Imagine an AI struggling to identify a simple triangle in an image. That's the surprising reality facing Multimodal Large Language Models (MLLMs)—AIs designed to understand both text and visuals. While they excel at tasks like image captioning and visual question answering, they often stumble when it comes to accurately describing geometric details—the positions of points, lengths of lines, angles, and the relationships between shapes. This limitation, known as low-level visual perception (LLVP), creates a significant hurdle for MLLMs in fields like math problem-solving, scientific image analysis, and even robotics. Researchers have introduced 'Euclid,' a novel approach to supercharge LLVP in MLLMs. Recognizing the scarcity of detailed geometric training data, the team built a synthetic data engine capable of generating countless examples of geometric shapes with high-fidelity descriptions. Think of it as an AI geometry tutor, crafting lessons with increasing complexity. This engine allowed them to study how different model architectures and training methods impacted an MLLM's ability to grasp geometric concepts. They found that Convolutional Neural Networks (CNNs), often used for image processing, were surprisingly more effective at preserving low-level visual information than the popular Vision Transformers (ViTs). Furthermore, a carefully structured learning curriculum—starting with simple shapes like lines and circles, then progressing to triangles and more complex figures—proved crucial. This 'learn-to-crawl-before-you-walk' approach enabled the MLLM to master challenging geometric concepts it couldn't learn when presented with complex shapes right away. Euclid, trained using this innovative approach, demonstrated remarkable performance on a new benchmark dataset called Geoperception. It handily outperformed state-of-the-art models in tasks like identifying points on lines and comparing line lengths, even though it had only been trained on synthetic data. While challenges remain, particularly with complex annotations and generalizing to diverse visual domains, Euclid represents a significant step toward enabling AIs to truly 'see' and understand geometry. This breakthrough has implications far beyond abstract math problems, paving the way for MLLMs to better interpret medical images, navigate complex environments, and unlock new possibilities in manufacturing and augmented reality.