CIFQA Records 95.54% Accuracy on Fixed-Deposit Queries via Deterministic Multi-Agent Routing
CIFQA decouples LLM language handling from exact financial computation through deterministic tools and specialized agents. On fixed-deposit queries the system records 95.54 percent accuracy on calculation-heavy items, exceeding both direct LLM baselines and larger models. The result isolates architectural separation, not scale, as the decisive factor for numerical correctness.
The arXiv paper introduces CIFQA as a five-agent pipeline separating query interpretation, routing, parameter extraction, computation planning, and response generation. Deterministic components handle exact rate lookup, tenure computation, rolling-year adjustment, and premature-withdrawal penalties. The framework was tested on a curated benchmark of fixed-deposit queries.
Direct LLM baselines supplied with complete formulas and rate cards still produced numerically incorrect answers. A 17B open-source model inside CIFQA outperformed larger frontier models given identical financial data. Ablation results isolate the deterministic tool layer as the primary driver of the accuracy gain.
The architecture demonstrates that numerical reliability in financial QA depends on explicit separation of symbolic execution from statistical generation rather than parameter count. Fixed-deposit logic serves as a verifiable test case for broader calculation-intensive domains.
Future deployments can substitute the fixed-deposit tool set with other rule-based financial engines without retraining the agent backbone.
AXIOM: CIFQA-style deterministic routing will appear in production fixed-income query systems within 18 months at greater than 92 percent accuracy on live customer logs.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.26114)
- [2]Supporting Source(https://arxiv.org/abs/2305.11738)