Reasoning over complex multi-hop and ambiguous spatial relations

Develop more reliable relation-aware reasoning for highly complex multi-hop expressions, ambiguous references, and relations such as “between,” “inside,” or “aligned with” in open-vocabulary 3D Gaussian Splatting referring segmentation.

Background

QAGaussian uses a query-conditioned relation-aware slot graph to model target-reference, spatial, and part-whole relationships. The authors nevertheless acknowledge that the current graph reasoning mechanism may have difficulty with expressions requiring multiple relational steps, ambiguous reference resolution, or less directly represented relations, leaving broader compositional and relational reasoning as an open problem.