Source-linked AI summary
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation
Shiyuan Li, Yixin Liu, Qingsong Wen, Chengqi Zhang, Shirui Pan
TL;DR
Existing MAS topology-design methods modify fixed templates, limiting task-specific composition and structure. ARG-DESIGNER conditionally generates graphs from scratch by selecting agents, roles, and links autoregressively. Across six benchmarks, it reports state-of-the-art performance with superior token efficiency and extensibility.
Problem
Existing graph-learning methods commonly modify predefined templates with fixed agents and hard-coded structures, creating redundant compositions and limiting extensibility.
Method
ARG-DESIGNER learns P(G|Q, R) and autoregressively constructs a collaboration graph from scratch, jointly selecting agent roles and communication links conditioned on a natural-language task query.
Results
Across six benchmarks, ARG-DESIGNER consistently outperforms existing methods while achieving state-of-the-art performance and superior token efficiency.
Takeaways & Limitations
Task-specific autoregressive graph generation supports bespoke MAS topologies with dynamically chosen agent numbers, extensible roles, and communication links.
Takeaways & Limitations
Existing methods trained on fixed template graphs have limited extensibility as new agent functionalities emerge.
Abstract
from arXiv · showhide
Multi-agent systems (MAS) based on large language models (LLMs) have emerged as a powerful solution for dealing with complex problems across diverse domains. The effectiveness of MAS is critically dependent on its collaboration topology, which has become a focal point for automated design research. However, existing approaches are fundamentally constrained by their reliance on a template graph modification paradigm with a predefined set of agents and hard-coded interaction structures, significantly limiting their adaptability to task-specific requirements. To address these limitations, we reframe MAS design as a conditional autoregressive graph generation task, where both the system composition and structure are designed jointly. We propose ARG-Designer, a novel autoregressive model that operationalizes this paradigm by constructing the collaboration graph from scratch. Conditioned on a natural language task query, ARG-Designer sequentially and dynamically determines the required number of agents, selects their appropriate roles from an extensible pool, and establishes the optimal communication links between them. This generative approach creates a customized topology in a flexible and extensible manner, precisely tailored to the unique demands of different tasks. Extensive experiments across six diverse benchmarks demonstrate that ARG-Designer not only achieves state-of-the-art performance but also enjoys significantly greater token efficiency and enhanced extensibility. The source code of ARG-Designer is available at https://github.com/Shiy-Li/ARG-Designer.
Introduction
MAS collaboration topology determines how agents exchange information, yet existing automated methods largely modify fixed templates with predefined agents and hard-coded links. ARG-DESIGNER instead generates task-specific graphs incrementally, jointly designing agent composition and communication structure.
- MAS performance varies substantially with how inter-agent communication is architected, making task-specific topology design a critical challenge.
- Existing methods commonly adapt predefined templates containing fixed agent roles and hard-coded interaction structures through pruning or edge adjustment.Examples include AgentPrune, AgentDrop, and G-Designer.
- Template modification creates redundant compositions because initialized roles and dense connections may remain unnecessary for a specific task.The resulting redundancy can reduce efficiency and contribute to sub-optimal decision-making during execution.
- Autoregressive graph generation constructs collaboration graphs from scratch by progressively selecting appropriate agents, naturally avoiding redundant agent-role compositions.The approach is inspired by incrementally recruiting team members according to expertise, availability, and evolving task needs.
- ARG-DESIGNER conditions generation on a natural-language task query and iteratively produces agents with communication links to previously generated nodes.The model is designed to improve flexibility and scalability across agent counts, roles, and interactions.
Problem Formulation
The paper models MAS topology design as conditional autoregressive graph generation: a graph is built node by node, with each step selecting an agent role and its links. This factorization makes task-adaptive composition and structure generation tractable.
- MAS as a Collaboration Graph: A MAS is represented as a directed acyclic collaboration graph whose nodes are role-specific LLM agents and whose edges define directed communication pathways.Each agent also maintains an internal state representing memory of prior actions and interactions.
- MAS Topology Design as a Graph Generation Task: Topology design learns a conditional distribution P(G|Q, R) over collaboration graphs given a task query Q and an extensive role pool R.The formulation replaces navigation through graph space with a learned generative model.
- Autoregressive Graph Generation: Autoregressive factorization decomposes whole-graph generation into a tractable sequence of conditional steps that incrementally add nodes and edges.This converts an intractable joint generation problem into manageable conditional decisions.
