Search papers, labs, and topics across Lattice.
This paper addresses the limitations of existing harmful meme detection systems by introducing fine-grained target identification, which not only assesses harmfulness but also specifies the type of target, the entity being targeted, and its contextual presence in the meme. The authors present a new dataset, Meme3W, which consolidates multiple harmful meme datasets with human-verified annotations, and introduce Joint Record Accuracy (JRA) as a stringent evaluation metric. The proposed HarmTrace framework enhances target-entity supervision and employs Conditional Target-identification Policy Optimization (CTPO) to significantly improve both harmfulness accuracy and JRA, achieving a notable increase from 17.58% to 52.51% on the Qwen3-VL-8B model.
HarmTrace reveals that fine-grained target identification can drastically improve harmful meme detection accuracy, closing a critical gap in multimodal AI understanding.
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.
Apple ML