Bridging the Gap in Complex Claims: A Dual-Adaptive Multi-Agent Approach and Evaluation Benchmark for Chinese Insurance Reasoning

ACL ARR 2025 May Submission6960 Authors

Abstract: Insurance claims reasoning is a complex process that necessitates the integration of multi-source evidence while ensuring regulatory compliance and fairness. While Large Language Models (LLMs) show promise, existing evaluation systems lack the rigor for high-stakes scenarios with real economic and legal implications. To address this, we introduce **InsClaimQA**, the first clause-to-conclusion dataset for rigorous insurance claims reasoning. InsClaimQA features multi-difficulty grading, real-world derivations, expert annotations for legal traceability, and mandated explainable reasoning. To meet these high demands, we propose **DAMA**, a modular **D**ual-**A**daptive **M**ulti-**A**gent framework. DAMA uses specialized agents, context-aware routing, and a closed-loop quality control system to ensure reliable and transparent decisions. Evaluations confirm InsClaimQA's quality, with 98.7% accuracy and 0.96 RAGAs fidelity for explanations. DAMA significantly improves decision accuracy by 8.15% and reduces financial risk by 57.4%, proving its practical reliability in critical insurance applications. Code and data are available in the supplementary materials.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: financial/business NLP, legal NLP

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: Chinese, English

Keywords: Insurance Claims Reasoning​​, Multi-Agent Framework​​, Large Language Models, Dataset Benchmark, Dynamic Routing​​

Submission Number: 6960