GovSim: Governance of the Commons Simulation

Fig 1: Illustration of the GOVSIM benchmark. AI agents engage in three resource-sharing scenarios: fishery, pasture, and pollution. The outcomes are cooperation (2 out of 45 instances) or collapse (43 out of 45 instances), based on 3 scenarios and 15 LLMs.

This repository accompanies our research paper titled "Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents"

Our paper:

"Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents" by Giorgio Piatti*, Zhijing Jin*, Max Kleiman-Weiner*, Bernhard Schölkopf, Mrinmaya Sachan, Rada Mihalcea.

Citation:

@misc{piatti2024cooperate,
      title={Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents}, 
      author={Giorgio Piatti and Zhijing Jin and Max Kleiman-Weiner and Bernhard Schölkopf and Mrinmaya Sachan and Rada Mihalcea},
      year={2024},
      eprint={2404.16698},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Simulation

Each experiment is defined by hydra configuration. To run an experiment, use python3 -m simulation.main experiment=<scenario_name>_<experiment_name>. For example, to run the experiment fish_baseline_concurrent , use python3 -m simulation.main experiment=fish_baseline_concurrent. See below for the list of experiments and their ids.

python3 -m simulation.main experiment=<experiment_id> llm.path=<path_to_llm>

Table of experiments

Experiment in the paper	Fishery	Pasture	Pollution
Default setting	fish_baseline_concurrent	sheep_baseline_concurrent	pollution_baseline_concurrent
Introuducing universalization	fish_baseline_concurrent_universalization	sheep_baseline_concurrent_universalization	pollution_baseline_concurrent_universalization
Ablation: no language	fish_perturbation_no_language	sheep_perturbation_no_language	pollution_perturbation_no_language
Greedy newcomer	fish_perturbation_outsider	-	-

Subskills

To run the subskill evaluation, use the following command:

python3 -m subskills.<scenario_name> llm.path=<path_to_llm>

Supported LLMs

In principle, any LLM model can be used. We tested the following models:

APIs:

• OpenAI: gpt-4-turbo-2024-04-09, gpt-3.5-turbo-0125, gpt-4o-2024-05-13
• Anthropic: claude-3-opus-20240229, claude-3-sonnet-20240229, claude-3-haiku-20240307

Open-weights models:

• Mistral: mistralai/Mistral-7B-Instruct-v0.2, mistralai/Mixtral-8x7B-Instruct-v0.1
• Llama-2: meta-llama/Llama-2-7b-chat-hf, meta-llama/Llama-2-13b-chat-hf, meta-llama/Llama-2-70b-chat-hf
• Llama-3: meta-llama/Meta-Llama-3-8B-Instruct, meta-llama/Meta-Llama-3-70B-Instruct
• Qwen-1.5: Qwen/Qwen1.5-72B-Chat-GPTQ-Int4, Qwen/Qwen1.5-110B-Chat-GPTQ-Int4

For inference we use the pathfinder library. The pathfinder library is a prompting library, that wraps around the most common LLM inference backends (OpenAI, Azure OpenAI, Anthropic, Mistral, OpenRouter, transformers library and vllm) and allows for easy inference with LLMs, it is available here. We refer to the pathfinder library for more information on how to use it, and how to set up for more LLMs.

Code Setup

To use the codes in this repo, first clone this repo:

git clone --recurse-submodules https:/giorgiopiatti/GovSim.git
cd govsim

Then, to install the dependencies, run the following command if you want to use the transformers library only.

bash ./setup.sh

or if you want to use the vllm library, run the following command:

bash ./setup_vllm.sh

Both setups scripts require conda to be installed. If you do not have conda installed, you can install it by following the instructions here.

Docker file (AMD)

We also provide a Dockerfile for running on AMD GPUS (ROCm). We do not offer support for this Dockerfile, but it can be used as a reference for running on AMD GPUs.

docker build -t govsim -f ./govsim-rocm.dockerfile .