Research map / Label-efficient and robust learning / Semi-supervised learning
Multi-Agent VLMs Guided Self-Training with PNU Loss for Low-Resource Offensive Content Detection
Han Wang, Deyi Ji, Junyu Lu, Lanyun Zhu, Hailong Zhang, Haiyang Wu, Liqun Liu, Peng Shu, Roy Ka-Wei Lee
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 self-training with CLIP + Qwen72B pseudo-labels to the CLIP classifier | CLIP-Large, n=100 · FHM | macro-F1: 59.24 -> 72.68 | Improved |
| add self-training with CLIP + Qwen72B pseudo-labels to the CLIP classifier | CLIP-Large, n=100 · Sent140 | macro-F1: 64.22 -> 77.11 | Improved |
| verify classifier pseudo-labels with MA-VLMs (CLIP + Qwen72B) instead of CLIP-only labels (ablation) | SelfTrain CLIP, n=100 · MAMI Task A | macro-F1: 67.03 -> 73.49 | Improved |
| verify classifier pseudo-labels with MA-VLMs on HSOL (ablation) | SelfTrain CLIP, n=100 · HSOL | macro-F1: 86.48 -> 86.69 | No clear effect |
| use MA-VLMs moderator/user prompt instead of zero-shot (ablation) | Qwen2.5-VL-72B · MAMI Task A | macro-F1: 79.17 -> 81.64 | Improved |
| use MA-VLMs prompt instead of zero-shot (ablation) | Qwen2.5-VL-72B · FHM | macro-F1: 74.46 -> 74.62 | No clear effect |
| add few-shot examples to the VLM prompt (ablation) | Qwen2.5-VL-72B · FHM | macro-F1: 74.46 -> 71.09 | Hurt |
| add chain-of-thought to the VLM prompt (ablation) | Qwen2.5-VL-72B · FHM and MAMI | macro-F1: 74.46 -> 74.43 (FHM) | No clear effect |
| negative PNU weight gamma = -0.1 (NU learning) (ablation) | SelfTrain CLIP, n=100 · MAMI Task A | macro-F1: 73.49 -> 59.98 | Hurt |
| set PNU gamma = 0.1 instead of 0.0 (add PU term) on MAMI (ablation) | SelfTrain CLIP, n=100 · MAMI Task A | macro-F1: 68.84 -> 73.49 | Improved |
| PNU gamma = 0.1 instead of 0.0 on FHM (ablation) | SelfTrain CLIP, n=100 · FHM | macro-F1: 72.68 -> 71.50 | Hurt |
| increase labels from n=100 to n=250 for SupOnly Qwen7B vs SelfTrain CLIP (ablation) | Qwen7B vs SelfTrain CLIP · FHM | macro-F1: SelfTrain 72.97 vs Qwen7B 75.88 at n=250 | Mixed |
Closest papers by meaning
- T2Vs Meet VLMs: A Scalable Multimodal Dataset for Visual Harmfulness Recognition (NeurIPS 2024 · similarity 0.83)
- MAT-Agent: Adaptive Multi-Agent Training Optimization (NeurIPS 2025 · similarity 0.83)
- MLLM-as-a-Judge for Image Safety without Human Labeling (CVPR 2025 · similarity 0.83)
- Hateful Meme Detection through Context-Sensitive Prompting and Fine-Grained Labeling (Student Abstract) (AAAI 2025 · similarity 0.83)
- ICM-Assistant: Instruction-tuning Multimodal Large Language Models for Rule-based Explainable Image Content Moderation (AAAI 2025 · similarity 0.82)
- Taming Self-Training for Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.82)
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.
