Evaluating the Capabilities of LLMs for Persuasive Dialogue

Large language models (LLMs) can generate apparently highly persuasive text, but does sounding persuasive mean arguing well? We introduce PERSUASIO, a multi-agent dialogue platform grounded in a formal argumentationbased theory of persuasion dialogues that adjudicates logical winners during free-text debates. Using this system, we generated 192 debates on a UK political topic between humans and LLMs, and evaluated 22 interlocutors through both automated adjudication and 9,702 crowdsourced pairwise judgements across 1,386 annotation instances. We observed a consistent decoupling between subjective and formal persuasiveness: LLMs dominated the subjective ranking yet performed substantially worse under argumentation-theoretic adjudication, where humans remained competitive. Multi-agent and retrieval-augmented variants further widened this divergence. These findings reveal a systematic gap between rhetorical fluency and formal argumentative strength in LLM-based persuasive dialogues.

Introduction. Persuasion has been studied for millennia, from Aristotle’s Rhetoric (Aristotle, 2010) to modern computational models of argumentation and persuasion (Hunter, 2016). Persuasive communication aims to shape or change the beliefs of others (Stiff and Mongeau, 2016). Recent work shows that LLMs can generate arguments that humans find compelling across political, health, and social domains (Bai et al., 2025; Matz et al., 2024; Salvi et al., 2025; Costello et al., 2024). However, it remains unclear whether perceived dialogical persuasiveness corresponds to logical persuasiveness. Existing evaluations typically assess either rhetorical appeal through human judgements (Tan et al., 2016; Habernal and Gurevych, 2016) or structural properties of arguments through automated analysis (Stab and Gurevych, 2014; Wachsmuth et al., 2017), but rarely both jointly. As a result, we lack direct evidence on whether model fluency and confident framing align with inferential rigour and logical consistency.

Discussion / Conclusion. Across analyses, a consistent divide emerges: LLMs dominate subjective persuasiveness yet fall sharply in logical ranking, while humans exhibit the opposite pattern. The LLMs and agentic variants studied appear optimised for rhetorical polish – fluency, structural clarity, confident framing – rather than formal argumentative rigour. Logical adjudication, grounded in formal coherence, rewards different properties than those that drive human perceptions of persuasiveness. The pronounced rank inversions among topperforming subjective systems perhaps highlights a fundamental limitation of LLMs for logical reasoning: their stochastic, autoregressive paradigm renders them inherently myopic, precluding strategic planning or deliberate reasoning over upcoming content. This limitation is potentially compounded by post-training procedures – namely instruction tuning and reinforcement learning from human feedback (Ouyang et al., 2022) – which optimise for rhetorical fluency and may thereby further obscure deficiencies in logical reasoning. Our results show that multi-agent orchestration and retrieval augmentation seemingly amplify this effect by enhancing perceived persuasiveness without reliably safeguarding logical integrity.