ReRef-3D: A Benchmark for Spatial Referring Expression-Guided 3D Object Placement
ReRef-3D is a benchmark testing how 3D vision-language models can follow spatial instructions to place objects correctly in 3D scenes. It uses over 34,000 complex instructions derived from real scenes. The goal is not just finding a coordinate but ensuring the new placement satisfies all spatial relationships and remains physically valid within the scene, revealing that physical validity is a major challenge for current models.