Can MLLMs Reason About Visual Persuasion? Evaluating the Efficacy and Faithfulness of Reasoning
The study tested whether prompting multimodal large language models (MLLMs) to reason before predicting visual persuasion outcomes consistently improves performance; it found that naive rationales are unreliable signals. Researchers developed a rationale supervision method using diverse teacher-generated rationales and introduced a three-dimensional faithfulness framework, concluding that prediction accuracy does not guarantee faithful reasoning.