Title: BigFinanceBench: A Workflow-Grounded Benchmark for Financial-Research Agents

Abstract:

In the realm of financial research, the credibility of an answer hinges on its auditability. For a conclusion to be decision-relevant, another analyst must be able to verify the underlying process: identifying the selected sources, specifying the accounting definitions and time periods, outlining the assumptions, and tracing the calculation methodology. Current financial benchmarks primarily focus on isolated subskills or final outputs, thereby failing to adequately measure the auditable derivation itself.

To address this gap, we present BigFinanceBench, a comprehensive benchmark comprising 928 expert-curated, open-ended financial research tasks. Each task is paired with a ground-truth reference answer and a point-weighted rubric that breaks down the derivation into independently verifiable steps. Distinctively, BigFinanceBench is workflow-grounded, assessing the entire derivation process rather than just the final result. Spanning 36,241 rubric points, the benchmark enables partial-credit scoring and facilitates the localization of failures within the analyst’s workflow.

Our evaluation of ten leading frontier and open-weight agents reveals significant room for improvement. The top-performing system achieved only a 58.8% rubric score, demonstrating that while final-answer accuracy serves as a useful metric, it is a lossy proxy for true derivation quality. Furthermore, the analysis highlights that model capabilities are not uniform across different financial workflows.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC