Reproduce “Towards Low-Resource Harmful Meme Detection with LMM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jianzhao Huang, Hongzhan Lin, Ziyan Liu, Ziyang Luo, Guang Chen, Jing Ma
Researchers developed LOREHM, an agency-driven, gradient-free framework for detecting harmful memes in low-resource environments, leveraging Large Multimodal Models (LMMs). It significantly outperformed state-of-the-art baselines by augmenting LMMs with external contextual signals and internal self-reflection, improving macro-F1 scores by up to 4.40% on benchmark datasets.
Key insights
- LMMs can be effectively transformed into intelligent agents for few-shot multimodal tasks without needing gradient updates or model re-training.
- Combining external signals from similar labeled data (RSA) with internal self-reflection and knowledge refinement (MIA) provides comprehensive reasoning for LMM agents.
- The proposed agency-driven paradigm allows LMMs to derive generalizable rules for harmfulness, mimicking human learning and mitigating overfitting in low-resource scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.