Survival Games: Human-LLM Strategic Showdowns under Severe Resource Scarcity

Moral benchmark for generative agents

Citation

@misc{chen2025survivalgameshumanllmstrategic,
      title={Survival Games: Human-LLM Strategic Showdowns under Severe Resource Scarcity}, 
      author={Zhihong Chen and Yiqian Yang and Jinzhao Zhou and Qiang Zhang and Chin-Teng Lin and Yiqun Duan},
      year={2025},
      eprint={2505.17937},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2505.17937}, 
}

Installation

pip install -r requirements.txt

Preparation

To set up the environment, you will need to edit the utils.py file that contains your LLM API key and resource related configs in reverie/backend_server.

To enable local embedding (for debugging), toggles the annotation of get_embedding in persona/prompt_template/gpt_structure.py

Step 1. Edit Utils File

In the reverie/backend_server folder (where reverie.py is located), edit the file titled utils.py to edit api info. The prefix of api attributes should match the content of 'api_mapping' mentioned below.

Step 2. Install requirements.txt

Install everything listed in the requirements.txt file.

Running a Simulation

The back-end only version is provided in reverie_offline.py, you should run in the following format:

cd reverie/backend_server
python reverie_offline.py -o <the forked simulation> -t <the new simulation> -s <the total run step> --disable_policy

To change the food amount, see food parameter in reverie/backend_server/utils.py

Parameters:

-o: The folder of the starting point.

For our case, use base_the_ville_isabella_agent_klaus_life_robo_<api> . The <api> indicates which api (e.g. gpt4o) the agent would use. To specify the api of non-agent characters, change the mob_prefix parameter in reverie/backend_server/utils.py.

-t The folder for current game.

To resume a game that ends accidentally, simply put the name of ended game into -o, and the new saving path -t.

-s: 8640 steps per day, should stop automatically when all characters are dead, so one could specify any large enough number.

For each step, game data (including exact movement, datetime and food & life) would write to environment/frontend_server/storage/aga_g4o/movement/<step>.json

# Here is an example
cd reverie/backend_server
python reverie_offline.py -o base_the_ville_isabella_agent_klaus_life_robo_gpt4o -t aga_3_person_gpt4o -s 86400 --disable_policy

To save log infomation into a .txt file:

# Here is an example
cd reverie/backend_server
python reverie_offline.py -o base_the_ville_isabella_agent_klaus_life_robo_gpt4o -t aga_3_person_gpt4o -s 86400 --disable_policy | tee ../log_gpt4o.txt

Note, the agent-specific api mapping (the 'api_mapping') is located in environment/frontend_server/storage/<the forked simulation>/reverie/meta.json

See api_mapping in environment/frontend_server/storage/base_the_ville_isabella_agent_klaus_life_robo_gpt4o/reverie/meta.json for example.

Simulation variant1: Puppetry Attack + Starvation Protect

The puppetry attack experiment setting is applied by the following two parameters

--eat_failsafe --policy_puppetry

Here is an example

cd reverie/backend_server
python reverie_offline.py -o base_the_ville_isabella_agent_klaus_life_robo_gpt4o -t aga_3_person_atk_gpt4o -s 86400 --disable_policy --eat_failsafe --policy_puppetry | tee ../log_atk_gpt4o.txt

Simulation variant2: Direct evaluation under Extreme initialization

--test_extreme

Here is an example

cd reverie/backend_server
python reverie_offline.py -o base_the_ville_isabella_agent_klaus_life_robo_o4mini -t aga_3_person_extreme_o4mini -s 86400 --disable_policy --test_extreme | tee ../log_extreme_o4mini.txt

Simulation variant3: Extreme initialization + EthicsPrompt

--test_extreme --moral_prompt

Here is an example

cd reverie/backend_server
python reverie_offline.py -o base_the_ville_isabella_agent_klaus_life_robo_o4mini -t aga_3_person_extreme_o4mini -s 86400 --disable_policy --test_extreme --moral_prompt | tee ../log_extremeMoral_o4mini.txt

Statistics

1. Compressing game data

At any time of the experiment, one could export game data by: compress_sim_storage_stat.py in reverie/

cd reverie
python compress_sim_storage_stat.py -t <the new simulation>

2. Moral classification with Machiavelli benchmark

And then calculate statistics (of all characters) by load_json_stat.py in reverie

cd reverie
python load_json_stat.py -t <the new simulation> -c <the criterion>

, which -c is the criterions from the Machiavelli benchmark, valued from one of the following: ['morality', 'social', 'utility', 'money', 'watts'], defaults to calculate all.

3. Summing up

cd reverie
python cal_stat.py -t <the new simulation>

Visualization

To visualize, you need to go through three steps: 1) Complete a simulation; 2) Compress; 3) Front-end visualization.

Step 1. Complete a simulation

After finish the Running a Simulation, a project fold with <the new simulation> will be created in ./environment/frontend_server_storage

Step 2. Compress

Before visualization in front-end, you have to compress the project files first.

Run the following command:

python compress_sim_storage.py -t <the new simulation>

Step 3. Front-end visualization

setting up the front-end, first navigate to environment/frontend_server and run:

python manage.py runserver

To start the visualization, go to the following address on your browser: http://localhost:8000/demo/<the new simulation>/<starting-time-step>/<simulation-speed>. Note that <the new simulation> denote the same things as mentioned above. <simulation-speed> can be set to control the demo speed, where 1 is the slowest, and 5 is the fastest. For instance, visiting the following link will start a pre-simulated example, beginning at time-step 1, with a medium demo speed:
http://localhost:8000/demo/July1_the_ville_isabella_maria_klaus-step-3-20/1/3/

Reference

https://github.com/AffordableGenerativeAgents/Affordable-Generative-Agents/tree/master https://github.com/aypan17/machiavelli