Researchers have developed a benchmark to evaluate the performance of large language model (LLM) agents on end-to-end spreadsheet tasks in finance. The MBABench evaluates over 18 agents' ability to construct spreadsheets from scratch and perform complex financial workflows such as modeling and scenario analysis. However, the results show that even top-performing agents fall short of professional standards, particularly when faced with high levels of complexity. --- Why it matters: This matters because current LLM agents are not yet reliable enough for real-world finance applications, where accuracy and quality are crucial. Improving their performance on end-to-end spreadsheet tasks is essential for enterprise adoption. Source: https://arxiv.org/abs/2605.22664