Simulating Prosocial Behavior and Social Contagion in LLM Agents under Institutional Interventions

This repository contains the official code and documentation for our paper, titled "Simulating Prosocial Behavior and Social Contagion in LLM Agents under Institutional Interventions."

Install Environment

To set up the environment and dependencies, run the following commands:

conda create -n ProSim python=3.10

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118

pip install -r requirements.txt

Running the Experiments

Study 1: Prosocial Behavior

python prosocial_behavior.py

Parameters

Parameter	Type	Description
--model_name	str	The name of the model to be used in the experiment.
--use_local_model	bool	If set, the local version of the model will be used.
--max_new_tokens	int	The maximum number of tokens that can be generated in a single response by the local model.
--temperature	float	Default: 1.0.
--num_threads	int	The number of threads to use for parallel execution.
--output_reason	bool	If set, the model will provide reasoning along with the generated response.
--num_agents	int	The number of LLM agents involved in the experiment.
--policy	str	The policy that governs the agents' behavior. Choices include "moral", "social", "enforce", "economic".
--chara_file_path	str	The path to the file containing the agents' character data.
--max_attempts	int	The maximum number of attempts per dialogue.

Note: Parameters that are shared across studies (e.g., --model_name, --num_agents, etc.) are only documented in Study 1 for brevity.

Study 2: Third-Party Punishment

python third_party_punishment.py

Parameters

Parameter	Type	Description
--num_rounds	int	The number of rounds each agent plays in the game.

Study 3: Inequity Information Spread with Prosocial Context

python inequality_info_spread.py

Parameters

Parameter	Type	Description
--explicit_spread	bool	If set, the experiment will explicitly spread unfair information.
--history_action	bool	If set, the prompt will include the historical actions of each agent.
--rate_inequality	bool	IIf set, the agent will also output its perception of inequality.
--num_rounds	int	The number of iterations agents will participate in during the experiment.
--activation_rate	float	The probability of activating edges in the small-world network during each round of the experiment.
--inequality_agent_counts	int	The initial number of agents that perceive inequality.
--inequality	str	The type of inequality being simulated in the experiment.
--chara_detail_file_path	str	The file path for the agent's detailed data.
--swn_file_path	str	The file path for the small-world network data.

Citation

If you use this code or results in your work, please cite the following paper:

@article{zhou2025simulating,
  title={Simulating Prosocial Behavior and Social Contagion in LLM Agents under Institutional Interventions},
  author={Zhou, Yujia and Wang, Hexi and Ai, Qingyao and Wu, Zhen and Liu, Yiqun},
  journal={arXiv preprint arXiv:2505.15857},
  year={2025}
}