MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph

Key points

• Although such evidence commonly resides in source-specific graphs, temporal snapshots, or local observations, existing graph reasoning benchmarks remain confined to a single self-contained structure, assessing only graph perception and reasoning while leaving the alignment and composition of evidence across views underexplored.1
• MVGraph comprises 8k instances spanning 8 graph-theoretic tasks, abstracted from two multi-graph scenarios (partial and temporal graph views).1
• Beyond end-to-end evaluation, we decompose multi-graph reasoning into four complementary capabilities---Perception, Alignment, Composition, and Reasoning---and design 29 diagnostic tasks with 64k instances to probe each one, along with golden intermediate representations for evaluating the intermediate process.1
• Across 15 LVLMs, we reveal the difficulty of multi-graph reasoning and identify cross-view edge alignment and relation-level composition as prominent bottlenecks through stage-oriented diagnostics.1

Sources (1)

• [1]MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 03:59 AM

  We introduce MVGraph, a benchmark for evaluating large vision-language models (LVLMs) on multi-graph reasoning where the evidence required by a query is distributed across multiple interrelated visual graph views.

  Although such evidence commonly resides in source-specific graphs, temporal snapshots, or local observations, existing graph reasoning benchmarks remain confined to a single self-contained structure, assessing only graph perception and reasoning while leaving the alignment and composition of evidence across views underexplored.

Extractive summary: sentences quoted from the sources.