AMAS: Adaptively Determining Communication Topology for LLM-based Multi-Agent System

Hui Yi Leong, Yuheng Li, Yuqing Wu, Wenwen Ouyang, Wei Zhu, Jiechao Gao, Wei Han

Introduction

Large Language Model (LLM) based Multi-Agent Systems (MAS) have emerged as a powerful paradigm for solving complex tasks by decomposing them into specialized roles. Traditionally, these systems rely on static topologies—predefined communication structures that dictate how agents interact. However, a fixed "org chart" is often suboptimal; a structure effective for mathematical reasoning may fail in creative writing or coding tasks. The AMAS (Adaptive Multi-Agent System) framework addresses this limitation by dynamically selecting the optimal collaboration topology for each individual query.

The AMAS Framework

AMAS moves beyond "one-size-fits-all" structures through a three-phase pipeline designed to optimize agent interaction on the fly.

The AMAS three-phase pipeline: Swarm Structure Initialization, Candidate Generation, and Dynamic Selection via the Graph Designer.
  1. Candidate Generation: The system first generates a set of $K$ high-performing candidate graph topologies using optimization techniques like the REINFORCE algorithm. These candidates represent diverse interaction patterns among agents.
  2. The Graph Designer: At the heart of AMAS is the "Graph Designer," a lightweight LLM specialized via Low-Rank Adaptation (LoRA). Given an input query and a candidate graph, it predicts a suitability score.
  3. Dynamic Selection: During inference, the Graph Designer evaluates the candidate graphs for the specific query and selects the top-ranked topology. The task is then executed following this optimized pathway.

By using a list-wise ranking loss during training, the Graph Designer learns the relative hierarchy of graph performance across different input types, rather than just binary classification.

Key Insights: The "No Silver Bullet" Discovery

A pivotal finding in the development of AMAS was that no single graph configuration consistently outperforms others across all samples. Even if "Graph A" has the highest average performance, it may be significantly outperformed by "Graph B" on specific subsets of problems.

Examples of different learned graph topologies (Candidate A and Candidate D) showing the structural diversity available for selection.

This discovery underscores that the "intelligence" of a multi-agent system resides not just in the individual agents, but in their organizational structure relative to the task context. AMAS treats this structure as a dynamic variable. Furthermore, the use of LoRA ensures that the Graph Designer adds minimal overhead—approximately 0.5% memory footprint—while maintaining latency comparable to static state-of-the-art (SOTA) systems.

Experimental Evaluation

The researchers evaluated AMAS across five diverse benchmarks using various backbone models, including Llama-3 (8B and 70B), GPT-3.5-turbo, Qwen-2.5, and DeepSeek-R1.

  • Benchmarks: Crossword (textual reasoning), Game-of-24 (math logic), MMLU (general knowledge), LLM-Eval-P (reasoning depth), and HumanEval (code generation).
  • Performance: AMAS consistently outperformed both single-agent (CoT, ToT, GoT) and multi-agent baselines (AutoGPT, AgentVerse, GPTSwarm). For instance, on the Crossword task with Llama-3 8B, AMAS achieved a score of 0.485, surpassing AgentVerse (0.452) and GPTSwarm (0.447).
  • Scalability: With a Llama-3 70B backbone, AMAS reached 0.847 accuracy on MMLU, demonstrating its ability to scale effectively with larger models.

Conclusion and Future Work

AMAS demonstrates that adaptive communication topologies significantly enhance the performance of LLM-based multi-agent systems. While highly effective, the authors note that future research should explore "ultra-large" models and more open-ended domains. Additionally, integrating safety-specific evaluations into the Graph Designer will be crucial for ensuring reliable agent behavior in complex, real-world applications.

Summary updated

BibTeX
@misc{leong2025amas,
  author = {Hui Yi Leong and Yuheng Li and Yuqing Wu and Wenwen Ouyang and Wei Zhu and Jiechao Gao and Wei Han},
  title = {{AMAS: Adaptively Determining Communication Topology for LLM-based Multi-Agent System}},
  year = {2025},
  eprint = {2510.01617},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.01617},
}
Figure