EnterpriseArena — can LLM agents be CFOs?
Collection
EnterpriseArena / CFO-Env (arXiv:2603.23638). • 2 items • Updated
Request access to EnterpriseArena (CFO-Env)
EnterpriseArena is released by The Fin AI for non-commercial research. Access is granted automatically after you complete this short form.
This environment is licensed under CC BY-NC-ND 4.0: non-commercial use only, no redistribution of modified versions. The company in the environment is fictional and its documents are anonymized. By accessing it you agree to these terms and to cite the paper below in any resulting publication.
Log in or Sign Up to review the conditions and access this dataset content.
EnterpriseArena (CFO-Env)
📄 Paper · 🌐 The Fin AI
EnterpriseArena is a benchmark for evaluating LLM agents on long-horizon enterprise resource allocation. The agent acts as the CFO of a fictional company in a 132-month enterprise simulator and must keep the company solvent while growing it — committing scarce cash, raising debt or equity, and deciding when to spend budget on gathering information. Introduced in Can LLM Agents Be CFOs? Benchmarking Long-Horizon Resource Allocation in an Uncertain Enterprise Environment (arXiv:2603.23638).
This repository contains the environment's starting state and the data that drives the simulator.
Why it is hard
- Long horizon: decisions compound over 132 monthly steps.
- Partial observability: the true state is only revealed through organizational tools that cost budget, so the agent must trade information against resources.
- Uncertainty: fundraising outcomes and macro conditions are stochastic.
In the paper's experiments with eleven advanced LLMs, only 16% of runs survive the full horizon, and larger models do not reliably outperform smaller ones.
The simulated company
Company XYZ is a fictional consumer-focused fintech in digital lending: it originates loans from its own balance sheet, earns interest income on the loan portfolio, and bears the full credit risk. Its starting position (t = 0) is described by board materials, financial statements and operating contracts.
Repository Layout
| Path | Content |
|---|---|
enterprise_env_t0/t0_board_materials/ |
company_overview.txt, cap_table.txt, governance_approvals.txt |
enterprise_env_t0/t0_financials/ |
Opening balance sheet, income statement and cash-flow statement (.xlsx) |
enterprise_env_t0/to_operations_payments/ |
Anonymized operating agreements (customer support, fraud monitoring, software development, hosting, sales & marketing) (.docx) |
data_collect/combined_economic_data_2015_2025.csv |
132 monthly rows of macro and market signals: GDP, CPI, UNRATE, SOFR, FEDFUNDS, Baa yield, 2/5/10/30-year Treasury yields, VIX, P/E and P/S ratios, gross and EBITDA margins, user-growth rates |
data_collect/fundraising_success_probabilities_2015_2025.csv |
Monthly probability that a debt or equity raise succeeds (P_debt, P_equity, plus adjusted variants) |
data_collect/simulator_logic.md |
Full specification of the simulator: ledger and state tracking, financial-statement generation, fundraising simulation, stochastic evaluation design |
Quick Start
from huggingface_hub import snapshot_download
import pandas as pd
path = snapshot_download("TheFinAI/CFO-Env", repo_type="dataset")
macro = pd.read_csv(f"{path}/data_collect/combined_economic_data_2015_2025.csv")
funding = pd.read_csv(f"{path}/data_collect/fundraising_success_probabilities_2015_2025.csv")
print(macro.shape, funding.shape)
print(open(f"{path}/enterprise_env_t0/t0_board_materials/company_overview.txt").read()[:500])
The simulator rules the environment implements are documented in data_collect/simulator_logic.md.
Intended Use & Limitations
- Intended use: research on LLM agents' planning and resource allocation under uncertainty.
- The company and its documents are fictional/anonymized; results do not transfer directly to real firms.
- The environment models one business type (balance-sheet consumer lending) over 2015–2025 conditions.
- Not investment or corporate-finance advice.
License
CC BY-NC-ND 4.0 — attribution, non-commercial use only, no derivatives.
Citation
@misc{han2026llmagentscfosbenchmarking,
title={Can LLM Agents Be CFOs? Benchmarking Long-Horizon Resource Allocation in an Uncertain Enterprise Environment},
author={Yi Han and Yan Wang and Lingfei Qian and Haohang Li and Yupeng Cao and Yueru He and Xueqing Peng and Nanhan Shen and Yitao Xu and Yankai Chen and Dongji Feng and Jimin Huang and Xue Liu and Jian-Yun Nie and Sophia Ananiadou},
year={2026},
eprint={2603.23638},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2603.23638},
}
- Downloads last month
- 25