大型语言模型代理能否担任首席财务官？动态企业环境中的资源配置基准

Large language models (LLMs) have enabled agentic systems capable of reasoning, planning, and acting across complex tasks, yet their effectiveness in resource allocation under uncertainty remains uncertain. Unlike short-horizon reactive decisions, resource allocation requires committing scarce resources over time while balancing competing objectives and maintaining flexibility for future needs. This research introduces EnterpriseArena, the first benchmark for evaluating agents on long-horizon enterprise resource allocation. It simulates CFO-style decision-making over a 132-month period, integrating firm-level financial data, anonymized business documents, macroeconomic and industry signals, and expert-validated operating rules. The environment is partially observable, designed to assess LLM agents' performance in dynamic enterprise contexts.