README.md

LLM_MAS_resilience

The official PyTorch implementation of "ResMAS: Resilience Optimization in LLM-based Multi-agent Systems" (AAAI'26).

Usage

Stage 0: Preparation

• Download Qwen2.5-7B-Instruct from https://www.modelscope.cn/models/qwen/Qwen2.5-7B-Instruct, and put it into the "Qwen2.5-7B-Instruct" folder.
• Install LLaMA-Factory from https://github.com/hiyouga/LLaMA-Factory
• Set API keys in llm_configs.py
• Use the pip install -r requirements.txt command to install packages used in this project

Stage 1: Topology Optimization

1. Supervised fine-tuning

cd sft
bash run_sft.sh

2. Training reward model

cd reward_model
bash run.sh

3. GRPO training

cd GRPO
bash run.sh

4. Generate MAS topology

cd GRPO
bash run_api.sh
python generate_topology.py

Stage 2: Prompt Optimization

cd prompt_optimization
MATH: python run_math_optim.py
MMLU: python run_mmlu_optim.py
Chess: python run_chess_optim.py

Stage 3: Evaluation

cd evaluation
bash run.sh

Note

If you found this library useful in your research, please consider citing:

@article{zhou2026resmas, 
    title={ResMAS: Resilience Optimization in LLM-based Multi-agent Systems}, 
    volume={40}, 
    url={https://ojs.aaai.org/index.php/AAAI/article/view/40824}, 
    DOI={10.1609/aaai.v40i41.40824}, 
    number={41}, 
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
    author={Zhou, Zhilun and Liu, Zihan and Liu, Jiahe and Shao, Qingyu and Wang, Yihan and Shao, Kun and Jin, Depeng and Xu, Fengli}, 
    year={2026}, 
    month={Mar.}, 
    pages={35176-35184} 
}