Multi-Agent Reinforcement Learning for Language-Based Social Deduction

Abstract

This study applies MARL to train LLM agents for language-based social deduction, where communication directly influences multi-agent outcomes. Speaking policies are optimized using rewards that capture the causal impact of messages on other agents’ beliefs, while listening models predict hidden state information from dialogue. Training on 12,000 simulated game episodes results in a 2.1× increase in win rate over standard RL baselines and demonstrates emergent strategic behaviors such as coordinated accusations and evidence sharing.