SketchAgentBench – evaluate LLMs on 3D spatial reasoning
7 points by wsxiaoys
SketchAgentBench evaluates 11 large language models (LLMs) on 3D spatial reasoning using a methodology that involves a single photo and prompt to generate 3D scenes. The evaluation included 25 photos and 275 runs, with each model’s output scored based on the final 3D scene quality. The results provide insights into the models’ capabilities in understanding and rendering spatial information.