Researchers introduced JUPE, a benchmark for visual-question-answering. It fixes answers while varying explanations. Fourteen multimodal large language models were tested as judges. Plausible explanations misled judges as much as persuasive arguments. Plain justifications reduced rejection of wrong answers by 12.7 percentage points. Chain-of-thought rationales reduced it by 13.8 percentage points. The effect appeared across domains, explanation generators and judging protocols. Mitigations offered limited, model-dependent protection. The paper calls this failure plausibility capture. It can inflate performance estimates when judges evaluate answers with explanations.

• 14 multimodal large language models tested
• 13.8 percentage points reduction with chain-of-thought rationales