CLAIR-Fin uses 9 agents to slash AI hallucinations in finance Q&A

A team of researchers from Bangladesh and Saudi Arabia has developed CLAIR-Fin, a novel adversarial multi-agent framework designed to combat hallucinations in AI-generated financial answers. The system breaks down complex financial questions into atomic claims stored in a Financial Claim Ledger, where each claim undergoes rigorous verification through four key innovations: Asymmetric Evidence Authority (which evaluates evidence based on claim type), Chain-of-Custody Verification (checking grounding during handoffs), Adaptive Rebuttal Cycle (scaling debate depth based on contention), and a terminal entailment audit with continuous Hallucination Risk Index.

When tested on BB-FinQA-X—a custom 500-question cross-modal financial evaluation set derived from Bangladesh Bank Annual Reports—CLAIR-Fin achieved an 88.9% faithfulness score, a 14% improvement over a single-pass retrieval baseline (78.0%) and outperforming stronger baselines like HyDE and Graph-RAG (≤87.4%). Notably, the system abstained from answering 5.4% of questions when evidence was insufficient, avoiding unsupported responses that plague traditional pipelines. This represents a significant step forward in reliable financial AI by addressing modality disagreements and inter-agent errors that typically slip through undetected until the final output.