Skip to content

salesforce/BOLAA

BOLAA: Benchmarking and Orchestrating LLM-augmented Autonomous Agents

News

Introduction

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: BOLAA 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: BOLAA-webshop

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

Installation

  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

Web Agent Simulation

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

Webshop Reward Table

HotpotQA Agent Simulation

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 Hotpot Reward Table

Citation

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}
}

Acknowledge