vision-language models (VLMs)

SpatialClaw, a training-free framework that uses code as the action interface for spatial reasoning, achieves 59.9% average accuracy across 20 benchmarks, outperforming the recent spatial agent by +11…