• [APR. 2024] Check new benchmark implementation with AgentLite.
• [March 2024] Our paper BOLAA is accepted to LLMAgent@ICLR workshop!
• [Aug. 2023] Initial Release of BOLAA paper and implementation code!

This is the repo for BOLAA paper. In this paper, we create benchmark on LLM-augmented Autonoumous Agents (LAA). We compare 6 different LAA architecture, including 5 existing intuitions and 1 new BOLAA agent. And all those agents are paired with different LLMs to compare the performance. BOLAA is able to communicate and orchestrate multiple specialitist agents: We tested on two types of enviroments: the webshop navigation environment, and HotPotQA enviroment. An example of the BOLAA web agent simulation on webshop enviroment is:

Besides BOLAA arch, we also devise five standard LAA arches, the Zeroshot (ZS), Zeroshot-Think (ZST), ReAct, PlanAct, PlanReAct as follows:

1. Setup the fastchat to use local open-source LLMs. Go to next step if you only test openai API.
2. Setup OPENAI API KEY in both webrun/config and hotpotqa_run/config. Skip this if you only test open-source LLMs.
3. Setup the webshop environment if you are testing web agent
4. Setup the agent_benchmarking environment as follows:

conda create -n agent_benchmark python=3.10 -y
conda activate agent_benchmark
pip install -r requirements.txt

python run_webagent.py --agent_name Search_Click_Control_Webrun_Agent --llm_name gpt-3.5-turbo --max_context_len 4000

other agent options can be found in test_webagent.sh. The implementation code for various web agents is in web_run

python run_hotpotqaagent.py --agent_name React_HotPotQA_run_Agent --llm_name gpt-3.5-turbo --max_context_len 4000

other agent options commands can be found in test_hotpotqa.sh. The implementation code for various web agents is in hotpotqa_run

If you find our paper or code useful, please cite

@misc{liu2023bolaa,
      title={BOLAA: Benchmarking and Orchestrating LLM-augmented Autonomous Agents}, 
      author={Zhiwei Liu and Weiran Yao and Jianguo Zhang and Le Xue and Shelby Heinecke and Rithesh Murthy and Yihao Feng and Zeyuan Chen and Juan Carlos Niebles and Devansh Arpit and Ran Xu and Phil Mui and Huan Wang and Caiming Xiong and Silvio Savarese},
      year={2023},
      eprint={2308.05960},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

• Part of our environment code reuse ReAct code.
• Our LLM API is based on Langchain
• We use the WebShop and HotPotQA for testing.