AgentFinVQA: A Deployable Multi-Agent Pipeline for Auditable Financial Chart QA

AgentFinVQA addresses the dual challenge of accuracy and auditability in financial chart analysis. Existing chart-QA models are often opaque, "black-box" systems that struggle with hallucinations and require proprietary API access, which is unsuitable for institutions with strict data residency requirements. The authors propose a multi-agent pipeline that decomposes complex chart queries into a sequence of specialized, auditable stages: planning, OCR extraction, legend grounding, deterministic color-area measurement, visual inspection, and final verification. Every step of this process is recorded in a Model Evaluation Packet (MEP), providing a transparent trace of how the final answer was derived.

The pipeline demonstrates significant performance improvements over standard zero-shot baselines. On the FinMME benchmark, AgentFinVQA achieved a +7.68 percentage point increase using a proprietary backbone (Gemini-3 Flash) and a +4.84 percentage point increase using an open-weights model (Qwen3.6-27B-FP8) served locally on a single A100 GPU. The system's verifier stage acts as a confidence signal, allowing for human-in-the-loop routing: answers marked as 'confirmed' by the verifier are significantly more accurate than those marked as 'revised,' enabling practitioners to focus their review efforts on the most error-prone outputs.

This research proves that high-accuracy, auditable financial chart QA is achievable without relying on external model providers. By utilizing a prompting-only, agentic architecture, the authors demonstrate that institutions can maintain data privacy while achieving performance gains comparable to proprietary systems. The inclusion of the MEP audit trail directly addresses regulatory needs for transparency, making the system a practical tool for high-stakes financial environments where trust and explainability are as important as raw accuracy.