arxiv_cl 90% Match Research Paper AI Safety Researchers,ML Engineers,AI Ethicists,Researchers in Multi-Agent Systems 20 hours ago

I Want to Break Free! Persuasion and Anti-Social Behavior of LLMs in Multi-Agent Settings with Social Hierarchy

reinforcement-learning › multi-agent

📄 Abstract

Abstract: As LLM-based agents become increasingly autonomous and will more freely interact with each other, studying the interplay among them becomes crucial to anticipate emergent phenomena and potential risks. In this work, we provide an in-depth analysis of the interactions among agents within a simulated hierarchical social environment, drawing inspiration from the Stanford Prison Experiment. Leveraging 2,400 conversations across six LLMs (i.e., LLama3, Orca2, Command-r, Mixtral, Mistral2, and gpt4.1) and 240 experimental scenarios, we analyze persuasion and anti-social behavior between a guard and a prisoner agent with differing objectives. We first document model-specific conversational failures in this multi-agent power dynamic context, thereby narrowing our analytic sample to 1,600 conversations. Among models demonstrating successful interaction, we find that goal setting significantly influences persuasiveness but not anti-social behavior. Moreover, agent personas, especially the guard's, substantially impact both successful persuasion by the prisoner and the manifestation of anti-social actions. Notably, we observe the emergence of anti-social conduct even in absence of explicit negative personality prompts. These results have important implications for the development of interactive LLM agents and the ongoing discussion of their societal impact.

Key Contributions

This work provides an in-depth analysis of persuasion and anti-social behavior among LLM agents in a simulated hierarchical social environment, inspired by the Stanford Prison Experiment. By analyzing 2,400 conversations across six LLMs, it reveals model-specific failures and finds that goal setting influences persuasiveness but not anti-social behavior, offering crucial insights into the emergent dynamics and potential risks of autonomous LLM interactions.

Business Value

Crucial for developing safer and more predictable AI systems, especially in multi-agent scenarios, by understanding potential negative emergent behaviors and designing mitigation strategies.

Paper Metadata

Innovation Type

Empirical study and analysis

Deployment Feasibility

High, as the findings inform the design and safety protocols for future multi-agent AI systems.

Limitations Addressed

The lack of understanding regarding the emergent behaviors, risks, and social dynamics of increasingly autonomous LLM agents interacting with each other.

Performance Gains

Provides qualitative and quantitative insights into LLM agent behavior, not direct performance gains on a task.

Technical Tags

multi-agent systemsLLM agentssocial hierarchypersuasionanti-social behaviorsimulated environmentconversational analysisemergent phenomena

Research Topics

AI SafetyMulti-Agent SystemsHuman-AI InteractionAI EthicsLarge Language Models

Methods & Architectures

Simulated social environmentLLM agent interaction analysisConversational data analysisExperimental scenario design LLaMA3Orca2Command-rMixtralMistral2GPT-4.1

Applications & Tasks

AI Safety Research Multi-agent AI development Simulated social interactions Understanding emergent behaviors in LLM agentsAnalyzing power dynamics and social interactionsIdentifying risks in autonomous agent systems Analyzing persuasion strategiesDetecting anti-social behaviorStudying social hierarchy effects on LLM interactions

Related Fields

AI SafetyMulti-Agent SystemsSocial PsychologyArtificial Intelligence EthicsNatural Language Processing

Keywords

LLM agentsmulti-agentsocial hierarchypersuasionanti-social behaviorsimulated environmentStanford Prison Experimentemergent behaviorAI safetyconversational AILLaMA3Orca2Mixtral

Academic Context

#AI Safety#Multi-Agent Systems#Human-AI Interaction#AI Ethics#Large Language Models

Commercial Potential

Target Industries

AI DevelopmentRoboticsGamingSimulation

Use Case Examples

Designing AI agents for complex simulations where social dynamics are important.Developing safety guidelines for autonomous agent interactions.Understanding potential manipulation or deception by AI systems.

Competitive Edge

This study provides a unique empirical investigation into the social dynamics and potential risks of LLM agents in a controlled, hierarchical setting, offering insights not typically found in standard LLM evaluations.

Market Opportunity

Growing importance of AI safety and multi-agent systems research.

Revenue Models

Research publicationsinsights for AI safety product development.

Resource Requirements

Compute Needs

High, for running multiple LLMs in conversational simulations.

Data Requirements

Simulated social environment setup, multiple LLM agents.

Deployment Constraints

Findings are based on simulations and may not perfectly generalize to all real-world scenarios.,Ethical considerations in simulating human experiments.

Scalability

The experimental setup can be scaled by increasing the number of agents, scenarios, or LLM models.

Regulatory Considerations

Ethical implications of simulating social experiments.Responsible AI development.

Production Readiness

Maturity Level

Research

View Full Paper Back to Papers