TerraLingua

Paper: Link

Dataset: https://huggingface.co/datasets/GPaolo/TerraLingua

Dataset dashboard: https://aianthropology.decisionai.ml/

A multi-agent simulation framework for studying emergent behavior, artifact creation, and cultural evolution.

LLM-powered agents (Claude or other models) interact in a shared 2D grid environment — foraging for resources, creating text artifacts, reproducing, and communicating — enabling research into how language-using agents develop social structure and culture over time.

After each experiment, the AI Anthropologist — itself an LLM agent — analyzes the simulation logs to annotate agent behaviors, infer group dynamics, classify artifacts, and trace cultural lineages, providing a qualitative and quantitative account of what emerged.

An overview of the TerraLingua system and of the AI-Anthropologist is shown in the figure below.

Installation

Requires Python 3.13+.

Using venv:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Using conda:

conda create -n terralinguia python=3.13
conda activate terralinguia
pip install -r requirements.txt

Copy .env.example to .env and fill in your API key(s):

cp .env.example .env

Running Experiments

Run directly with main.py using CLI flags:

python main.py --exp_name my_experiment --init_agents 10 --max_ts 200 --model claude-haiku-4-5

Or use run_experiment.sh, a fully annotated template with all available options documented:

bash run_experiment.sh

Logs are written to logs/<exp_name>/.

Reproducing paper experiments

The paper_experiment_scripts/ folder contains the exact scripts used to run each experiment from the paper. All scripts must be run from the project root:

bash paper_experiment_scripts/run_core.sh

Supported Agent Models

Pass any of the following keys via --model:

Key	Provider	Notes
claude-haiku-4-5	Anthropic	Fast, cost-effective
claude-sonnet-4-6	Anthropic	Default
o4-mini	OpenAI
o3-mini	OpenAI
gpt-5.1	OpenAI
gpt-5-mini	OpenAI
QWEN2.5	Local (vLLM)	Qwen2.5-32B-Instruct
QWEN3	Local (vLLM)	Qwen3-32B
DeepSeek-R1-32	Local (vLLM)	DeepSeek-R1-Distill-Qwen-32B
DeepSeek-R1-70	Local (vLLM)	DeepSeek-R1-Distill-Llama-70B

Local models (vLLM)

Local models require a running vLLM server. Start one (or more) on any of the default ports (9000–9003, 9010–9012):

vllm serve Qwen/Qwen3-32B --port 9000

Then pass the ports via --ports (defaults to 9000 9001 9002 9003 9010 9011 9012):

python main.py --model QWEN3 --ports 9000 9001

TerraLingua will auto-discover which ports are hosting the requested model and load-balance across them.

Data Analysis

Analysis is performed by the AI Anthropologist, a post-hoc LLM-based framework that annotates agent behaviors, infers group dynamics, classifies artifacts, and traces cultural lineages. See analysis_scripts/AI_ANTHROPOLOGIST.md for a detailed description of the pipeline.

Scripts follow a numbered order and must be run from the project root (they import from core and analysis_scripts as packages):

Script	Description
001_llm_agent_analyser.py	Annotate agent logs with LLM-generated behavior labels
002_make_graph.py	Build interaction graphs and compute network metrics
003_llm_group_analyser.py	Group-level behavioral analysis
004_artifact_analysis.py	Compute artifact complexity metrics
005_artifact_classification.py	Classify artifacts into behavioral categories
006_artifact_philogeny.py	Analyze artifact genealogy and conceptual ancestry

python analysis_scripts/001_llm_agent_analyser.py

Data Visualization

Notebooks in analysis_scripts/notebooks/ mirror the analysis pipeline:

Notebook	Description
n000_general_stats.ipynb	Overall experiment statistics
n001_llm_agent_analyser.ipynb	Per-agent behavior visualization
n002_graph_analysis.ipynb	Interaction network plots
n003_llm_group_analysis.ipynb	Group dynamics
n004_artifact_analysis.ipynb	Artifact complexity over time
n005_artifact_categories.ipynb	Classification results
n006_artifact_phylogeny.ipynb	Artifact lineage trees
n007_interactive_phylogeny.ipynb	Interactive phylogeny explorer

jupyter notebook analysis_scripts/notebooks/

Citation

If you use TerraLingua in your research, please cite:

@techreport{paolo26terralingua,
title = "TerraLingua: Emergence and Analysis of Open-Endedness in LLM Ecologies",
author = "Giuseppe Paolo and Jamieson Warner and Hormoz Shahrzad and Babak Hodjat and Risto Miikkulainen and Elliot Meyerson",
year = 2026,
month = jan,
institution = "Cognizant AI Lab",
url = "https://www.researchgate.net/publication/402263491_TerraLingua_Emergence_and_Analysis_of_Open-endedness_in_LLM_Ecologies",
doi = "10.13140/RG.2.2.25551.55206",
number = "2026-01",
}