EvoAgent is a generic method to automatically extend expert agents to multi-agent systems via the evolutionary algorithm.
Specifically, we consider the existing agent frameworks as the initial individual and then apply a series of evolutionary operators (e.g., mutation, crossover, selection, etc.) to generate multiple agents with diverse agent settings.
Why EvoAgent 🤔?
EvoAgent can be generalized to any LLM-based agent framework, and can automatically extend the existing agent framework to multi-agent systems without any extra human designs. Experimental results across various tasks have shown that EvoAgent can automatically generate multiple expert agents and significantly enhance the task-solving capabilities of LLM-based agents.🌍
We formulate the procedure of EvoAgent as a four-stage pipeline:
STEP 1: Initialization. Start with a predefined framework as the initial agents.
STEP 2: Crossover and Mutation. Employ the evolutionary operators to generate agents.
STEP 3: Selection. Employ a quality-check module to ensure agents retain characteristics and introduce variations.
STEP 4: Results Update. Generate results using child agents, then integrate these with previous results via LLMs.
Here, we choose the debate scenario used in MetaGPT, which includes two debaters with different opinions, leading to dull and repetitive content generation.
Instead of manually assigning new roles, we applied EvoAgent to extend each debate team to more agents with diverse settings, increasing the variety of opinions and the quality of the debate.
To align previous experiences, we select three NLP knowledge-intensive and reasoning-intensive tasks and one multi-modal task. We find that:
1. By utilizing multiple generated agents, EvoAgent can greatly improve LLM performances in both NLP knowledge and reasoning tasks.
2. EvoAgent can provide consistent improvements among each LLM, proving its strong generalization by using diverse generated agents.
We evaluate 30 scientific tasks in ScienceWorld to demonstrate the capability of EvoAgent in solving tasks in more challenging open-world environments. Results show that
EvoAgent can also extend interactive agents to multi-agent systems in solving complete scientific tasks in dynamic, open-world environments and consistently improve the performance of LLMs.
We also select TravelPlanner, a benchmark designed to evaluate language agents in real-world complex planning with multiple constraints. We find that:
1. EvoAgent can generate specialized agents, such as those focused on culinary experiences, transportation, and attractions.
2. By using EvoAgent to automatically generate multiple agents and forming a multi-agent collaboration paradigm, we can develop higher-quality plans that better meet user preferences.
@misc{yuan2024EvoAgent,
title={EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms},
author={Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, Deqing Yang},
year={2024},
archivePrefix={arXiv},
primaryClass={cs.CL}
}