This repository is the implementation of our paper EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation.
Large language models (LLMs) have demonstrated an impressive ability to role-play humans and replicate complex social dynamics. While large-scale social simulations are gaining increasing attention, they still face significant challenges, particularly regarding high time and computation costs. Existing solutions, such as distributed mechanisms or hybrid agent-based model (ABM) integrations, either fail to address inference costs or compromise accuracy and generalizability. To this end, we propose EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation. EcoLANG operates in two stages: (1) language evolution, where we filter synonymous words and optimize sentence-level rules through natural selection, and (2) language utilization, where agents in social simulations communicate using the evolved language. Experimental results demonstrate that EcoLANG reduces token consumption by over 20%, enhancing efficiency without sacrificing simulation accuracy.
Language evolution is conducted on the dev set of Synthetic-Persona-Chat dataset.
Social simulation is validated on PHEME and Metoo & Roe movements of HiSim dataset.
- PHEME: To effectively observe the communication (content), we preserved the reactions in each thread of PHEME and removed replies that could not reconstruct the propagation chain (i.e., tweets that could not be matched to a parent tweet). This resulted in retaining 196 threads (i.e., 196 scenarios), which were then formatted for OASIS simulation. See
./data/phemefor the processed data. - HiSim: Due to the Twitter's terms of service, we can not release the raw text data or the user data. Please contact the author (via xymou20@fudan.edu.cn) if you need the processed data in the OASIS simulation format.
EcoLANG requires three conda environments:
- ecolang-vllm: for vLLM deployment;
- ecolang-evo: for language rule evolution;
- ecolang-simu: for validation of social simulation.
To support custom vocabulary for generation, modifications need to be made to parts of the vllm, camel and oasis code. Please refer to the zip files in (./pkg).
- for llm serving
conda create -n ecolang-vllm python=3.10
conda activate ecolang-vllm
pip install -r requirements_vllm.txt
# install vllm 0.6.4.post2.dev394+g2bfdbf2a
# If you cannot find this version of vllm package, try to unzip the vllm pkg of ./pkg in the path of site-packages of ecolang-vllm env. You can run 'pip show pip' to find the path.
# cd pkg
# unzip vllm.zip -d /opt/miniconda3/envs/ecolang-vllm/lib/python3.10/site-packages/
# prepare self-defined logits_processor
cd pkg/logits_processors
pip install -e .- for language evolution (simulation based on conversations using AutoGen)
conda create -n ecolang-evo python=3.10
conda activate ecolang-evo
git clone https://github.com/xymou/EcoLANG.git
cd EcoLANG
pip install -r requirements_evo.txt
# prepare self-defined logits_processor
cd pkg/logits_processors
pip install -e .- for simulation (simulation based on OASIS)
conda create -n ecolang-simu python=3.10
conda activate ecolang-simu
cd EcoLANG
pip install -r requirements_simu.txt
# install oasis
cd pkg
unzip oasis.zip
cd oasis
pip install -e .
# install camel
cd ..
# unzip the camel pkg of ./pkg in the path of site-packages of ecolang-simu env.
# unzip -o camel.zip -d /opt/miniconda3/envs/ecolang-simu/lib/python3.10/site-packages/
# prepare self-defined logits_processor
cd logits_processors
pip install -e .- code structure
EcoLANG/
│
├── configs/ # overall configs of simulation
├── scen_configs/ # config files of each scenario of a dataset (automatically generated when running simulation)
├── data/ # dataset
├── rules/ # initialized and evolved rules
├── pkg/ # modified vllm, camel and oasis code for this work
├── output/ # simulation results
├── score_ckpt/ # save intermediate evaluation results
└── code/
├── scripts/ # sh command set for simulation
├── run_evolution.py # main file of language evolution
├── prepare_config.py # prepare configs for each scenario
├── chat_simulation.py # chat simulation for language evolution
├── diversification.py # cross and mutation in the evolution
├── myselect.py # evaluation and selection in the evolution
----------------------
├── run_oasis_eval.py # main file of social simulation of PHEME
├── run_oasis_hisim_eval.py # main file of social simulation of HISIM
├── twitter_simulation.py # start an OASIS simulation of a single scenario in PHEME
├── twitter_simulation_hisim.py # start an OASIS simulation of a single scenario in HISIM
├── collect_simu_res.py # collect simulation result
└── evaluation_prl.py # evaluation protocol of each datasetEvolve the language rules through simulations of syn-persona-chat dataset. Here are some examples showing how to run the code.
The compressed vocabs are in ./data/vocab.
Remember to give your base_url and api_key in the scripts.
- serve the LLMs using vLLM
conda activate ecolang-vllm
cd code
bash scripts/local_sevrer_vllm.sh- run the evolution code
conda actiavte ecolang-evo
cd code
bash scripts/evo_persona_hisim_vocab.sh- serve the LLMs using vLLM
conda activate ecolang-vllm
cd code
bash scripts/local_sevrer_vllm.sh- run the simulation code
conda activate ecolang-simu
cd code
bash scripts/run_hisim_vocab_topdown_rule.sh # run ecolang (vocab+rule)
bash scripts/run_pheme_vocab_topdown_rule.sh # run ecolang (vocab+rule)The evaluation process will be carried out in the simulation code above. More details can be found in code/evaluation_prl.py
Remember to give your base_url and api_key in the scripts.
Please consider citing this paper if you find this repository useful:
@article{mou2025ecolang,
title={EcoLANG: Efficient and Effective Agent Communication Language Induction for Social Simulation},
author={Xinyi Mou and Chen Qian and Wei Liu and Xuanjing Huang and Zhongyu Wei},
journal={arXiv preprint arXiv:2505.06904},
year={2025}
}