Multi-Agent Dialectical Refinement for Enhanced Argument Classification

arXiv cs.CL / 3/31/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper addresses limitations in argument mining, where supervised methods require costly domain-specific fine-tuning and LLM-based approaches can misclassify ambiguous components like Claims versus Premises.
It proposes MAD-ACC (Multi-Agent Debate for Argument Component Classification), using a Proponent–Opponent–Judge multi-agent setup to perform dialectical refinement over uncertain text.
By forcing agents to defend opposing interpretations and then adjudicate, the method reduces structural ambiguity errors that often occur in single-agent self-correction and mitigates sycophancy.
Experiments on the UKP Student Essays corpus show MAD-ACC reaches a Macro F1 of 85.7%, outperforming single-agent reasoning baselines while remaining training-free for the domain.
The framework is positioned as more explainable than black-box classifiers because it produces human-readable debate transcripts that justify classification decisions.

Abstract

Argument Mining (AM) is a foundational technology for automated writing evaluation, yet traditional supervised approaches rely heavily on expensive, domain-specific fine-tuning. While Large Language Models (LLMs) offer a training-free alternative, they often struggle with structural ambiguity, failing to distinguish between similar components like Claims and Premises. Furthermore, single-agent self-correction mechanisms often suffer from sycophancy, where the model reinforces its own initial errors rather than critically evaluating them. We introduce MAD-ACC (Multi-Agent Debate for Argument Component Classification), a framework that leverages dialectical refinement to resolve classification uncertainty. MAD-ACC utilizes a Proponent-Opponent-Judge model where agents defend conflicting interpretations of ambiguous text, exposing logical nuances that single-agent models miss. Evaluation on the UKP Student Essays corpus demonstrates that MAD-ACC achieves a Macro F1 score of 85.7%, significantly outperforming single-agent reasoning baselines, without requiring domain-specific training. Additionally, unlike "black-box" classifiers, MAD-ACC's dialectical approach offers a transparent and explainable alternative by generating human-readable debate transcripts that explain the reasoning behind decisions.