Spreadsheets are the lingua franca of modern finance, underpinning valuation, forecasting, audit and many other core workflows. As large language model (LLM) agents advance toward automating knowledge work, spreadsheet construction emerges as a natural and critical testbed. Existing benchmarks mostly use fill‑in‑the‑blank tasks that require only minor edits to already‑complete sheets, missing the most time‑consuming steps such as modeling decisions, format design and sheet layout. MBABench-2.0 addresses this gap with 101 original spreadsheet‑building tasks; the hardest one is estimated to take a modeler with over three years of experience about 33 hours. Unlike code or math, where a unit test can verify the output, the quality of a financial spreadsheet—readability, maintainability, auditability—is not easily verifiable. We introduce 129 criteria across 12 categories to evaluate quality beyond mere accuracy. Experiments on more than 20 agents, 8 models and 4 harness types show that even frontier agents rarely produce spreadsheets that are both correct and meet our quality standards, with the best agent achieving only a 16% pass rate. The benchmark highlights a substantial gap in current agents' spreadsheet‑building capabilities.
Review