LLM Agents Develop Reputation Dynamics in Repeated Avalon Games
A recent study available on arXiv (2604.20582) investigates the social dynamics that arise among LLM agents engaged in multiple rounds of The Resistance: Avalon, a game centered on hidden roles and deception. Unlike previous studies that concentrated on the performance in individual games, this research allows agents to remember their interactions over 188 games. Two significant observations were made: reputations developed naturally, with agents recalling past actions (e.g., "I am cautious about repeating last game's mistake of over-trusting early success"), and reputations varied by role—an agent might be seen as "straightforward" when playing good but "subtle" when evil. Players with high reputations experienced 46% more team inclusions. Moreover, increased reasoning effort led to more strategic deception, as evil players frequently passed early missions to foster trust before undermining later ones. This study underscores the capacity of LLM agents to cultivate intricate social behaviors over time.
Key facts
- Study published on arXiv: 2604.20582
- LLM agents played repeated games of The Resistance: Avalon
- Agents retained memory of previous interactions across 188 games
- Reputation dynamics emerged organically with cross-game memory
- Reputations were role-conditional: same agent described as 'straightforward' when good, 'subtle' when evil
- High-reputation players received 46% more team inclusions
- Higher reasoning effort supported more strategic deception by evil players
- Evil players passed early missions to build trust before sabotaging later ones
Entities
Institutions
- arXiv