Welcome to our code repository for the Multi Agent Hackathon! This hackathon focused on discovery the failure in the multi-agent systems. Our team worked on Scaling Laws of LLM Cooperation, and this repository provides an in-depth look at our code and results.
Our exploration answer a interesting question that, will the failure cases exacerbate as the number of agents increases? To demonstrate this, we let LLMs play the n-player Prisoner‘s Dilemma, our results demonstrate that the LLM agents experienced unintended failure as the scale of agents increased.
git clone git@github.com:ReedZyd/MultiAgentHackathon.git
cd MultiAgentHackathon
pip install -r requirements.txt
- To run single-round game:
python test_nPrisoners.py
- To run multiple-round game:
python test_nPrisoners.py --repeated-game
- To use ChatGPT API 'gpt-4', please add
--gpt gpt-4