From Logical Mismatch to Logical Alignment: The Integration of the Logic of Evidential Reasoning and the Logic of Multi-Agent Collaboration Framework

Jiawen Zhang, Jinze Sang, Chenggong Zhao, Juan Li, Jianzhong Shi

Keywords: Logical Alignment; Evidential Reasoning; MAS

TL;DR: To achieve logic alignment, we introduce EMAr, a unified Multi-Agent argumentation framework whose design is deeply rooted in legal theory.

Abstract: Evidential reasoning is inherently complex, involving intricate subtasks such as legal entity recognition and charge classification, which provide fact-finding for Legal Judgment Prediction (LJP). Traditional Large Language Models (LLMs) often yield suboptimal performance due to the stringent legal reasoning logic required in the domain. While Multi-Agent Systems (MAS) are promising for complex tasks, existing MAS frameworks rely on generic coordination logic, which creates a detrimental "logic misalignment" with the specific requirements of legal evidence reasoning. To achieve logic alignment, we introduce EMAr, a unified Multi-Agent argumentation framework whose design is deeply rooted in legal theory. Its structure integrates adversarial debate and recursive reasoning to systematically align with the primary paradigms of evidence logic. We validated this framework on the LFPBench dataset. Our results show a significant improvement, achieving an Accuracy of 51.22% and a micro-F1 score of 51.61%. This validation confirms that integrating domain-specific evidence reasoning logic directly into the corresponding agent collaboration framework is crucial for substantially enhancing LJP performance.

Submission Number: 109