Enhancing Conflict Resolution in Language Models via Abstract Argumentation
Zhaoqun Li, Xiaotong Fang, Chen Chen, Mengze Li, Beishui Liao
TL;DR
This work tackles the challenge of conflict resolution in large language models by marrying abstract argumentation with symbolic computation. It introduces a large, diverse AAF benchmark with algorithm explanations and trains LLMs (Llama-3 and Qwen2.5) via supervised fine-tuning and RLHF to compute grounded ($\text{grd}$) and complete ($\text{com}$) extensions, comparing against chain-of-thought baselines. Key findings show process explanations significantly improve generalization and transparency, enabling robust conflict resolution while maintaining explainability through self-generated illustrations. The approach advances practical decision-making tasks that require defeasible reasoning and paves the way for dynamic, probabilistic extensions and real-world grounding of argumentation frameworks.
Abstract
In recent years, large language models (LLMs) have made significant advancements in developing human-like and engaging dialogue systems. However, in tasks such as consensus-building and persuasion, LLMs often struggle to resolve conflicts arising from incomplete or inconsistent information, revealing their limitations in real-world applications. Given these limitations, abstract argumentation, a specialized logical framework designed to resolve conflicts and inconsistencies, becomes particularly relevant. In this paper, we aim to enhance the conflict-solving capabilities of LLMs by leveraging formal abstract argumentation, integrating language model learning with symbolic computation. To achieve this, we develop and curate a dataset comprising diverse abstract argumentation frameworks, accompanied by detailed explanations of the argument acceptability computation process. Subsequently, we fine-tune LLMs on this dataset, focusing on abstract conflict resolution tasks. As a comparative baseline, LLMs are also evaluated using a chain-of-thought approach, however, they fail to solve the conflict-based arguments effectively. Our experiments demonstrate that process explanations play a crucial role in learning. Models trained with explanations exhibit superior generalization accuracy compared to those trained solely on question-answer pairs. Furthermore, leveraging LLMs' self-explanation capabilities, our approach provides detailed illustrations that mitigate the lack of transparency typically associated with neural networks.