Research map / Multi-agent systems and games / LLM agents
MAS-Architect: Declarative Multi-Agent System Design via Separation of Concerns
Jing Huang, Lidong Zhang, Mutian Bao, Yadong Li, Xingzhong Xu, Jinjian Zhang, Jie Liu, Ming Kong, Qiang Zhu
Read the paperCheck your idea against 137,217 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 MAS-Architect, a framework that automates multi-agent system design through a code-based declarative approach that separates system topology from individual agent implementation.
Problem
Automated design of multi-agent systems (Auto-MAS) for complex reasoning often suffers from structural rigidity and entangles topology design with the implementation of individual agents.
Approach
Separates MAS topology design from agent implementation using a code-based declarative formulation, in contrast to existing Auto-MAS approaches that entangle the two.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SFT architectural distillation (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · MATH | accuracy: 82.1 -> 83.3 (+1.2) | Improved |
| add SFT architectural distillation (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · GPQA | accuracy: 55.1 -> 56.1 (+1.0) | Improved |
| add SFT architectural distillation (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · HotpotQA | accuracy: 45.5 -> 47.9 (+2.4) | Improved |
| add RL architectural exploration after SFT (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · MATH | accuracy: 83.3 -> 89.5 (+7.4) | Improved |
| add RL architectural exploration after SFT (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · GPQA | accuracy: 56.1 -> 64.1 (+9.0) | Improved |
| add RL architectural exploration after SFT (ablation) | Qwen3-4B-Instruct-2507 MAS-Architect · HotpotQA | accuracy: 47.9 -> 61.5 (+16.0) | Improved |
| add SFT architectural distillation (ablation) | Qwen3-30B-A3B-Instruct-2507 MAS-Architect · MATH | accuracy: 88.7 -> 89.8 (+1.1) | Improved |
| add SFT architectural distillation (ablation) | Qwen3-30B-A3B-Instruct-2507 MAS-Architect · GPQA | accuracy: 70.2 -> 71.7 (+1.5) | Improved |
| add SFT architectural distillation (ablation) | Qwen3-30B-A3B-Instruct-2507 MAS-Architect · HotpotQA | accuracy: 65.5 -> 67.2 (+1.7) | Improved |
| add RL architectural exploration after SFT (ablation) | Qwen3-30B-A3B-Instruct-2507 MAS-Architect · MATH | accuracy: 89.8 -> 92.5 (+3.8) | Improved |
| add RL architectural exploration after SFT (ablation) | Qwen3-30B-A3B-Instruct-2507 MAS-Architect · HotpotQA | accuracy: 67.2 -> 70.8 (+5.3) | Improved |
| use MAS-Architect on unseen GSM-Hard instead of MAS-GPT | Qwen3-4B-Instruct-2507 · GSM-Hard | accuracy: 39.9 -> 64.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MAS-GPT — MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems (ICML 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.91)
- Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.90)
- The Illusion of Multi-Agent Advantage (NeurIPS 2026 · similarity 0.90)
- MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.89)
- EvoMAS: Evolutionary Generation of Multi-Agent Systems (ICML 2026 · similarity 0.88)
- MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems (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.
