LLM-Based Agent Society Investigation: Collaboration and Confrontation in Avalon Gameplay

TL;DR

<3-5 sentence high-level summary> The paper investigates how large language model–based agents behave in Avalon, a social deduction game, by deploying a multi-module framework that includes memory storage and summarization, analysis, planning, action, response generation, and experiential learning. Agents are prompted with system roles and guided through memory-augmented reasoning and learning from experience, with ablation studies showing the importance of each module. Experimental results demonstrate strong performance against a baseline and reveal rich social behaviors such as leadership, persuasion, camouflage, and teamwork/confrontation, along with quantifiable metrics like WR, QER, FVR, and LAR. The work provides actionable insights into designing adaptive AI agents for dynamic social interactions and makes the code publicly available for replication and extension.

Abstract

This paper explores the open research problem of understanding the social behaviors of LLM-based agents. Using Avalon as a testbed, we employ system prompts to guide LLM agents in gameplay. While previous studies have touched on gameplay with LLM agents, research on their social behaviors is lacking. We propose a novel framework, tailored for Avalon, features a multi-agent system facilitating efficient communication and interaction. We evaluate its performance based on game success and analyze LLM agents' social behaviors. Results affirm the framework's effectiveness in creating adaptive agents and suggest LLM-based agents' potential in navigating dynamic social interactions. By examining collaboration and confrontation behaviors, we offer insights into this field's research and applications. Our code is publicly available at https://github.com/3DAgentWorld/LLM-Game-Agent.

Figures (12)

• Figure 1: Our framework has six modules: summary, analysis, planning, action, response, and experiential learning. This design follows human thinking, helps LLM agents play Avalon effectively, and reveals their social behaviors.
• Figure 2: (a): Comparison of the engaging quests rate when playing evil side. Higher engaging quests rate means more opportunities for the player to influence the outcome of the game. (b): Comparison of the failure vote rate when playing evil side. Baseline is worse.
• Figure 3: (a): The leadership behavior. Players with higher Leader Approval Rate get more agreements from other players when deciding a quest team. (b) and (c): The persuasion behavior. Self-recommendation Rate: players with higher Self-recommendation Rate are more will to engage in quests. Self-recommendation Success Rate: players more likely to gain the trust of other players has higher Self-recommendation Success Rate.
• Figure 4: The camouflage behavior when playing different roles: at first round of each game, the distribution of the players choose Self-Disclosure, Camouflage or Withholding Identity.
• Figure 5: The teamwork and confrontation behaviors when playing different roles. Each subfigure shows the attitude distribution of the player portraying specific role (on the top) towards players in other roles (on the left).