AI summaryⓘ
The authors extend harmful meme detection by not only deciding if a meme is harmful but also identifying who or what it targets and where in the meme this appears. They introduce a new dataset called Meme3W that combines existing harmful meme data with detailed human-verified labels about targets. They propose a new strict evaluation metric, Joint Record Accuracy (JRA), to measure if the model gets all harmfulness and target details right together. To improve performance, the authors develop HarmTrace, a training approach that better teaches models to identify targets only when they correctly detect harmful memes. HarmTrace significantly improves detailed harmful meme detection on tested models.
multimodal harmful meme detectiontarget identificationMeme3W datasetJoint Record Accuracy (JRA)large multimodal language modelssupervised fine-tuningConditional Target-identification Policy Optimization (CTPO)Virtual Positive Anchor (VPA)harmfulness classificationentity-aware learning
Authors
Yujia Li, Yiqun Zhang, Zihan Cheng, Yijie Huang, Tenglong Ye, Zihan Wang, Xiaocui Yang, Shi Feng, Yifei Zhang, Daling Wang
Abstract
Multimodal harmful meme detection is typically formulated as image--text harmfulness classification. A model may correctly predict harmfulness while misidentifying the attacked target or its supporting evidence. We therefore extend harmful meme detection with fine-grained target identification, asking what type of target is attacked, who is targeted, and where the target appears in the meme. The model predicts harmfulness for every meme and, for harmful memes, outputs the target category, target entity, textual mention, and visual region. To support this task, we introduce Meme3W, which unifies multiple public harmful meme datasets and provides human-verified annotations for harmful instances. We further introduce Joint Record Accuracy (JRA), a strict record-level metric requiring the harmfulness label and all target-identification fields to be jointly correct. Experiments with representative multimodal large language models reveal a substantial gap between harmfulness accuracy and JRA. To narrow this gap, we propose HarmTrace, an anchor-calibrated decoupled optimization framework. HarmTrace strengthens target-entity supervision through entity-aware supervised fine-tuning. It then applies Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification advantages, restricting target-identification optimization to label-correct responses for harmful examples. CTPO uses a Virtual Positive Anchor (VPA) as a fully correct reference for target-identification advantage normalization. HarmTrace improves both JRA and harmfulness accuracy across the evaluated backbones, with JRA on the Qwen3-VL-8B backbone increasing from 17.58\% to 52.51\%. Our code is publicly available at https://github.com/llly1234/HarmTrace-for-Harmful-Memes.