RAMAR: retrieval-augmented multi-agent reasoning for zero-shot sarcasm detection

The automated detection of sarcasm presents a significant challenge for Natural Language Processing, requiring a sophisticated comprehension of contextual nuances, external world knowledge, and the implicit incongruity between literal semantics and intended meaning. While Large Language Models (LLMs) have demonstrated remarkable zero-shot reasoning abilities, their efficacy is frequently constrained by 1. Knowledge Limitation and 2. Monolithic Reasoning, which struggle to deconstruct the multifaceted nature of sarcastic utterances. To surmount these limitations, we introduce RAMAR, a novel Retrieval-Augmented Multi-Agent Reasoning framework for zero-shot sarcasm detection. RAMAR operationalizes a structured, three-stage deliberative reasoning process. First, a Knowledge-Informed Contextual Retrieval stage dynamically grounds the model by sourcing relevant, high-value contextual examples from a knowledge base. Second, a Multi-Agent Collaborative Analysis stage deploys specialized agents, each tasked with deconstructing the input from a distinct analytical perspective (e.g., semantic, rhetorical, knowledge-based). Finally, an Evidence-Weighted Adjudication mechanism synthesizes the agents’ diverse analyses, prioritizing analyses grounded in high-relevance retrieved evidence. Comprehensive experiments on four benchmark datasets show that RAMAR markedly outperforms state-of-the-art baselines (achieving an average Macro-F1 of 78.01%), setting a new standard for robust, explainable, and effective zero-shot sarcasm detection.