Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation

Conference: AAAI 2026 Oral
arXiv: 2507.18224
Code: https://github.com/Shiy-Li/ARG-Designer
Area: Graph Learning Keywords: Multi-Agent Topology Design, Autoregressive Graph Generation, Collaboration Graph, Curriculum Learning, Scalable Agents

TL;DR

This paper proposes ARG-Designer, which reformulates multi-agent system topology design as a conditional autoregressive graph generation task. Rather than pruning from template graphs, the model incrementally generates agent nodes and communication edges from scratch. ARG-Designer achieves state-of-the-art performance across 6 benchmarks (average 92.78%), reduces token consumption by approximately 50% compared to G-Designer, and supports role expansion without retraining.

Background & Motivation

Background: The effectiveness of LLM-based multi-agent systems critically depends on the collaborative topology — how agents are organized and exchange information. Automated topology design has become a research focus, with representative works including AgentPrune (edge pruning), AgentDropout (stochastic dropping), and G-Designer (graph autoencoder learning).

Limitations of Prior Work: Existing methods follow a "template graph modification" paradigm — starting from a predefined fully-connected or dense template and adapting it to the task via edge reweighting or pruning. Two key limitations arise: (1) Redundant composition: the template predefines all possible agent roles, so even after pruning, irrelevant agents or edges may remain; (2) Limited scalability: models are trained on fixed templates and cannot generalize to newly added agent roles or dynamically changing agent pools.

Key Challenge: The search space of the template modification paradigm is constrained by the predefined template, making truly task-tailored topologies unattainable; yet expanding the template to cover all possible roles is prohibitively expensive.

Goal: How to construct, from scratch, a customized multi-agent topology containing only the necessary agents and optimal communication links?

Key Insight: Drawing an analogy to real-world team formation — rather than hiring all possible members and then downsizing, one incrementally recruits suitable members according to task requirements. This motivates the autoregressive graph generation paradigm: iteratively adding nodes and edges until the topology is complete.

Core Idea: Transform multi-agent topology design from "template modification" to "conditional autoregressive graph generation," constructing from scratch a collaboration graph that is optimal in terms of agent count, roles, and connectivity.

Method

Overall Architecture

The MAS is modeled as a directed acyclic graph \(\mathcal{G} = (\mathcal{V}, \mathcal{E})\), where nodes are role-assigned agents and edges are communication links. ARG-Designer is a conditional generative model \(P(\mathcal{G}|\mathcal{Q}, \mathcal{R})\), conditioned on task query \(\mathcal{Q}\) and role pool \(\mathcal{R}\), which autoregressively factorizes into a sequence of node and edge generation steps: $\(P(\mathcal{G}|\mathcal{Q}, \mathcal{R}) = \prod_{i=1}^{|\mathcal{V}|} P(v_i|\mathcal{G}_{<i}, \mathcal{Q}, \mathcal{R}) \cdot \prod_{j=1}^{i-1} P(e_{ji}|v_i, \mathcal{G}_{<i}, \mathcal{Q})\)$

Key Designs

1. Node Generator:

   • Function: At each step, selects the role of the next agent or outputs END to terminate generation.
   • Mechanism: A GRU aggregates the historical embeddings of previously generated agents \(\mathbf{f}_{\text{hist}}^{(i)} = \text{GRU}_{\text{prev}}([\mathbf{z}_{r_1}, \ldots, \mathbf{z}_{r_{i-1}}])\), which is fused with the task embedding \(\mathbf{f}_{\mathcal{Q}}\) via dynamic gating to produce a context embedding, followed by a GRU hidden state update. Critically, role selection employs metric learning (dot-product) rather than a fixed classifier — the hidden state is projected into a "node intent" embedding and dot-producted against the role embedding matrix to obtain selection probabilities.
   • Design Motivation: The metric learning mechanism is key to scalability — new roles can be appended as embedding rows without modifying or retraining the model, naturally supporting dynamic expansion of the role pool.

2. Edge Generator:

   • Function: Determines from which existing agents \(v_j\) \((j < i)\) the newly added agent \(v_i\) should receive information.
   • Mechanism: An edge GRU is initialized with the final hidden state of the node GRU, and iterates over existing nodes to predict the existence probability of each edge via Sigmoid: \(P(e_{j,i}=1) = \text{Sigmoid}(s_{\text{edge}}^{(i,j)})\).
   • Design Motivation: Conditioning edge generation on the node generation history and full context enables the model to learn complex structural dependencies, rather than treating each edge independently.

3. Curriculum Learning Training Strategy:

   • Function: Two-stage training — first learn to "generate correct and valid topologies," then learn to "generate compact and efficient topologies."
   • Mechanism: Stage 1 (Exploration): Complex configurations (many agents, dense connections) are used to generate successful topologies as training data \(\mathcal{D}_{\text{exp}}\), allowing the model to learn basic collaborative patterns. Stage 2 (Efficiency): Three data sources are mixed — efficient topologies from simple configurations \(\mathcal{D}_{\text{simple}}\), compact topologies obtained by systematically pruning the dense graphs from Stage 1 that remain successful \(\mathcal{D}_{\text{pruned}}\), and replay data \(\mathcal{D}_{\text{replay}}\) to prevent forgetting.
   • Design Motivation: Directly learning compact topologies suffers from a cold-start problem (the model has no prior sense of "correct" collaboration); learning dense-first then compact is a natural curriculum, and efficiency fine-tuning reduces token consumption by 30–34% without sacrificing performance.

