From Claims to Stance: Zero-Shot Detection with Pragmatic-Aware Multi-Agent Reasoning

Stance detection aims to identify whether a text expresses a favorable, opposing, or neutral attitude toward a given target and has become increasingly important for analyzing public discourse on social media. Existing approaches, ranging from supervised neural models to prompt-based large language models (LLMs), face two persistent challenges: the scarcity of annotated stance data across diverse targets and the difficulty of generalizing to unseen targets under pragmatic and rhetorical variation. To address these issues, we propose PAMR (Pragmatic-Aware Multi-Agent Reasoning), a zero-shot stance detection framework that decomposes stance inference into structured reasoning steps. PAMR orchestrates three LLM-driven agents—a linguistic parser that extracts pragmatic markers and canonicalizes claims, an NLI-based estimator that produces calibrated stance probabilities through consensus voting, and a counterfactual and view-switching auditor that probes robustness under controlled rewrites. A stability-aware fusion integrates these signals, conservatively abstaining when evidence is uncertain or inconsistent. Experiments on SemEval-2016 and COVID-19-Stance show that PAMR achieves macro-F1 scores of 71.9% and 73.0%, surpassing strong zero-shot baselines (FOLAR and LogiMDF) by +2.0% and +3.1%. Ablation results confirm that pragmatic cues and counterfactual reasoning substantially enhance robustness and interpretability, underscoring the value of explicit reasoning and pragmatic awareness for reliable zero-shot stance detection on social media.