Herculean：面向金融智能的智能体基准测试

# Herculean：金融智能的智能体基准
来源：https://arxiv.org/abs/2605.14355
作者：Xueqing Peng (https://arxiv.org/search/cs?searchtype=author&query=Peng,+X),Zhuohan Xie (https://arxiv.org/search/cs?searchtype=author&query=Xie,+Z),Yupeng Cao (https://arxiv.org/search/cs?searchtype=author&query=Cao,+Y),Haohang Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+H),Lingfei Qian (https://arxiv.org/search/cs?searchtype=author&query=Qian,+L),Yan Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+Y),Vincent Jim Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+V+J),Huan He (https://arxiv.org/search/cs?searchtype=author&query=He,+H),Xuguang Ai (https://arxiv.org/search/cs?searchtype=author&query=Ai,+X),Linhai Ma (https://arxiv.org/search/cs?searchtype=author&query=Ma,+L),Ruoyu Xiang (https://arxiv.org/search/cs?searchtype=author&query=Xiang,+R),Yueru He (https://arxiv.org/search/cs?searchtype=author&query=He,+Y),Yi Han (https://arxiv.org/search/cs?searchtype=author&query=Han,+Y),Shuyao Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+S),Yuqing Guo (https://arxiv.org/search/cs?searchtype=author&query=Guo,+Y),Mingyang Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+M),Yilun Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+Y),Youzhong Dong (https://arxiv.org/search/cs?searchtype=author&query=Dong,+Y),Xiaoyu Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+X),Yankai Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+Y),Ye Yuan (https://arxiv.org/search/cs?searchtype=author&query=Yuan,+Y),Qiyuan Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+Q),Fuyuan Lyu (https://arxiv.org/search/cs?searchtype=author&query=Lyu,+F),Haolun Wu (https://arxiv.org/search/cs?searchtype=author&query=Wu,+H),Yonghan Yang (https://arxiv.org/search/cs?searchtype=author&query=Yang,+Y),Zichen Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+Z),Yuyang Dai (https://arxiv.org/search/cs?searchtype=author&query=Dai,+Y),Fan Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+F),Rania Elbadry (https://arxiv.org/search/cs?searchtype=author&query=Elbadry,+R),Ayesha Gull (https://arxiv.org/search/cs?searchtype=author&query=Gull,+A),Muhammad Usman Safder (https://arxiv.org/search/cs?searchtype=author&query=Safder,+M+U),Nuo Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+N),Fengbin Zhu (https://arxiv.org/search/cs?searchtype=author&query=Zhu,+F),Tianshi Cai (https://arxiv.org/search/cs?searchtype=author&query=Cai,+T),Zimu Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+Z),Polydoros Giannouris (https://arxiv.org/search/cs?searchtype=author&query=Giannouris,+P),Yuechen Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+Y),Zhiwei Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+Z),Mohsinul Kabir (https://arxiv.org/search/cs?searchtype=author&query=Kabir,+M),Yuyan Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+Y),Yixiang Zheng (https://arxiv.org/search/cs?searchtype=author&query=Zheng,+Y),Yangyang Yu (https://arxiv.org/search/cs?searchtype=author&query=Yu,+Y),Weijin Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+W),Wenbo Cao (https://arxiv.org/search/cs?searchtype=author&query=Cao,+W),Anke Xu (https://arxiv.org/search/cs?searchtype=author&query=Xu,+A),Peng Lu (https://arxiv.org/search/cs?searchtype=author&query=Lu,+P),Jerry Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+J),Fengran Mo (https://arxiv.org/search/cs?searchtype=author&query=Mo,+F),Mingquan Lin (https://arxiv.org/search/cs?searchtype=author&query=Lin,+M),Prayag Tiwari (https://arxiv.org/search/cs?searchtype=author&query=Tiwari,+P),Yijia Zhao (https://arxiv.org/search/cs?searchtype=author&query=Zhao,+Y),Victor Gutierrez Basulto (https://arxiv.org/search/cs?searchtype=author&query=Basulto,+V+G),Xiao\-Yang Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+X),Kaleb E Smith (https://arxiv.org/search/cs?searchtype=author&query=Smith,+K+E),Jiahuan Pei (https://arxiv.org/search/cs?searchtype=author&query=Pei,+J),Arman Cohan (https://arxiv.org/search/cs?searchtype=author&query=Cohan,+A),Jimin Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+J),Yuehua Tang (https://arxiv.org/search/cs?searchtype=author&query=Tang,+Y),Alejandro Lopez\-Lira (https://arxiv.org/search/cs?searchtype=author&query=Lopez-Lira,+A),Xi Chen (https://arxiv.org/search/cs?searchtype=author&query=Chen,+X),Xue Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+X),Junichi Tsujii (https://arxiv.org/search/cs?searchtype=author&query=Tsujii,+J),Jian\-Yun Nie (https://arxiv.org/search/cs?searchtype=author&query=Nie,+J),Sophia Ananiadou (https://arxiv.org/search/cs?searchtype=author&query=Ananiadou,+S)

View PDF (https://arxiv.org/pdf/2605.14355)

> **摘要：**随着AI智能体的进步，核心问题已不再是它们能否解决孤立的、定义明确的金融任务，而是它们能否可靠地执行金融专业工作。现有的金融基准仅提供了这种能力的部分视角，因为它们主要评估静态能力，如问答、检索、摘要和分类。我们引入了Herculean，这是首个面向智能体金融智能的技能基准，涵盖了四个具有代表性的工作流，包括交易（Trading）、对冲（Hedging）、市场洞察（Market Insights）和审计（Auditing）。每个工作流都实例化为一个基于MCP的标准技能环境，拥有自己的工具、交互动态、约束条件和成功标准，从而能够对异构智能体系统进行一致的端到端评估。在前沿智能体的测试中，我们发现智能体在交易和市场洞察方面表现相对较好，但在对冲和审计方面则明显吃力，后两者需要长期协调、状态一致性和结构化验证。总体而言，我们的结果揭示了当前智能体在高风险金融工作流中，将金融推理转化为可靠工作流执行方面存在关键差距。

## 提交历史

来自：Xueqing Peng Dr [view email](https://arxiv.org/show-email/92e15dc2/2605.14355) **\[v1\]** 2026年5月14日 星期四 04:30:49 UTC (3,734 KB)