Answers are grounded in this paper and cite their sources.
Debates represent a fundamental form of human communication, yet computational approaches to predicting debate outcome have largely overlooked a central factor – argument quality. Leveraging the language understanding capabilities of large language models (LLMs), we propose a novel method that predicts debate outcome using LLMs’ evaluation of argument quality. Adopting a functional perspective, we instruct an LLM to assess the quality of an argument along three dimensions – support for the speaker’s position, response to the opposition, and language style. The LLM’s evaluation, which accounts for both local and global context, is then used to predict debate outcome via machine learning. Using a dataset of face-to-face debates, we demonstrated that LLM-generated argument quality measures yielded significant performance gain, lifting the accuracy by 2.2% to 22.2% proportionally compared to various baselines while offering granular argument-level quality insights. Further analyses confirmed the reliability and validity of these measures, highlighting their multi-dimensional, context-sensitive nature. A comparison of high- and low-quality arguments further enhanced the interpretability and face validity of LLM-based argument assessment. Our work establishes argument quality as a pivotal construct for debate analysis, advances persuasion theories, sheds light on the situational contingencies in adversarial argumentation, and highlights the promise of LLMs as evaluative agents in computational argumentation.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: