Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Tool-meme: Tool-Learning Framework for Multimodal Meme Detection

AbstractWith the explosive growth of social media, multimodal memes (text and images) have become a key carrier of misinformation dissemination and emotional manipulation because they strongly depend on cultural background, satirical metaphor, and current events. Existing detection methods mostly rely on end-to-end deep learning models, but they face two fundamental limitations: first, static models are difficult to capture the dynamic semantic evolution of memes; second, fixed architectures cannot adapt to meme heterogeneity. This work proposes a Tool-Learning Framework for Multimodal Meme Detection (Tool-meme), a Large Multimodal Model (LMM)-driven tools-learning framework. The framework uses multimodal representations only as input features, dynamically orchestrates tool chains through retrieval-augmented generation and meta-cognitive planning, and uses reinforcement learning to maximize tool-use efficiency. On HarM, FHM, and MAMI, Tool-meme outperforms all evaluated direct-LMM and agentic baselines under frozen-executor setting. Additional transfer and deployment analyses show consistent gains across tested executor families while lowering tool calls and cost vs. exhaustive execution, with explicit evidence chains for inspection.

More from EMNLP 2026

All EMNLP 2026 papers →