arXiv:2509.24473v3 Announce Type: replace-cross 
Abstract: Spatial intelligence spans a rich suite of abilities, including visualising and transforming shapes, mentally rotating objects, judging relational positions and containment, and estimating numerosity. However, it still remains a critical unresolved challenge for Multimodal Large Language Models (MLLMs). To fill this gap, we propose to treat Euclidean geometry problem-solving as a surrogate task. Specifically, we meticulously constructed a curated multimodal dataset, called Euclid30K, comprising approximately 30K plane and solid geometry problems. Furthermore, to enable the model to learn and apply Euclidean principles from these geometry problems, we fine-tuned seven model variants (spanning 3--72B parameters) from the Qwen2.5VL, Qwen3VL, and RoboBrain2.0 families using Group Relative Policy Optimization (GRPO), inspiring the models to identify shapes, count, and relate entities, and perform multi-step deductive reasoning using Euclidean principles. Our experiments demonstrate that the resulting models achieve substantial zero-shot gains across four spatial reasoning benchmarks (Super-CLEVR, Omni3DBench, VSI-Bench, and MindCube) without any task-specific adaptations. Notably, after training on the Euclid30K, the mean VSI-Bench accuracy rose from 36.6\% to 41.8\% (+5.2\%), and the mean MindCube accuracy rose from 31.4\% to 38.1\% (+6.7\%). To our knowledge, this is the first systematic study showing that geometry-centric fine-tuning can confer vision-language models with broadly transferable spatial skills. Code and Euclid30K dataset can be found in \href{https://zgca-ai4edu.github.io/Euclids_Gift}{this}.

يتناول المقال تطوير Euclid30K، وهو مجموعة بيانات متعددة الوسائط مصممة لتعزيز الذكاء المكاني في نماذج اللغة متعددة الوسائط (MLLM). من خلال اعتبار حل مسائل الهندسة الإقليدية كوظيفة بديلة، قام المؤلفون بضبط سبع متغيرات نموذجية من عائلات Qwen2.5VL وQwen3VL وRoboBrain2.0. تهدف هذه المقاربة إلى تحسين قدرات النماذج في التعرف على الأشكال، والعد، والاستدلال الاستنتاجي متعدد الخطوات.

El artículo aborda el desarrollo de Euclid30K, un conjunto de datos multimodal diseñado para mejorar la inteligencia espacial en Modelos de Lenguaje Multimodal (MLLM). Al tratar la resolución de problemas de geometría euclidiana como una tarea sustituta, los autores ajustaron siete variantes de modelos de las familias Qwen2.5VL, Qwen3VL y RoboBrain2.0. Este enfoque busca mejorar las habilidades de los modelos en identificación de formas, conteo y razonamiento deductivo en múltiples pasos.

L'article traite du développement d'Euclid30K, un ensemble de données multimodal conçu pour améliorer l'intelligence spatiale des modèles de langage multimodal (MLLM). En considérant la résolution de problèmes de géométrie euclidienne comme une tâche de substitution, les auteurs ont affiné sept variantes de modèles des familles Qwen2.5VL, Qwen3VL et RoboBrain2.0. Cette approche vise à améliorer les capacités des modèles en matière d'identification de formes, de comptage et de raisonnement déductif en plusieurs étapes.

The article discusses the development of Euclid30K, a multimodal dataset designed to enhance spatial intelligence in Multimodal Large Language Models (MLLMs). By treating Euclidean geometry problem-solving as a surrogate task, the authors fine-tuned seven model variants from the Qwen2.5VL, Qwen3VL, and RoboBrain2.0 families. This approach aims to improve the models' abilities in shape identification, counting, and multi-step deductive reasoning.

Euclid's Gift: Enhancing Spatial Perception and Reasoning in Vision-Language Models via Geometric Surrogate Tasks

One More Thing in AI – Your Shortcut to AI Mastery

Euclid's Gift: Enhancing Spatial Perception and Reasoning in Vision-Language Models via Geometric Surrogate Tasks

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

Magicley AI

The Visualizer

Zemith-3bda3b

Guidejar-4eb95b

Ready to build your own newsroom?