Awesome

Learning to Break: Knowledge-Enhanced Reasoning in Multi-Agent Debate System

Paper

Running experiments

Preparation

tqdm
openai
vllm
torch

Run MADKE

• We used the GPT3.5 and GPT4 apis provided by Microsoft Azure to complete the experiment. You need to specify an endpoint, deployment id, and api key.
• We used VLLM to deploy Qwen1.5-32B-Chat and Qwen1.5-72B-Chat, and then called the API. Run the corresponding model with the following code.

python -m vllm.entrypoints.openai.api_server --model /home/llm_models/Qwen1.5-32B-Chat  --served-model-name qwen1.5-32b-chat

python madke.py --test_data_path ./data/triviaqa/triviaqa_test_retrieval_all_evidence.json \
--test_output_path ./results/predictions/triviaqa/triviaqa_test_500_google_chatgpt_a2r3.json \
--endpoint YOUR_ENDPOINT \
--deployment_id YOUR_DEPLOYMENT_ID \
--api_key YOUR_API_KEY 

Run Evaluation

python evaluation.py \
--endpoint YOUR_ENDPOINT \
--deployment_id YOUR_DEPLOYMENT_ID \
--api_key YOUR_API_KEY 

Citation

If you find this work helpful, we kindly request that citations refer to the arXiv version:

@misc{wang2024learningbreak,
      title={Learning to Break: Knowledge-Enhanced Reasoning in Multi-Agent Debate System}, 
      author={Haotian Wang and Xiyuan Du and Weijiang Yu and Qianglong Chen and Kun Zhu and Zheng Chu and Lian Yan and Yi Guan},
      year={2024},
      eprint={2312.04854},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2312.04854}, 
}