Research map / Multi-agent systems and games / LLM agents
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation
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.
Reframes LLM multi-agent topology design as conditional autoregressive graph generation, using a GRU-based node/edge generator trained with a two-stage curriculum to build task-specific agent graphs from scratch.
Problem
Existing automated MAS topology design starts from fixed template graphs and prunes them, causing redundant agent composition and limited extensibility to new agent roles.
Approach
Construct the collaboration graph node-by-node from scratch conditioned on the task query: a GRU node generator selects roles from an extensible role pool via metric-learning retrieval, and an edge GRU decides incoming links, trained with a cold-start on dense graphs then efficiency fine-tuning on pruned/simple graphs.
Key results
- ARG-Designer averages 92.78% across six benchmarks, versus 90.04% for the strongest learned baseline G-Designer.
- On AQuA, ARG-Designer reaches 86.45% accuracy, 4.85 points above G-Designer.
- On GSM8K, ARG-Designer uses about 4.1e6 tokens, roughly 50% fewer than G-Designer while scoring higher accuracy (94.40%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add efficiency fine-tuning phase (vs w/o fine-tune) (ablation) | ARG-Designer · MMLU | accuracy (%): 88.23 -> 89.54 | Helped |
| add efficiency fine-tuning phase (vs w/o fine-tune) (ablation) | ARG-Designer · GSM8K | accuracy (%): 94.70 -> 94.40 | Hurt |
| add efficiency fine-tuning phase (vs w/o fine-tune) (ablation) | ARG-Designer · Average over MMLU, GSM8K, HumanEval | accuracy (%): 91.28 -> 91.89 | Helped |
| remove task query embedding (ablation) | ARG-Designer · MMLU | accuracy (%): 89.54 -> 86.93 | Hurt |
| add historical role embedding (GRU_prev) (ablation) | ARG-Designer · Average over MMLU, GSM8K, HumanEval | accuracy (%): 90.64 -> 91.89 | Helped |
| use ARG-Designer autoregressive generation instead of G-Designer template approach | MAS topology design · Six benchmarks average | accuracy (%): 90.04 -> 92.78 | Helped |
| switch from G-Designer to ARG-Designer on AQuA | MAS topology design · AQuA | accuracy (%): 81.60 -> 86.45 | Helped |
| robustness to system prompt attack under ARG-Designer | MAS topology design · Robustness attack | performance degradation (%): 2.15 degradation | Helped |
Surprising findings
- The w/o fine-tune variant is highly competitive, scoring 91.28 average versus 91.89 for the full model, with fine-tuning mainly reducing token cost.
Limitations
- Relies on GPT-4o API backbone; results depend on proprietary model behavior.
- Training data constructed synthetically from successful configurations, with a limited number of training queries.
- Experiments limited to six benchmarks, mostly reasoning and math, with a single backbone.
Experimental setup
Datasets: MMLU GSM8K MultiArith SVAMP AQuA HumanEval
Benchmarks: MMLU GSM8K AQuA MultiArith SVAMP HumanEval
Metrics: accuracy Pass@1 token consumption
Baselines: CoT Self-Consistency Chain Tree Complete Graph Random Graph LLM-Debate AgentPrune AgentDropout G-Designer
Builds on: G-Designer AgentPrune AgentDropout GraphRNN LLM-Debate
Closest papers by meaning
- CARD: Towards Conditional Design of Multi-agent Topological Structures (ICLR 2026 · similarity 0.90)
- AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.90)
- MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.89)
- G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.89)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
- Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 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.