🤖 AI Summary
Current evaluations of financial large language models (LLMs) rely excessively on static benchmarks and lack comprehensive validation across the full system stack. This work proposes the first LLM full-stack verification framework tailored to financial scenarios, encompassing data, model, retrieval, generation, agent behavior, governance, and deployment layers, advocating for verification as an ongoing engineering practice. By introducing a multi-rater LLM-as-a-judge mechanism integrated with scoring rubrics, consistency checks, and auditability, the framework uncovers system failure modes that static benchmarks fail to capture. The study defines critical failure types and advances novel directions—including system-aware benchmarks, agent trajectory validation, rater alignment protocols, and lifecycle-oriented verification standards—thereby shifting the evaluation paradigm from score-driven metrics toward evidence-based readiness for real-world decision-making.
📝 Abstract
Large language models are increasingly deployed in financial applications that combine retrieval, proprietary data, tool use, orchestration logic, monitoring, and human escalation. Yet evaluation often remains model-centric: benchmark scores, task accuracy, or one-off qualitative reviews are treated as evidence of readiness. In financial settings, this is insufficient. We take the position that financial LLM systems should not be approved for production based on benchmark performance alone. They require system-level validation evidence across the application stack: data, model design, retrieval and generation performance, agent behavior, governance, and implementation. Drawing on industry experience validating GenAI applications in financial institutions, we outline a multi-layer validation view and explain why hybrid evaluation is necessary. We discuss where LLM-as-a-judge methods are useful and why they require controls such as multiple judges, rubrics, agreement, and auditability checks. We also highlight failure modes poorly captured by static benchmarks, including retrieval failures, unfaithful generation, tool misuse, escalation errors, and operational instability. Our position is that financial LLM validation should be an ongoing system discipline rather than a one-time model scoring exercise. Validation should produce decision-ready evidence, not only scores. We conclude with a research agenda for system-aware benchmarks, agent trace validation, judge alignment protocols, and lifecycle validation standards.