Sophistry to Logic: Mitigating Persuasiveness-Fluency Feature Entanglement via Latent Variable Estimation

ACL ARR 2026 January Submission6138 Authors

Keywords: Persuasive and Debate Generation, Persuasion Modeling, Preference Optimization

Abstract: Effective persuasion relies on two distinct pillars: \textit{argumentative logic} and \textit{rhetorical fluency}. However, existing persuasive generation methods often conflate these dimensions by optimizing joint reward signals, where surface-level fluency dominates logical substance. We term this phenomenon \textbf{Feature Entanglement}, a pathology where models prioritize the surface fluency—producing ``well-formatted hallucinations''—over the underlying mechanics of perusasion. To address this, we propose \textbf{$\text{P}^3$}, a framework designed to decouple these attributes via latent variable modeling. The framework operates in three stages: (1) \textbf{\underline{P}ersuasiveness Reward Estimation} employs an Expectation-Maximization (EM) algorithm to explicitly distinguish latent persuasiveness from superficial fluency; (2) \textbf{\underline{P}ersuasiveness Sample Mining} leverages these disentangled signals to filter out rhetorical noise; and (3) \textbf{\underline{P}ersuasiveness Strategy Optimization} introduces Persuasion Augment Policy Optimization (PAPO), a novel objective that uses decoupled scores to dynamically scale policy updates. Experimental results demonstrate that a 13B parameter model trained with $\text{P}^3$ surpasses the efficient commercial models (e.g., Gemini 1.5 Flash and Claude 3 Haiku) in both automatic and human evaluated persuasiveness.

Paper Type: Long

Research Area: Natural Language Generation

Research Area Keywords: domain adaptation,text-to-text generation,model architectures

Contribution Types: NLP engineering experiment, Theory

Languages Studied: English

Submission Number: 6138