NVIDIA published Spatial-IQ, a BSD-3-Clause licensed reproducibility artifact for the paper "Spatial-IQ: Deconstructing Spatial Intelligence via Hierarchical Capability Tests" (Rim et al., arXiv:2607.22864). The benchmark evaluates multimodal LLMs on 3D spatial reasoning using procedurally generated stacked-block scenes rendered in Isaac Sim 4.5, with each view carrying 11 atomic spatial sub-tasks plus a composite block-counting target that requires inferring occluded supporting blocks. Models can answer in three modalities — free-response text, image-option MCQ, and image editing — scored by exact match/Pearson correlation, letter match, and silhouette score against ground-truth masks, with a separate OCR consensus path (Tesseract, PaddleOCR, Qwen2.5-VL) for image-output models.
- github.com2026-07-31