ทันเอไอ

CIFQA uses a multi-agent system to reduce errors in financial calculations

Researchers propose a framework in which LLMs interpret questions while rule-based tools handle calculations, reporting an overall accuracy of 90.87%.

Kunjesh Parekh, Anil Kumar Tiwari, and Divya Saxena published the CIFQA paper on arXiv under number 2608.26114. It aims to address a weakness of Large Language Models, or LLMs, in answering financial questions that require multistep calculations: models can generate answers that sound plausible even when the figures are wrong.

CIFQA stands for Calculation-Intensive Financial Query Answering. It is a Multi-agent framework built around separating “language understanding” from “numerical processing,” instead of having an LLM interpret the problem, select formulas, perform calculations, and produce the final answer on its own.

The system includes specialized agents for interpreting questions, selecting workflows, extracting parameters, planning calculation steps, and composing answers. Calculations and rule-based conditions are delegated to Python tools that operate deterministically, producing results through the same steps when given the same inputs. This reduces reliance on the model’s numerical guesswork.

The researchers tested the system on Fixed Deposit questions involving interest rates, terms, numerical formulas, and various conditions. They reported that CIFQA achieved 95.54% accuracy on calculation-intensive questions and 90.87% overall accuracy, outperforming large general-purpose LLMs in the same evaluation.

The results suggest that in tasks where even minor errors can change a financial answer, designing an architecture that lets models systematically call calculation tools may matter more than simply increasing model size. However, all figures were reported by the researchers from Fixed Deposit tasks, so it is too early to conclude that CIFQA can maintain the same accuracy for loans, taxes, investments, or country-specific financial rules.