Research map / Large language models / Agents, code and math
NoReGeo: Non-Reasoning Geometry Benchmark
Irina Abdullaeva, Anton Vasiliuk, Elizaveta Goncharova, Temurbek Rahmatullaev, Zagorulko Ivan, Maxim Kurkin, Andrey Kuznetsov
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| give full diagram images instead of dotted images to VLMs (ablation) | VLMs · NoReGeo, basic coordinate tasks (mean over models) | accuracy (%): 19.2 -> 53.4 | Improved |
| give full images instead of dotted images (ablation) | Qwen2.5-VL-7B-Instruct · NoReGeo numeric basic coordinate tasks | accuracy (%): 3.4 -> 79.0 | Improved |
| give full images instead of text-only for unstable circle properties (ablation) | Qwen2.5-VL-7B-Instruct · NoReGeo unstable simple circle properties | accuracy (%): 0.0 -> 99.0 | Improved |
| give full images instead of dotted images for parallelism, perpendicularity and geometric transformations (ablation) | VLMs · NoReGeo PAP and GTR categories | accuracy (%) | No clear effect |
| give full images instead of dotted images (ablation) | InternVL-3, G-LLaVA-13, URSA · NoReGeo | accuracy (%) | No clear effect |
| use math-specialized fine-tuned text-only models | Qwen2.5-Math-7B-Instruct · NoReGeo classification and unstable tasks | accuracy (%) | Hurt |
| fine-tune a linear probe on full images instead of freezing the CLIP backbone (ablation) | CLIP ViT-B/32 probe · full-image test, 8 binary tasks | accuracy (%): 92.6 -> 97.4 | Improved |
| train probe on full images, test on dot images (ablation) | CLIP ViT-B/32 probe · dot-image test | accuracy (%): 97.4 -> 58.6 | Hurt |
| retrain probe on dots instead of full images (ablation) | CLIP ViT-B/32 probe · dot-image test | accuracy (%): 58.6 -> 83.9 | Improved |
| train probe on one task family and evaluate on a related family (ablation) | CLIP ViT-B/32 probe · cross-task transfer | accuracy (%): 86.0 on right-angle transfer from parallel-line probes | Improved |
| fine-tune models on geometry alone to gain native geometric understanding (ablation) | LLMs and VLMs · NoReGeo | accuracy (%) | No clear effect |
Closest papers by meaning
- Do Large Language Models Truly Understand Geometric Structures? (ICLR 2025 · similarity 0.92)
- SolidGeo: Measuring Multimodal Spatial Math Reasoning in Solid Geometry (NeurIPS 2025 · similarity 0.90)
- GeoX: Geometric Problem Solving Through Unified Formalized Vision-Language Pre-training (ICLR 2025 · similarity 0.89)
- Hilbert-Geo: Solving Solid Geometric Problems by Neural-Symbolic Reasoning (CVPR 2026 · similarity 0.88)
- Geoint-R1: Formalizing Multimodal Geometric Reasoning with Dynamic Auxiliary Constructions (CVPR 2026 · similarity 0.88)
- G-LLaVA: Solving Geometric Problem with Multi-Modal Large Language Model (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
