See, Explain, and Intervene: A Few-Shot Multimodal Agent Framework for Hateful Meme Moderation

TL;DR

This paper addresses the challenge of moderating hateful memes by proposing a few-shot, multimodal agent framework that jointly performs detection, explanation, and intervention. It trains task-specific agents on curated datasets (MemeCap, HatReDAug, MemeSense) to generate silver data and then applies a few-shot prompting scheme with large models (GPT-4o, Intern-VL3, Pixtral) to achieve end-to-end triad performance. Evaluation on extended FHM and MAMI datasets demonstrates state-of-the-art macro-F1 for classification and high-quality explanations and interventions, highlighting the viability of low-resource, generalizable moderation in production. The work also provides a detailed analysis of linguistic properties, coherence, and sentiment in the generated explanations and interventions, discusses limitations, and outlines ethical considerations and data-release plans for the research community.

Abstract

In this work, we examine hateful memes from three complementary angles - how to detect them, how to explain their content and how to intervene them prior to being posted - by applying a range of strategies built on top of generative AI models. To the best of our knowledge, explanation and intervention have typically been studied separately from detection, which does not reflect real-world conditions. Further, since curating large annotated datasets for meme moderation is prohibitively expensive, we propose a novel framework that leverages task-specific generative multimodal agents and the few-shot adaptability of large multimodal models to cater to different types of memes. We believe this is the first work focused on generalizable hateful meme moderation under limited data conditions, and has strong potential for deployment in real-world production scenarios. Warning: Contains potentially toxic contents.

Figures (6)

• Figure 1: Overview of our novel task formulation.
• Figure 2: Overview of fine-tuning task specific agents and using them for silver data generation of FHM and MAMI datasets.
• Figure 3: Token count, type token ratio, and perplexity along with error bars at 95% confidence interval. Here, cp: correct positive (circle), cn: correct negative (square), wp: wrong positive (triangle), wn: wrong negative (diamond), g:GPT-4o (maroon), i:Intern-VL3 (purple), and p:Pixtral (green).
• Figure 4: Coherence analysis based on semantic similarity for all considered models and datasets. Here, cp: correct positive, wp: wrong positive, g: GPT-4o, i: Intern-VL3, and p: Pixtral.
• Figure 5: Bar charts presenting the distribution of sentiments for all models and across both datasets. Here, cp: correct positive, cn: correct negative, wp: wrong positive, wn: wrong negative, ex: explanation, in: intervention.