IRAC: Retrieval-Augmented Multi-Agent Collaboration for Multimodal Sarcasm Detection

Multimodal Sarcasm Detection (MSD) aims to identify ironic intent in image–text pairs, where sarcasm typically arises from inconsistencies or semantic incongruity between linguistic and visual cues. However, most MSD research still relies on coarse binary labels, and even recent efforts that enrich annotations seldom identify which modality drives the irony—that is, its underlying source. This limitation leads models to fit shallow statistical correlations rather than understanding deep cross-modal interactions, while traditional supervised methods suffer from restricted generalization in open-world scenarios. In this paper, we present MIRS, a large-scale benchmark for Multimodal Irony Reasoning and Source-attribution, containing 11,000 samples with fine-grained four-layer hierarchical annotations. To tackle the demands of zero-shot explainable analysis—requiring both precise detection and human-aligned diagnostics—we propose IRAC (Irony-oriented Retrieval-Augmented multi-agent Collaboration). IRAC employs a specialized swarm of agents (Textual, Visual, and Cross-modal) and an orchestrator optimized via Group Relative Policy Optimization (GRPO) with irony-specific rewards. Experimental results demonstrate that IRAC achieves state-of-the-art performance (86.92% Acc) and significantly reduces hallucination rate, providing a new paradigm for explainable multimodal irony analysis.