Can Thinking Models Think to Detect Hateful Memes?

TL;DR

This paper explores the use of thinking-based multimodal large language models with reinforcement learning to improve the detection of hateful memes through enhanced reasoning and explanation quality.

Contribution

It introduces a novel GRPO-based training framework, extends datasets with rationales, and achieves state-of-the-art results in hateful meme detection.

Findings

Improved accuracy and F1 scores on Hateful Memes benchmark

Enhanced explanation quality with step-by-step reasoning

Effective reinforcement learning approach for multimodal reasoning

Abstract

Hateful memes often require compositional multimodal reasoning: the image and text may appear benign in isolation, yet their interaction conveys harmful intent. Although thinking-based multimodal large language models (MLLMs) have recently advanced vision-language understanding, their capabilities remain underexplored for hateful meme analysis. We propose a reinforcement learning based post-training framework that improves reasoning in thinking-based MLLMs through task-specific rewards and a novel Group Relative Policy Optimization (GRPO) objective. Specifically, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful meme understanding, (ii) extend an existing hateful meme dataset by generating weakly or pseudo-supervised chain-of-thought rationales via distillation, and (iii) introduce a GRPO-based objective that jointly optimizes meme classification and explanation quality to encourage fine-grained, step-by-step reasoning. Experiments on the Hateful Memes benchmark show that our approach achieves state-of-the-art performance, improving accuracy and F1 by approximately 1 percent and explanation quality by approximately 3 percent. We will publicly release our code, dataset extensions, and evaluation resources to support reproducibility.