Not Every Agent Needs the Strongest Model: Unified Prompt-Model Optimization for Heterogeneous LLM-based Multi-Agent Systems

Mingxiu Cai; Haoyang Xian; Xinyu Ma; Yutong Wang; Zhenxi Song; Jing Li; Xuebo Liu

cs.MAOct 22, 2026

Not Every Agent Needs the Strongest Model: Unified Prompt-Model Optimization for Heterogeneous LLM-based Multi-Agent Systems

Authors: Mingxiu Cai, Haoyang Xian, Xinyu Ma, Yutong Wang, Zhenxi Song, Jing Li, Xuebo Liu

Organizations: Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen

Abstract

Large language model (LLM)-based multi-agent systems (MAS) rely on role-specific prompts and execution models to solve complex tasks collaboratively. However, prompt optimization typically keeps model assignments fixed, leaving differences in role requirements and their associated inference costs outside the search. We present UniMASO, a Unified framework for heterogeneous Multi-Agent System Optimization through the joint adaptation of prompts and model assignments. UniMASO evaluates candidate updates through local, downstream, and global feedback, using cost-aware Pareto search to retain useful quality-cost trade-offs. Topology-aware co-evolution then coordinates these updates according to agent roles and dependencies. Cross-level trajectory memory further supplies historical successes and failures to guide subsequent optimization. Experiments on five benchmarks and five execution topologies show that UniMASO achieves competitive task performance at substantially lower deployment cost than the strongest homogeneous configurations. Under centralized execution, it improves the average score from 69.79 to 70.40 while reducing token-weighted deployment cost by 85.2% relative to DeepSeek-V4-Pro with MASPO. These results support joint prompt-model adaptation as an effective approach to efficient multi-agent collaboration. The code is available at https://github.com/Achilles008/UniMASO_main.

Figures & tables

Method	Prompt	Model
MIPRO ( Opsahl-Ong et al., 2024 )	✓	✗
GPTSwarm ( Zhuge et al., 2024 )	✓	✗
MASS ( Zhou et al., 2026 )	✓	✗
MASPOB ( Hong et al., 2026 )	✓	✗
MASPO ( Wang et al., 2026 )	✓	✗
MasRouter ( Yue et al., 2025 )	✗	✓

Table 1: Comparison of prompt and model optimization paradigms.

Figure 1: Homogeneous vs. Heterogeneous MAS.

Figure 2: Overview of UniMASO. Cost-aware multi-agent optimization proposes and evaluates prompt candidates from execution traces; heterogeneous execution runs the current prompt–model configuration on the realized topology; topology-aware adaptation reconciles role, dependency, and system views to update model assignments; and trajectory memory records cross-level feedback and retrieves historical evidence for subsequent rounds.

Model	Method	AGIEval		GPQA		BFCL		LiveCodeBench		HotpotQA		Avg.
		Acc.	Cost	Acc.	Cost	Acc.	Cost	pass@1	Cost	EM	Cost	Result	Cost
Qwen3.5-4B	Vanilla MAS	62.11	7.3	67.17	5.3	88.40	4.8	18.32	3.8	58.00	5.8	58.80	5.38
	+TPE	59.77	8.2	66.67	6.1	88.80	5.4	19.08	4.2	58.20	6.5	58.50	6.08
	+SPO	63.28	8.3	69.70	5.9	88.60	5.5	20.61	4.3	58.00	6.6	60.03	6.12
	+MASPOB	63.67	8.2	68.69	5.8	89.40	5.3	19.85	4.3	58.60	6.5	60.04	6.02
	+MASPO	63.28	8.2	69.19	5.9	89.60	5.5	21.37	4.4	58.40	6.4	60.36	6.08

Table 2: Task quality and deployment cost under the centralized topology. Prompt-only baselines use homogeneous models; UniMASO jointly optimizes prompts and model assignments.

Reference: Proposed Framework
Method	AGIEval		GPQA		BFCL		LiveCodeBench		HotpotQA		Avg.
	Acc.	Cost	Acc.	Cost	Acc.	Cost	pass@1	Cost	EM	Cost	Result	Cost

UniMASO (Full)	82.42	56.86	78.28	33.02	92.40	12.21	35.11	13.27	63.80	39.26	70.40	30.92
(I) Component Ablations

Table 3: UniMASO ablations and parameter sensitivity under centralized execution.

Target	Memory	Acc.	Cost
AGIEval-MATH	None	80.08	51.35
	Own	82.42	56.86
	Other	81.64	52.93
GPQA-Diamond	None	76.77	31.58
	Own	78.28	33.02
	Other	77.78	29.43

Table 4: Cross-task memory transfer.

