Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models
A fully automated, concept-driven pipeline that bridges the compositional gap by generating photorealistic compositional benchmarks at scale and uncovers a task-dependent trade-off: visio-linguistic context aids relational reasoning but hinders attribute binding through visual clutter.
Cristian Sbrolli, T. Yamasaki, Matteo Matteucci
· 0 citations