Researchers have developed StateSight, a benchmarking tool for evaluating vision-language models' ability to reconstruct latent spatial structures from single images. The tool assesses three tasks: cube-net opposite-face reasoning, occluded cube-tower counting, and 4-neighbor connected-component counting. OpenAI's GPT-5.5 and Claude Sonnet 5 models were tested on StateSight, achieving moderate accuracy but struggling with image-state reconstruction and reasoning procedures. A human baseline outperformed both models on every task. --- Why it matters: This matters because it highlights the limitations of current vision-language models in understanding spatial structures from images, which is crucial for applications like visual question answering and multimodal reasoning. Source: https://arxiv.org/abs/2608.20414