Herculean：面向金融智能的智能体基准
人工智能
2026-05-22 v2 计算与语言
摘要
随着 AI 智能体的改进，核心问题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，这是首个涵盖四个代表性工作流（包括交易、对冲、市场洞察和审计）的智能体金融智能技能基准。每个工作流都被实例化为一个标准化的基于 MCP 的技能环境，具有自己的工具、交互动态、约束和成功标准，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存在显著困难，这些工作流对长程协调、状态一致性和结构化验证至关重要。总体而言，我们的结果指出了当前智能体在将金融推理转化为高风险金融工作流中可靠工作流执行方面存在关键差距。
引用
@article{arxiv.2605.14355,
title = {Herculean: An Agentic Benchmark for Financial Intelligence},
author = {Xueqing Peng and Zhuohan Xie and Yupeng Cao and Haohang Li and Lingfei Qian and Yan Wang and Vincent Jim Zhang and Huan He and Xuguang Ai and Linhai Ma and Ruoyu Xiang and Yueru He and Yi Han and Shuyao Wang and Yuqing Guo and Mingyang Jiang and Yilun Zhao and Youzhong Dong and Xiaoyu Wang and Yankai Chen and Ye Yuan and Qiyuan Zhang and Fuyuan Lyu and Haolun Wu and Yonghan Yang and Zichen Zhao and Yuyang Dai and Fan Zhang and Rania Elbadry and Ayesha Gull and Muhammad Usman Safder and Nuo Chen and Fengbin Zhu and Tianshi Cai and Zimu Wang and Polydoros Giannouris and Yuechen Jiang and Zhiwei Liu and Mohsinul Kabir and Yuyan Wang and Yixiang Zheng and Yangyang Yu and Weijin Liu and Wenbo Cao and Anke Xu and Peng Lu and Jerry Huang and Mingquan Lin and Prayag Tiwari and Yijia Zhao and Victor Gutierrez Basulto and Xiao-Yang Liu and Kaleb E Smith and Jiahuan Pei and Arman Cohan and Jimin Huang and Yuehua Tang and Alejandro Lopez-Lira and Xi Chen and Xue Liu and Junichi Tsujii and Jian-Yun Nie and Sophia Ananiadou},
journal= {arXiv preprint arXiv:2605.14355},
year = {2026}
}