Research map / Multi-agent systems and games / LLM agents

Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

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 MASS, a multi-stage optimizer that interleaves block-level prompt optimization, pruned topology search, and workflow-level prompt optimization to design multi-agent systems on Gemini 1.5 Pro/Flash.

Problem

Designing effective multi-agent systems requires choosing both agent prompts and topologies, a combinatorial and prompt-sensitive search that prior automated design methods handle poorly by ignoring prompt optimization.

Approach

Analyze MAS design factors and find prompts are the most influential component and only a small fraction of topology blocks help; then interleave local-to-global optimization (block prompt warm-up, pruned topology search using influence-weighted rejection sampling, workflow-level prompt tuning).

Key results

• MASS reaches 78.79 average on Gemini 1.5 Pro across 8 tasks versus 70.26 for Multi-Agent Debate and 69.72 for ADAS.
• On Gemini 1.5 Flash, MASS averages 74.30 versus 65.91 for Multi-Agent Debate, with MATH rising from 66.7 (CoT) to 81.0.
• On Mistral-Nemo-12B, MASS averages 55.9 versus 46.9 for the best baseline (Multi-Agent Debate) over four tasks.

Surprising findings

• Only a small fraction of topology building blocks improve performance; e.g., on HotpotQA only debate gives gains (about 3%) while others fail or degrade results.
• Multi-agent debate and self-refine with default prompts saturate quickly as inference cost scales, while prompt-optimized single agents scale better per token.

Limitations

• Comparison with AFlow is not fully fair: AFlow uses Claude 3.5 Sonnet as optimizer and some AFlow runs timed out, so results are reference-only.
• Experiments use random subsets of validation and test splits (e.g., 100-200 test items), and the topology space is fixed and hand-designed.
• Relies on proprietary Gemini models; generalization is checked only on Claude 3.5 Sonnet and Mistral Nemo for limited tasks.

Experimental setup

Datasets: MATH DROP HotpotQA MuSiQue 2WikiMultiHopQA MBPP HumanEval LiveCodeBench

Benchmarks: MATH DROP HotpotQA MuSiQue 2WikiMultiHopQA MBPP HumanEval LiveCodeBench

Metrics: accuracy F1 pass@1

Baselines: CoT CoT-SC Self-Refine Multi-Agent Debate ADAS AFlow GPTSwarm MIPROv2

Builds on: MIPRO DSPy ADAS AFlow self-refine multi-agent debate

Closest papers by meaning

• AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (NeurIPS 2026 · similarity 0.90)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.90)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.89)
• MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks (ICML 2026 · similarity 0.89)
• MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.89)
• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.89)