FinReportBench: Benchmarking LLMs for Institution-Grade Financial Reports

A novel benchmark named FinReportBench has been launched to assess and enhance the production of institution-grade financial reports by large language models (LLMs). Researchers, noticing persistent shortcomings in report identity, institutional elements, source discipline, and visual presentation through expert evaluations, created this benchmark. They formulated a 35-item rubric utilizing expert partial orders, multimodal evidence, and decision boundary audits, focusing on deliverability, report identity, and institutional thoroughness. The benchmark comprises 10,000 balanced financial research records in Chinese and English, leading to 244 bilingual tasks across three research subjects and two input levels. Each task differentiates the public query, reconstructed research path, and concealed source packet. Three independent judging groups replicate the expert partial order at nearly ceiling rates, confirming the benchmark's reliability. This research, which is crucial for the financial sector, is documented in a paper on arXiv (arXiv:2608.04374), submitted for cross-announcement.