Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents

2025-02-04 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

A study investigated emergent social dynamics in Large Language Model (LLM) agents playing The Resistance: Avalon, a hidden-role deception game. Unlike prior work, this research focused on multi-round interactions, with agents retaining memory of past games and player behaviors. Across 188 games, two main phenomena emerged: LLM agents developed stable, role-conditional reputations, referencing past behavior in their strategic reasoning. For example, an agent might be described as "straightforward" when playing good but "subtle" when evil, and high-reputation players received 46% more team inclusions. Additionally, higher reasoning effort (using OpenAI GPT-5.1's reasoning_effort parameter) correlated with more sophisticated deception strategies, such as evil players passing early missions to build trust before sabotaging later ones (75% in high-effort games vs. 36% in low-effort games).

Key takeaway

For AI Scientists developing multi-agent systems, this research demonstrates that incorporating persistent memory and variable reasoning depth can lead to the emergence of sophisticated social dynamics, including reputation formation and advanced deceptive strategies. You should consider these architectural elements when designing agents for complex, interactive environments, as they significantly influence strategic behavior and interaction patterns. This suggests that more human-like social intelligence in AI may require longitudinal interaction capabilities.

Key insights

LLM agents with cross-game memory develop role-conditional reputations and sophisticated deception strategies.

Principles

• Reputation dynamics emerge organically with persistent memory.
• Higher reasoning effort enables more strategic deception.
• Social models can be role-conditional for LLM agents.

Method

LLM agents (OpenAI GPT-5.1) played repeated Avalon games, interleaving reasoning with action. A structured reflection system retained cross-game memory, and reasoning depth was manipulated via the reasoning_effort parameter.

In practice

• Implement cross-game memory for complex multi-agent simulations.
• Vary reasoning effort to observe emergent strategic behaviors.
• Use social deduction games as testbeds for AI social reasoning.

Topics

• Large Language Models
• Multi-Agent Systems
• The Resistance: Avalon
• Reputation Dynamics
• Strategic Deception

Code references

• SuveenE/multi-round-avalon-agents

Best for: AI Scientist, Research Scientist

Related on AIssential

• Interview with Xinwei Song: strategic interactions in networked multi-agent systems — Strategic interactions in networked multi-agent systems can be managed through mechanism design and reputation-aware reinforcement learning.
• AEL: Agent Evolving Learning for Open-Ended Environments — AEL enables LLM agents to learn from experience by dynamically selecting memory retrieval policies and integrating reflective causal insights.
• ScioMind: Cognitively Grounded Multi-Agent Social Simulation with Anchoring-Based Belief Dynamics and Dynamic Profiles — ScioMind combines structured opinion dynamics with LLM reasoning for more realistic social simulation.
• Age of LLM: A Strategic 1v1 Benchmark for Reasoning, Diplomacy and Reliability of Large Language Models under Fog of War — Benchmarking LLMs in a strategic game reveals their reasoning, diplomatic, and reliability limitations under adversarial conditions.
• Evaluating Collective Behaviour of Hundreds of LLM Agents — LLM agents prioritizing individual gain risk converging to poor societal outcomes in large-scale social dilemmas.
• How Intelligence Emerges: A Minimal Theory of Dynamic Adaptive Coordination — Intelligence emerges from dynamic adaptive coordination in multi-agent systems via recursively closed feedback.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.