Research map / Multi-agent systems and games / LLM agents
G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes G-Designer, which uses a variational graph auto-encoder over an agent network plus a task virtual node to generate query-adaptive, sparse LLM multi-agent communication topologies, trained via policy gradient on a small query set.
Problem
Choosing an effective, token-efficient, and robust communication topology for an LLM multi-agent system is hard because existing topologies are static and input-independent.
Approach
Model agents as nodes of a graph (profile/role/tool embeddings) with a task-specific virtual node, then use a VGAE-based encoder-decoder with a nuclear-norm sparsity regularizer and anchor regularization to decode a task-adaptive, sparse communication graph, optimized by policy gradient on task utility.
Key results
- Achieves 84.50% on MMLU and 89.90% pass@1 on HumanEval, with average 89.84 across six benchmarks, best in five of six.
- On GSM8K, reaches 95.0 accuracy versus 88.1 for DyLAN, with overall token use 8.5e6 versus 2.2e7 for DyLAN.
- Under a system prompt attack on one of five agents on MMLU, accuracy drops only 0.3% (84.5 to 84.2), while GPTSwarm drops 0.3% and DyLAN 6.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use G-Designer instead of GPTSwarm | GPT-4 multi-agent system · MMLU | accuracy: 83.98 -> 84.50 | Helped |
| Use G-Designer instead of PHP | GPT-4 · GSM8K | accuracy: 95.50 -> 95.07 | Hurt |
| Use G-Designer instead of PHP | GPT-4 multi-agent system · MultiArith | accuracy: 98.10 -> 98.30 | Helped |
| Use G-Designer instead of LLM-Debate | GPT-4 multi-agent system · SVAMP | accuracy: 90.56 -> 91.85 | Helped |
| Use G-Designer instead of Complete Graph | GPT-4 multi-agent system · AQuA | accuracy: 79.21 -> 79.47 | Helped |
| Use G-Designer instead of DyLAN | GPT-4 multi-agent system · HumanEval | pass@1: 89.70 -> 89.90 | Helped |
| Add sparsity regularization (ablation) | G-Designer · MMLU clean | accuracy: 84.1 -> 84.5 | Helped |
| Add sparsity regularization (ablation) | G-Designer · GSM8K clean | accuracy: 94.4 -> 95.0 | Helped |
| Add anchor structure (ablation) | G-Designer · MMLU clean | accuracy: 84.0 -> 84.5 | Helped |
| Add anchor structure (ablation) | G-Designer · GSM8K clean | accuracy: 94.7 -> 95.0 | Helped |
| Add NodeEncoder (ablation) | G-Designer · MMLU clean | accuracy: 83.2 -> 84.5 | Helped |
| Add NodeEncoder (ablation) | G-Designer · GSM8K clean | accuracy: 92.8 -> 95.0 | Helped |
Surprising findings
- Under the paper's own adversarial setup, GPTSwarm shows a comparably small 0.3% drop, so robustness gains are not unique to G-Designer on that metric.
- Replacing the anchor chain topology with an empty anchor is less harmful than removing the task virtual node, indicating the task node drives most of the task-adaptivity.
Limitations
- Relies on the GPT-4 family and API-based utilities; results depend on proprietary models.
- Only a small number of training queries (40 or 80) are used, and test sets for MMLU (153) and HumanEval (164) are small, so variance may be high.
- Reported ablation and robustness numbers are on MMLU and GSM8K only, and the attack setting is a single simulated system prompt attack.
Experimental setup
Datasets: MMLU GSM8K MultiArith SVAMP AQuA HumanEval
Benchmarks: MMLU GSM8K MultiArith SVAMP AQuA HumanEval
Metrics: accuracy pass@1 prompt token consumption wall-clock time accuracy drop under prompt attack
Baselines: CoT ComplexCoT Self-Consistency PHP Chain Star Tree Complete Graph Random Graph AutoGen MetaGPT LLM-Debate
Builds on: variational graph auto-encoder GPTSwarm DyLAN chain-of-thought
Closest papers by meaning
- Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · similarity 0.89)
- Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.89)
- Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck (AAAI 2024 · similarity 0.88)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.88)
- AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.