CIFQA：确定性工具接地多智能体LLM框架用于金融问答
原标题：CIFQA: A Deterministic Tool-Grounded Multi-Agent LLM Framework for Financial Query Answering
AI 摘要
印度理工学院焦特布尔分校的研究团队提出了CIFQA，一个确定性的工具接地多智能体LLM框架，用于计算密集型金融问答。该框架将语言理解与数值执行分离，通过LLM智能体负责查询理解、路由、参数提取等，而确定性Python工具执行金融计算。在定期存款查询基准上，CIFQA达到95.54%的计算密集型查询准确率和90.87%的总体准确率，显著优于直接LLM基线，且17B开源模型在框架内表现优于GPT-5.3、Gemini 3等更大模型，表明架构设计比模型规模更重要。
正文节选
CIFQA: A Deterministic Tool-Grounded Multi-Agent LLM Framework for Financial Query Answering Kunjesh Parekh1∗ Dr. Anil Kumar Tiwari1 Dr. Divya Saxena1 1School of Artificial Intelligence and Data Science, Indian Institute of Technology Jodhpur, Rajasthan 342030, India ∗Corresponding author: P23ai0003@iitj.ac.in Co-author: akt@iitj.ac.in Co-author: divyasaxena@iitj.ac.in Abstract Calculation-intensive financial question answering requires not only language understanding, but also exact e