Research map / Large language models / Language, safety and interpretability
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
Jingbiao Mei, Mingsheng Sun, Jinghong Chen, Pengda Qin, Yuhong Li, Da Chen, Bill Byrne
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add policy-manual augmented SFT warmup (SFT-PM) (ablation) | GRPO, Qwen2.5-VL-7B · HatefulMemes | attack F1: 61.2 -> 70.8 | Improved |
| add GRPO curriculum learning (fine-grained before binary) (ablation) | ExPO-HM, Qwen2.5-VL-7B · HatefulMemes | binary F1: 75.8 -> 78.4 | Improved |
| add CDE reward (ablation) | ExPO-HM, Qwen2.5-VL-7B · HatefulMemes | binary F1 / LLM-judge: 81.1 vs 78.4; LLM 5.8 -> 6.2; CDE 0.056 -> 0.026 | Improved |
| use Explain-then-Detect post-training (SFT/DPO/GRPO) instead of direct SFT | Qwen2.5-VL-7B · HatefulMemes | binary F1: 75.0 -> 74.5 (GRPO), 73.6 (DPO) | Hurt |
| add GRPO to Explain-then-Detect SFT | Qwen2.5-VL-7B · HatefulMemes | reasoning LLM-judge: 5.0 -> 5.2 | No clear effect |
| use DPO instead of zero-shot | Qwen2.5-VL-7B · HatefulMemes | reasoning LLM-judge: 5.0 -> 4.9 | Hurt |
| binary-only SFT warmup vs no warmup before GRPO-CL+CDE (ablation) | ExPO-HM · HatefulMemes | binary F1 / attack F1: attack 69.3 -> 66.8 | Hurt |
| policy-manual warmup vs fine-grained label warmup (ablation) | ExPO-HM · HatefulMemes | target F1: 73.4 -> 77.2 | Improved |
| ExPO-HM vs GRPO | ExPO-HM · HatefulMemes | attack F1: 61.2 -> 75.6 | Improved |
| ExPO-HM on 3B model vs GRPO | ExPO-HM · PrideMM | binary F1: 72.1 -> 75.6 | Improved |
Closest papers by meaning
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.88)
- TRACE: Textual Relevance Augmentation and Contextual Encoding for Multimodal Hate Detection (AAAI 2026 · similarity 0.88)
- Meme Trojan: Backdoor Attacks Against Hateful Meme Detection via Cross-Modal Triggers (AAAI 2025 · similarity 0.86)
- A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings (NeurIPS 2025 · similarity 0.85)
- Spanning the Spectrum of Hatred Detection: A Persian Multi-Label Hate Speech Dataset with Annotator Rationales (AAAI 2024 · similarity 0.85)
- Multi-Modal Discussion Transformer: Integrating Text, Images and Graph Transformers to Detect Hate Speech on Social Media (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
