arxivcs.CVcs.AI2026-07-09
MultiView-Bench: A Diagnostic Benchmark for World-Centric Multi-View Integration in VLMs
Hantao Zhang, Jinru Sui, Ed Li, Dirk Bergemann, Zhuoran Yang
Recent benchmarks for VLMs largely assess single- or limited-view perception, leaving untested the core cognitive ability to integrate observations across viewpoints into a coherent, world-centric (allocentric) 3D mental model. We introduce MultiView-Bench, a diagnostic benchmark…