Ten Papers Chart How Vision-Language Models Gained Spatial Reasoning Skills

Vision-language models (VLMs) have long been able to describe images in detail but struggled with basic spatial tasks like estimating distances or understanding 3D geometry. A researcher has mapped a coherent line of work spanning 2024 to 2026, tracing how ten papers progressively addressed this gap between discrete language tokens and continuous 3D space. The thread covers approaches ranging from large-scale synthetic spatial training data and depth plugins to reinforcement learning with verifiable rewards and real-time 3D reconstruction during inference. A benchmark called BLINK first quantified the problem, showing humans outperformed GPT-4V 95.7% to 51.3% on tasks solvable at a glance. The full analysis, including diagrams and code for each paper, has been published as a ten-chapter course on the author's blog.