Multimodal large language models (MLLMs) are increasingly being used as judges in AI training and evaluation pipelines, often assessing answers alongside explanations. A new preprint on arXiv, titled A Plausible Explanation Is Enough to Mislead Multimodal Judges, examines whether these judges can be swayed by the quality of the explanation rather than the correctness of the answer.
The abstract notes that while AI judges are known to be sensitive to textual factors, it remains unclear how much a plausible-sounding explanation can influence their verdicts. The paper's title suggests a clear finding: a convincing explanation may be sufficient to mislead the judge, even if the underlying answer is flawed.
This raises practical concerns for any pipeline that relies on MLLM judges to score or rank model outputs. If explanations can override substantive correctness, evaluation results may reflect rhetorical quality rather than true performance. The preprint is a single source, so no contrasting findings are available, but the implication is direct: multimodal judging needs safeguards against persuasive but misleading rationales.