What it measures
Financial questions require answers grounded in company filings and supporting evidence.
How to read the result
Evaluate whether the cited evidence supports the calculation or conclusion, rather than judging a plausible financial answer in isolation.
Limits and comparison traps
Retrieval access, document coverage and answer grading matter. The public release is a sample of the broader benchmark and needs its own denominator.
Published evaluation
Source-reported percentages on the 150-question human-evaluated sample, rather than the full dataset. Retrieval configurations remain visible. Continuation lines in the printed table are not duplicate evaluations.
Where this fits in an evaluation suite
Synthetic research rewards do not establish live investment performance. Preserve transaction costs, held-out regimes, constraints, and the exact policy behind a reported result.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Source-reported percentages on the 150-question human-evaluated sample, rather than the full dataset. Retrieval configurations remain visible. Continuation lines in the printed table are not duplicate evaluations.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.
PRIMARY EVIDENCE & FURTHER READING
- FinanceBench: A New Benchmark for Financial Question Answering — published result table ↗
- Primary benchmark documentation or model evaluation report ↗
Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.