跳转至

文章背景与核心概要

本文探讨了大语言模型（LLM）在跨越较长时段的多步骤、战略性商业决策中的表现能力。尽管LLM通常擅长短期模式识别和自然语言任务，但它们在连贯的长期战略规划方面的熟练程度在很大程度上仍未得到充分探索。

为了填补这一空白，作者引入了一个可复现的、开源的管理模拟器，用于对五种领先的AI模型——Gemini、ChatGPT、Meta AI、Mistral AI和Grok进行基准测试。在该模拟中，每个LLM在12个月的周期内扮演模拟零售公司的经理。通过使用结构化的月度业务报告，模型做出关键的运营决策，包括定价、订单规模、营销预算、招聘、融资和预测。性能评估结合了定量财务指标（利润、收入、市场份额）以及对战略连贯性和适应性的定性评估。

AI Playing Business Games: Benchmarking Large Language Models on Managerial Decision-Making in Dynamic Simulations

AI Playing Business Games: Benchmarking Large Language Models on Managerial Decision-Making in Dynamic Simulations

Summary

Summary

本文探讨了大语言模型（LLM）在跨越较长时段的多步骤、战略性商业决策中的表现能力。尽管LLM通常擅长短期模式识别和自然语言任务，但它们在连贯的长期战略规划方面的熟练程度在很大程度上仍未得到充分探索。

This paper investigates the capabilities of Large Language Models (LLMs) in multi-step, strategic business decision-making over extended time horizons. While LLMs typically excel at short-term pattern recognition and natural language tasks, their proficiency in coherent, long-term strategic planning remains largely unexplored.

为了填补这一空白，作者引入了一个可复现的、开源的管理模拟器，用于对五种领先的AI模型——Gemini、ChatGPT、Meta AI、Mistral AI和Grok进行基准测试。在该模拟中，每个LLM在12个月的周期内扮演模拟零售公司的经理。通过使用结构化的月度业务报告，模型做出关键的运营决策，包括定价、订单规模、营销预算、招聘、融资和预测。性能评估结合了定量财务指标（利润、收入、市场份额）以及对战略连贯性和适应性的定性评估。

To address this gap, the author introduces a reproducible, open-access management simulator to benchmark five leading AI models—Gemini, ChatGPT, Meta AI, Mistral AI, and Grok. In this simulation, each LLM acts as the manager of a simulated retail company over a 12-month period. Using structured monthly business reports, the models make critical operational decisions, including pricing, order sizing, marketing budgets, hiring, financing, and forecasting. Performance is evaluated using quantitative financial metrics (profit, revenue, market share) alongside qualitative assessments of strategic coherence and adaptability.

Metadata

Metadata

  • arXiv ID: arXiv:2509.26331 [cs.AI]
  • Authors: Berdymyrat Ovezmyradov
  • Submitted: September 30, 2025 (Last revised: August 5, 2026)
  • Primary Subject: Artificial Intelligence (cs.AI)
  • ACM Classification: I.2.1
  • Document Details: 34 pages, 7 figures, 3 tables
  • arXiv ID: arXiv:2509.26331 [cs.AI]
  • Authors: Berdymyrat Ovezmyradov
  • Submitted: September 30, 2025 (Last revised: August 5, 2026)
  • Primary Subject: Artificial Intelligence (cs.AI)
  • ACM Classification: I.2.1
  • Document Details: 34 pages, 7 figures, 3 tables

Methodology & Experimental Design

Methodology & Experimental Design

  • Environment: A transparent, month-by-month spreadsheet-based management simulation of a retail company.
  • Duration: 12 simulated months per evaluation cycle.
  • Input Data: Structured prompts containing complete business reports from the previous period.
  • Decision Variables:
  • Pricing and Sales Forecasts
  • Order Size and Inventory Management
  • Marketing Budget and Training Expenses
  • R&D Expenses
  • Hiring and Dismissals
  • Loans and Income Forecasts
  • Evaluation Metrics:
  • Quantitative: Profit, revenue, market share, and key performance indicators (KPIs).
  • Qualitative: Strategic coherence, market adaptability, and the logical rationale provided for decisions.
  • Environment: A transparent, month-by-month spreadsheet-based management simulation of a retail company.
  • Duration: 12 simulated months per evaluation cycle.
  • Input Data: Structured prompts containing complete business reports from the previous period.
  • Decision Variables:
  • Pricing and Sales Forecasts
  • Order Size and Inventory Management
  • Marketing Budget and Training Expenses
  • R&D Expenses
  • Hiring and Dismissals
  • Loans and Income Forecasts
  • Evaluation Metrics:
  • Quantitative: Profit, revenue, market share, and key performance indicators (KPIs).
  • Qualitative: Strategic coherence, market adaptability, and the logical rationale provided for decisions.

Full-Text & Resources

Full-Text & Resources