Researchers from the University of Maryland and AWS have introduced LEGO-Anything, a project where coding agents build editable 3D scenes from a single photo by writing and refining code for Blender step by step. To measure performance, the team developed LEGO-Bench, a benchmark containing 208 images from simulator scenes that provide an exact ground truth for automated scoring. Testing six GPT configurations revealed that while all models reliably delivered working scene artifacts, their accuracy varied widely, with the best model scoring 53.4 percent on indoor scenes and 39.6 percent on outdoor scenes. The analysis showed that models struggle with unreliable self-assessment, often failing to judge whether their geometric revisions actually improved the scene. To fix this, the researchers created LEGO-Plugin, an extension that replaces self-judgment with concrete measurements and improves model accuracy, particularly for weaker agents. Why it matters: This research highlights the limitations of AI self-assessment in spatial tasks and provides a framework to make automated 3D scene reconstruction more reliable.