MorAI at QIAS 2025: Collaborative LLM via Voting and Retrieval-Augmented Generation for Solving Complex Inheritance Problems
Year: 2025
Type: conference-paper
Abstract: Collaborative approaches have proven effective in addressing complex problems, from human and socio-economic challenges to multiagent systems.These methods rely on the principle that combining perspectives enhances problem-solving.In this paper, we propose a collaborative large language models (LLM) framework to solve Islamic inheritance problems, which demand precise mathematical reasoning and strict adherence to legal rules for fair distribution among heirs.The system implements a collaborative voting mechanism involving multiple LLMs, namely ALLaM-7B-Instruct-preview, Deepseek-reasoner, and Gemini-2.5-Flash.Each independently answered multiple-choice inheritance questions.The final answer is determined by majority vote.To improve accuracy and domain grounding, we integrate Retrieval-Augmented Generation (RAG).A curated database of solved inheritance cases in JSON format is indexed using TF-IDF.For each query, the most similar cases are retrieved and appended as contextual information to the prompt before being submitted to the LLMs.Experimental results demonstrate that this collaborative RAG-enhanced framework outperforms individual LLMs.The ensemble achieved 88% accuracy, surpassing the best-performing single models: the finetuned ALLaM-7B-Instruct-preview (79.50%),Deepeek-reasoner (71.80%), and Gemini-2.5-Flash(83.50%).
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037601