Loss & Training

The overall objective is the negative log-likelihood loss: $\(\mathcal{L}_{\text{total}} = \alpha \cdot \mathcal{L}_{\text{node}} + (1-\alpha) \cdot \mathcal{L}_{\text{edge}}\)$ with \(\alpha=0.2\) (edge generation is weighted more heavily). Effective models can be trained with as few as 40–60 queries. Teacher Forcing is applied to accelerate training.

Key Experimental Results

Main Results

Accuracy comparison across 6 benchmarks (GPT-4o as the underlying LLM):

Method	MMLU	GSM8K	AQuA	MultiArith	SVAMP	HumanEval	Avg.
Vanilla (Single Agent)	80.39	82.30	71.06	93.09	86.55	71.39	80.80
LLM-Debate	84.96	91.40	77.65	96.36	90.11	84.70	87.53
AgentPrune	85.07	91.10	80.51	94.65	90.58	86.75	88.09
G-Designer	86.92	93.80	81.60	96.50	93.10	88.33	90.04
ARG-Designer	89.54	94.40	86.45	98.93	95.63	91.74	92.78

Ablation Study

Configuration	MMLU	GSM8K	HumanEval	Avg.
ARG-Designer	89.54	94.40	91.74	91.89
w/o fine-tune	88.23	94.70	90.91	91.28
w/o task emb.	86.93	93.10	89.26	89.76
w/o hist. emb.	88.23	93.60	90.08	90.64

Key Findings

• ARG-Designer achieves state of the art on all 6 benchmarks, outperforming G-Designer by 2.74% and Vanilla single-agent by 11.98% on average.
• Token efficiency: ARG-Designer consumes only 4.1M tokens on GSM8K, approximately 50% fewer than G-Designer — because topologies generated from scratch are more compact and avoid the redundancy inherent in template modification.
• Task embedding is the most critical component: removing it causes an average drop of 2.13%, demonstrating that conditioning on the task query is central to generating customized topologies.
• Robustness: under prompt injection attacks, performance drops by only 2.15% (the lowest among compared methods), as the generated topologies naturally exhibit distributed risk and redundant communication paths.
• Scalability validation: adding a "lawyer" role without retraining, the model correctly identifies its relevance to legal questions and dynamically generates a lawyer-centric collaboration graph.

Highlights & Insights

• The paradigm shift from "template modification" to "autoregressive generation" is the core contribution of this paper. This is not merely a methodological innovation but a reconceptualization at the problem formulation level — treating MAS topology design as conditional graph generation rather than graph editing, opening up a vastly larger design space.
• The metric learning approach to role selection is particularly elegant: while the conventional approach uses a fixed-dimension classifier (number of roles = number of classes), ARG-Designer uses similarity matching in embedding space, enabling the role pool to be dynamically extended at inference time.
• The curriculum learning strategy is worth generalizing: first learn "what works" (dense + correct), then learn "what is efficient" (compact + correct), with replay data between stages to prevent forgetting. This paradigm is applicable to other "correctness-first, efficiency-second" learning scenarios.

Limitations & Future Work

• Node generation ordering has a known influence on autoregressive models — the paper does not thoroughly discuss the impact of different ordering strategies, which is a well-recognized issue in the graph generation literature.
• Training data is generated through trial-and-error (executing topologies and checking success), making data construction costly (requiring large numbers of LLM calls); the paper trains on only 40–60 queries, and scalability to larger settings remains unexplored.
• Experiments use only GPT-4o as the underlying LLM; performance in heterogeneous LLM settings (different nodes using LLMs of varying capability) is not investigated.
• The number of communication rounds \(K=3\) is a fixed hyperparameter with no adaptive mechanism.

Related Work & Insights

• vs. G-Designer: G-Designer employs a graph autoencoder to learn topologies on predefined templates; ARG-Designer generates topologies autoregressively from scratch. The former is constrained by the template and cannot alter agent count or role set, while the latter can dynamically determine both.
• vs. AgentPrune/AgentDropout: These methods perform "subtraction" (removing edges/agents from a dense graph); ARG-Designer performs "addition" (constructing from an empty graph). Addition more naturally avoids redundancy.
• vs. BAMAS: BAMAS focuses on budget-constrained LLM selection and topology selection (from 4 predefined options); ARG-Designer supports unconstrained generation of arbitrary topology structures. The two are complementary — BAMAS can handle budget-constrained LLM selection upstream, while ARG-Designer generates the optimal topology downstream.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The paradigm shift (template modification → autoregressive generation) is a significant conceptual contribution; the metric learning design for scalable role selection is elegant.
• Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive evaluation across 6 benchmarks with ablations, robustness analysis, scalability tests, and case studies.
• Writing Quality: ⭐⭐⭐⭐⭐ Problem formulation is clear; comparisons with existing paradigms are intuitively illustrated; the logical flow is coherent.
• Value: ⭐⭐⭐⭐⭐ Opens a new direction for automated multi-agent topology design with an elegant framework and strong empirical support.

Related Papers

• [AAAI 2026] A Graph-Theoretical Perspective on Law Design for Multiagent Systems
• [ACL 2026] CIA: Inferring the Communication Topology from LLM-based Multi-Agent Systems
• [ICML 2026] RADAR: Redundancy-Aware Diffusion for Multi-Agent Communication Structure Generation
• [ICLR 2026] Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration
• [ACL 2026] Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs