EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms

EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms

The paper "EvoAgent: Towards Automatic Multi-Agent Generation via Evolutionary Algorithms" presents an innovative approach for automatically generating multi-agent systems from existing LLMs using evolutionary algorithms (EAs). The primary motivation behind this work is to overcome the limitations imposed by human-designed frameworks that currently dominate the design of LLM-based autonomous agents.

Abstract

The authors introduce EvoAgent as a general method to extend expert agents into multi-agent systems, thereby improving the efficacy of LLM-based agents for complex tasks. The approach leverages evolutionary operators such as mutation, crossover, and selection to generate multiple diverse agents from initial parent agents. EvoAgent's adaptability allows it to be applied to any existing LLM-based framework without additional human intervention. The paper claims that EvoAgent significantly enhances task-solving capacities across various tasks, as evidenced by experimental results.

Key Contributions

1. Generic Multi-Agent Generation Method: EvoAgent automates the extension of specialized agents to multi-agent systems using a generalized evolutionary algorithm infrastructure. This method is applicable universally across different LLM-based agent frameworks.
2. Evolutionary Processing for Agent Generation: EvoAgent simulates human societal evolution by considering each agent as an individual whose characteristics can evolve across generations. This evolutionary process allows the automatic formation of diverse and effective agents.
3. Extensive Empirical Evaluation: The effectiveness, scalability, and generality of EvoAgent are demonstrated through a series of tasks, including knowledge-based question answering, multi-modal reasoning, interactive scientific solving, and real-world planning scenarios.

Methodology

The methodology revolves around a four-stage pipeline:

1. Initialization: Starting from a pre-defined agent framework, which serves as the initial parent agents. Variables to be evolved are also defined during this stage.
2. Crossover and Mutation: Parent agents are used to generate initial results, which are then analyzed. The evolutionary operators - crossover and mutation - are used to produce candidate child agents with modified settings.
3. Selection: Employing a quality-check module based on LLMs to select the high-quality agents from the pool of candidates. This ensures that only agents with the most promising configurations survive.
4. Result Update: The newly generated agents are used to produce results, which are integrated with previous iterations to form the final output.

Experimental Evaluation

The experiments span multiple datasets and settings:

• NLP and Multi-Modal Tasks: EvoAgent is tested on benchmarks such as Logic Grid Puzzle, Trivia Creative Writing, and the MMMU dataset. It significantly improves performance metrics by generating diverse expert agents tailored to task-specific needs.
• Interactive Scientific Tasks: Using ScienceWorld, EvoAgent demonstrates enhanced problem-solving capabilities in dynamic environments compared to single-agent frameworks.
• Real-World Planning: EvoAgent considerably improves complex planning tasks in the TravelPlanner benchmark, showcasing its ability to handle intricate real-world constraints.

Results and Analysis

The obtained experimental results illustrate the following:

• Performance Gains: EvoAgent uniformly outperforms traditional CoT and refinement methods, providing substantial improvements in task completion metrics across various tasks.
• Scalability: By efficiently managing a diverse set of expert agents, EvoAgent demonstrates scalability in generating high-quality results without the bottlenecks associated with manual agent design.
• Generalization: The robustness of EvoAgent is evident from its applicability across different LLM backbones and task domains.

Future Implications

This research holds significant implications for the development of autonomous systems:

• Scalability in Autonomous Systems: EvoAgent shows that scalable and automated multi-agent systems can be achieved without extensive human intervention, paving the way for the deployment of more sophisticated and capable AI systems.
• Enhanced Collaboration and Specialization: By generating diverse expert agents, EvoAgent enables more nuanced and specialized problem-solving capabilities, mimicking human collaborative approaches more closely.
• Expansion to New Domains: The flexibility of EvoAgent suggests its potential applicability in new domains where traditional methods struggle, including advanced simulation environments or real-time strategy games.

Conclusion

The paper presents EvoAgent as a robust and scalable solution to automatically generate multi-agent systems via evolutionary algorithms. By addressing the confines of human-designed frameworks, this approach enhances the problem-solving capabilities of LLM-based agents in a wide array of complex tasks. The empirical evidence firmly establishes EvoAgent's potential to revolutionize multi-agent system design, significantly contributing to the fields of AI and autonomous systems. Future research may explore the integration of EvoAgent with cutting-edge long-context LLMs, further expanding its applicability and efficacy in even more sophisticated and resource-intensive environments.