Learning to Evolve: A Self-Improving Framework for Multi-Agent Systems via Textual Parameter Graph Optimization
Summary: arXiv:2604.20714v1 Announce Type: new
Abstract
Designing and optimizing multi-agent systems (MAS) is a complex, labor-intensive process often referred to as “Agent Engineering.” Existing automatic optimization methods primarily focus on flat prompt tuning, which lacks the structural awareness needed to debug the intricate web of interactions inherent in MAS. More critically, these optimizers are static and do not learn from experience to improve their own optimization strategies. To address these gaps, we introduce Textual Parameter Graph Optimization (TPGO), a framework that enables a multi-agent system to learn to evolve.
Introduction
TPGO first models the MAS as a Textual Parameter Graph (TPG), where agents, tools, and workflows are modular, optimizable nodes. This innovative approach allows for a more nuanced understanding of the interactions and dependencies between various components of the system.
Key Features of TPGO
- Modular Design: Agents, tools, and workflows are represented as nodes in a graph, facilitating easier manipulation and optimization.
- Textual Gradients: The framework derives structured natural language feedback from execution traces, enabling it to pinpoint failures and suggest granular modifications.
- Group Relative Agent Optimization (GRAO): A novel meta-learning strategy that learns from historical optimization experiences, allowing the system to propose effective updates based on past successes and failures.
How TPGO Works
The core of TPGO is its ability to analyze historical data to refine its optimization strategies. By leveraging textual gradients, the framework identifies specific areas for improvement and adapts its approach accordingly. This self-improving capability is particularly beneficial in complex environments where traditional optimization methods fall short.
Experimental Results
Extensive experiments conducted on complex benchmarks such as GAIA and MCP-Universe demonstrate that TPGO significantly enhances the performance of state-of-the-art agent frameworks. The results indicate that TPGO achieves higher success rates through automated, self-improving optimization compared to existing methods.
Conclusion
In summary, Textual Parameter Graph Optimization presents a groundbreaking approach to multi-agent systems optimization. By introducing a self-improving framework that leverages textual gradients and a meta-learning strategy, TPGO addresses the limitations of traditional optimization methods. As the field of multi-agent systems continues to evolve, frameworks like TPGO will play a crucial role in enhancing the efficiency and effectiveness of agent-based applications.
Future Directions
- Exploration of additional domains for TPGO application.
- Improvement of the GRAO mechanism for better learning efficiency.
- Integration with other emerging AI technologies for enhanced performance.