RoTA-Bench Exposes Weak 3D Spatial Reasoning in VLMs

Xiamen University and NTU released RoTA-Bench, which tests VLM spatial reasoning via interactive 3D-to-2D projection alignment with varied rotation steps and IoU≥0.90 thresholds. Humans score near-perfect while the best model reaches only about 17%.

2026-10-16 ~ 2026-10-17 · 2 related posts

• RoTA-Bench: VLMs show large gaps vs humans in interactive 3D projection alignment — MAC-AutoML · 2026-10-16
• RoTA-Bench: humans solve ~100% of 3D rotation alignment tasks, best model GPT-5.6-Sol only ~17% — _akhaliq · 2026-10-17