- Autoregressive Graph Generation: At each generation step, node generation predicts the next agent role while edge generation establishes its connections to existing agents.Sampling a special END token allows the model to determine the total number of agents dynamically.
- Design Advantages: The resulting paradigm supports task-adaptive construction, dynamic role composition, extensibility, and scalable generation compared with fixed templates.Table 1 compares degrees of freedom across MAS design paradigms using full, partial, and no-support indicators.
ARG-DESIGNER for MAS Topology Design
ARG-DESIGNER uses a hierarchical autoregressive architecture to generate MAS topologies by selecting agent roles and communication links step by step. It trains on exploration and efficiency data, then generates graphs from task queries until an END token or agent limit is reached.
- Model Architecture: ARG-DESIGNER separates topology generation into a node generator for agent roles and an edge generator for communication links.The architecture uses GRU-based sequence modeling to capture generation history and long-range dependencies.
- Model Architecture: The node generator combines task information with generation history to predict the next role from an extensible role-embedding matrix.A metric-learning retrieval mechanism allows new roles to be added at inference time without retraining.
- Model Architecture: At each step, the edge generator evaluates existing nodes and predicts whether to create incoming communication links to the newly generated agent.Edge decisions are conditioned on the node-level state and previous edge decisions.
- Training and Inference Strategy: The training pipeline uses curriculum learning with an exploration dataset for functionally correct, diverse graphs and an efficiency dataset for simpler communication-efficient graphs.Efficiency data combines minimal verified configurations, pruned successful graphs, and replay data; the loss combines node and edge generation terms.
- Training and Inference Strategy: During inference, ARG-DESIGNER generates a graph node by node, sampling roles and edges until it selects END or reaches the maximum agent count.The process starts from an empty graph and returns the generated collaboration graph without ground-truth guidance.
Experiments
ARG-DESIGNER is evaluated across six benchmarks against multiple baselines, with results emphasizing performance, token efficiency, robustness, and extensibility.
- Evaluation Setup: ARG-DESIGNER is evaluated on MMLU, GSM8K, MultiArith, SVAMP, AQuA, and HumanEval, using training batches of B ∈{40, 60} queries.The benchmarks cover general reasoning, mathematical reasoning, and code generation.
- Performance Comparison: ARG-DESIGNER achieves the best performance across all six benchmarks and surpasses G-Designer by 4.85% on AQuA, reaching 86.45% accuracy.It also improves over debate-based methods by 8.8% on AQuA and 2.66% on GSM8K.
- Token Efficiency: ARG-DESIGNER balances performance and efficiency by generating task-specific topologies that avoid unnecessary complexity and reduce token consumption.Figure 3 presents the performance–token-cost trade-off.
- Robustness and Extensibility: Under a system prompt attack, ARG-DESIGNER shows the least performance degradation, at 2.15%, indicating robustness relative to fixed and naive topologies.The study also introduces new roles without retraining and demonstrates a legal-task topology centered on a newly added Lawyer role.
Related Work
Prior MAS topology methods rely on graph-generation orderings or modifications of predefined structures, whereas ARG-DESIGNER constructs task-specific collaboration graphs from scratch.
- Autoregressive Graph Generation: Autoregressive graph models factorize graph probability into conditional sequences for nodes and edges, but their performance depends critically on node ordering.BFS, ad-hoc orderings, and canonical ordering schemes each introduce different modeling challenges.
- Existing MAS Topology Design: Existing MAS topology methods include fixed chains, trees, fully connected graphs, pruning-based approaches, and graph autoencoders for task-adaptive structures.These approaches differ in communication structure but generally begin from predefined agent sets or templates.
- Template Graph Modification: Template graph modification starts from fixed agents and hard-coded interactions, then adapts topology through operations such as edge reweighting or pruning.The shared paradigm is described as limiting system composition and extensibility.
- Autoregressive Graph Generation: ARG-DESIGNER departs from template modification by jointly selecting agents and communication links while generating the collaboration graph from scratch.Its design dynamically selects roles from an extensible pool in a task-adaptive manner.
Conclusion
The paper reformulates MAS topology design as autoregressive graph generation and introduces ARG-DESIGNER to construct bespoke collaboration graphs from natural-language task queries.
- Conclusion: ARG-DESIGNER dynamically determines agent numbers, selects roles from an extensible pool, and establishes communication links tailored to task demands.The approach addresses redundant composition and limited extensibility in template graph modification methods.
- Conclusion: Across six benchmarks, ARG-DESIGNER consistently outperforms existing methods while maintaining superior token efficiency.The conclusion reports state-of-the-art performance together with improved efficiency.