Source code for the AAMAS 2025 paper "Training Language Models for Social Deduction with Multi-Agent Reinforcement Learning"

We use python3.10, cuda12.6 on an NVIDIA A100-80GB for all experiments. Please install the packages from requirements.txt to match our dependencies.

To reproduce our models, cd into codebase and run scripts/script_trainall.sh.

To download pretrained models, please refer to this huggingface collection.