Research map / Multi-agent systems and games / LLM agents
CompeteAI: Understanding the Competition Dynamics of Large Language Model-based Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CompeteAI, a framework and GPT-4-driven virtual town simulation where two LLM restaurant agents compete for 50 LLM customer agents, and analyzes resulting market strategies, Matthew effect, and quality dynamics.
Problem
Competition between LLM-based agents is under-studied compared with cooperation, and realistic competitive simulations for studying market and sociological dynamics are missing.
Approach
A general framework for competitive LLM-agent studies (environment selection, setup, simulation, analysis) instantiated as a restaurant-vs-customer virtual town where restaurant agents manage operations via APIs and customer agents with diverse traits choose and review restaurants, enabling micro- and macro-level analysis against economic and sociological theories.
Key results
- Winner-take-all (>80% customers for one restaurant after day 5) occurred in 66.7% of single-customer runs versus 16.7% of group runs.
- Menu similarity between the two restaurants stayed around 36%, indicating a dynamic equilibrium of differentiation and imitation.
- Dish score average increased from Day 1 to Day 15 by 0.26 for R1 and 0.22 for R2; quality improved in at least one restaurant in 86.67% of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Grouping customers (group vs individual customers) | Restaurant competition with GPT-4 agents · Simulation (2 restaurants, 50 customers, 15 days) | Winner-take-all frequency: 66.7% single vs 16.7% group | Helped |
| Competition between restaurants | Restaurant agents' dish quality · Simulation (2 restaurants, 50 customers, 15 days) | Frequency average dish score improves over time | Helped |
| Competition between restaurants (imitation and differentiation) | Restaurant menus · Simulation (2 restaurants, 50 customers, 15 days) | Menu similarity between rivals: stays around 36% | No clear effect |
| Individual customers vs group customers | Customer decision making · Customer decision reasons | Share of reputation-based reasons: 29.42 individual vs 10.71 group | Mixed |
Surprising findings
- Group customers were more open to new dishes and less reputation-driven, which diminished winner-take-all dynamics.
- Behaviors align with economic and sociological theories, but the authors could not determine whether they stem from memorized theory in training data.
Limitations
- Small scale: only two restaurants and 50 customers due to GPT-4 API cost and limits.
- Text-only interactions; no multimodal inputs.
- Results specific to GPT-4-0613 and may change with API updates.
Experimental setup
Metrics: customer flow dish score menu similarity winner-take-all frequency decision reason distribution
Builds on: Generative Agents CAMEL AutoGen resource dependence theory
Closest papers by meaning
- Cooperation, Competition, and Maliciousness: LLM-Stakeholders Interactive Negotiation (NeurIPS 2024 · similarity 0.88)
- When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets (ICML 2026 · similarity 0.88)
- Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (NeurIPS 2025 · similarity 0.87)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
- Evaluating LLMs in Open-Source Games (NeurIPS 2025 · similarity 0.87)
- Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.