Principled Evaluation of Multi-Label Persuasion in Advertisements with Large Vision-Language Models

The automatic detection of persuasive strategies in advertisements presents a uniquely multimodal challenge at the intersection of vision, language, and social cognition. While recent advances in Large Vision-Language Models (LVLMs) offer promising capabilities for such tasks, current approaches often rely on restrictive evaluation schemes that do not reflect the inherently multi-label nature of persuasive messaging. In this work, we reframe persuasion strategy detection as a genuine multi-label classification problem and propose a principled evaluation framework to enhance interpretability and robustness. We apply this approach to both image and video datasets, examine their characteristics, and introduce novel input-agnostic baselines that achieve macro F1-scores of 0.082 and 0.289 on the respective test sets. As part of this analysis, we study label co-occurrence patterns and dataset ambiguities, providing insights that inform both model interpretation and future dataset design. To assess the native capabilities of LVLMs, we benchmark three open-source models—PaliGemma, PaliGemma2, and Qwen2.5-VL—on the image persuasion dataset. Under zero-shot conditions, we demonstrate that querying each strategy individually with a logit-based decision threshold outperforms guided text generation. The best-performing zero-shot model, Qwen2.5-VL, achieves a macro F1-score of 0.227 and a sample F1-score of 0.234 on the image test set, and 0.381 and 0.400 respectively on the video test set. We further explore and compare two lightweight fine-tuning strategies that update only small subsets of model parameters while keeping the remaining weights frozen: fine-tuning of the image-to-text tokens linear projection and Low Rank Adaptation (LoRA) of the language model. Linear projector fine-tuning yields a top macro F1-score of 0.396 on the image test set, marking a substantial improvement over zero-shot performance. To evaluate cross-modal generalization, we apply fine-tuned image models to the video dataset. Our experiments reveal that, while projection-based fine-tuning enables partial knowledge transfer from image to video (macro F1=0.416 on a testing subset), LoRA adaptation severely disrupts cross-modal performance (macro F1=0.189). Finally, we perform a per-strategy performance analysis, looking into annotator- and data-centric factors that may influence LVLM performance. These findings highlight the viability of open-weight LVLMs for fine-grained persuasion analysis and suggest efficient pathways for domain-specific adaptation under realistic resource constraints.