CIFQA Uses Deterministic Multi-Agent LLMs for Reliable Financial Answers

Summary

Researchers introduce CIFQA, a deterministic tool-grounded multi-agent framework for calculation-intensive financial question answering. Specialized agents interpret queries, extract parameters, plan computations, and generate responses, while Python tools execute exact calculations and rules. CIFQA reached 95.54% accuracy on calculation-intensive fixed-deposit queries and 90.87% overall, outperforming direct LLM baselines. A 17B open-source backbone also exceeded substantially larger frontier models, highlighting the importance of architecture for numerical reliability.