MVGraph Benchmark Evaluates LVLM Reasoning Across Multi-View Visual Graphs
October 12, 2026
MVGraph tests Large Vision-Language Models on 8,000 instances requiring evidence composition across multiple interrelated visual graph views. The benchmark decomposes performance into perception, alignment, composition, and reasoning to identify failures in multi-graph alignment.
HOW THIS AFFECTS YOU
●
researcherYou can use this benchmark to evaluate how well your models compose information across fragmented visual inputs.