Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

DynaSwarm: Dynamically Graph Structure Selection for LLM-based Multi-agent System

Created by
  • Haebom

Author

Hui Yi Leong, Yuqing Wu

Outline

This paper proposes DynaSwarm, a dynamic framework for enhancing large-scale language model (LLM)-based multi-agent systems (MAS) to overcome the limitations of existing MASs that rely on static, manually designed collaborative graph structures. DynaSwarm leverages two key innovations: (1) graph structure optimization via the actor-critic reinforcement learning (A2C) mechanism, which improves stability compared to existing reinforcement learning (RL) methods, and (2) a dynamic graph selector that adaptively selects the optimal graph structure for each input sample via parameter-efficient LLM fine-tuning. This allows DynaSwarm to dynamically route queries through a network of specialized agents by exploiting sample-specific features, rather than relying on a fixed graph architecture that applies to all samples. Furthermore, we propose a method for fine-tuning a demo searcher to maximize the effectiveness of in-context learning (ICL). Extensive experiments on question answering, mathematical reasoning, and coding tasks demonstrate that DynaSwarm consistently outperforms state-of-the-art single-agent and MAS baseline models across multiple LLM backbones.

Takeaways, Limitations

•
Takeaways:
◦
We emphasize the importance of dynamic graph structures in LLM-based MAS.
◦
Achieve stable performance improvements over existing RL methods through A2C-based graph structure optimization.
◦
Increase adaptability to various tasks through dynamic graph selection based on sample features.
◦
We present a demo searcher fine-tuning technique that improves performance by leveraging in-context learning (ICL).
◦
It outperforms state-of-the-art models across a variety of tasks (question answering, mathematical reasoning, and coding).
•
Limitations:
◦
This paper does not address the specific Limitations. Further analysis and experiments are needed to elucidate Limitations. For example, a more detailed analysis is needed on the computational cost and convergence speed of the A2C algorithm, the complexity of dynamic graph selection, and generalization performance for specific types of problems.
👍
Made with Slashpage