in your research, please cite the original paper: | |
| ```bibtex | |
| @article{geofocus2026, | |
| title={GeoFocus: Blending Efficient Global-to-Local Perception for Multimodal Geometry Problem-Solving}, | |
| author={...}, | |
| journal={arXiv preprint arXiv:2602.08524}, | |
| year={2026} | |
| } | |
| ``` |