Method	Coordination		Efficiency
	Comm. $\uparrow$	Plan. $\uparrow$	Cost-of-Pass $\downarrow$
Centralized MAS	68.45	64.39	1.154
+TPE	71.81	66.86	1.357
+SPO	69.90	64.15	1.271
+MASPOB	71.54	67.08	1.245
+MASPO	72.84	68.41	1.241

Table 5: Coordination and efficiency.

Target dataset	Method	Opt. pool	Accuracy $\uparrow$ (%)	Cost $\downarrow$ (M)
AGIEval-MATH	Vanilla MAS	None	76.17	217.3
	MASPO	AGIEval-MATH	78.12	249.6
	UniMASO (in-domain)	AGIEval-MATH	82.42	56.9
	UniMASO (cross-dataset)	MATH-500	81.25	58.7
MATH-500	Vanilla MAS	None	79.20	262.9
	MASPO	MATH-500	84.80	339.5

Table 6: Cross-dataset transfer of UniMASO between AGIEval-MATH and MATH-500. The Vanilla MAS and MASPO baselines use DeepSeek-V4-Pro for all task agents. Accuracy and deployment cost are measured on the target evaluation set.

Method	AGIEval-MATH			GPQA-Diamond			BFCL-V3			HotpotQA			LiveCodeBench
	Input	Output	Cost	Input	Output	Cost	Input	Output	Cost	Input	Output	Cost	Input	Output	Cost
UniMASO	5.92	0.085	51.9	4.14	0.072	37.0	1.93	0.065	19.0	4.89	0.070	42.8	6.45	0.077	55.6
MASPO	7.48	0.103	1188.4	5.21	0.086	843.1	2.37	0.074	421.2	6.19	0.082	979.9	8.17	0.094	1278.0

Table 7: Optimization-side LLM usage and cost of MASPO and UniMASO under centralized execution. Input and output tokens and relative weighted-token cost are reported in millions (M). The accounting includes optimizer and evaluator calls and UniMASO’s Co-Evolver calls.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Samples	Metric
AGIEval-MATH	256	Accuracy
GPQA-Diamond	198	Accuracy
BFCL-V3	500	Accuracy
LiveCodeBench-v6	131	Pass@1
HotpotQA	500	Exact match
Total	1,585	–

Appendix

Table 8: Final-evaluation sample counts and metrics by benchmark.

\mathcal{C}_{i,r,t}\leftarrow\mathcal{E}_{\mathrm{loc}}(\mathcal{B}_{i},\mathcal{T}_{i,r,t},\mathcal{G},\mathcal{M}).

Appendix

Algorithm 1 UniMASO: Joint Prompt–Model Optimization

Topology	Model	AGIEval		GPQA		BFCL		LiveCodeBench		HotpotQA		Avg.
		Acc.	Cost	Acc.	Cost	Acc.	Cost	p@1	Cost	EM	Cost	Result	Cost
Independent	Qwen3.5-4B	62.11	4.9	66.16	4.6	86.60	2.5	19.08	2.8	57.80	3.6	58.35	3.68
	Qwen3.5-9B	59.38	7.6	68.69	8.5	87.80	6.6	23.66	5.7	57.60	7.0	59.43	7.07
	Qwen3.5-27B	69.92	35.9	68.69	34.1	93.40	23.1	29.01	21.6	59.60	30.9	64.12	29.12
	DeepSeek-V4-Flash	67.19	43.9	69.70	27.1	93.00	28.1	36.64	29.4	56.40	59.0	64.59	37.50
	DeepSeek-V4-Pro	69.53	140.8	70.20	83.8	93.40	83.7	40.46	68.9	58.20	178.8	66.36	111.20

Appendix

Table 9: Performance comparison of UniMASO against homogeneous model assignments across five execution topologies. For each topology, we compare five homogeneous configurations with the heterogeneous configuration selected by UniMASO, reporting both task quality and deployment cost.

Model	Capability	$c^{\mathrm{in}}$	$c^{\mathrm{out}}$
Qwen3.5-4B	Very weak	1	4
Qwen3.5-9B	Medium	2	8
Qwen3.5-27B	Strong	8	32
DeepSeek-V4-Flash	Very strong	24	72
DeepSeek-V4-Pro	Extremely strong	72	216

Appendix

Table 10: Model capability levels and relative token-cost coefficients.

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,032 papers

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems
May 7, 2026Zhexuan Wang, Xuebo Liu, Li Wang +4Multi-Agent LLM Systems Multi-Agent System Optimization
Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen, China.
MAS-PromptBench: When Does Prompt Optimization Improve Multi-Agent LLM Systems?
Jun 22, 2026Juyang Bai, Laixi ShiMulti-Agent LLM Systems Multi-Agent System Optimization
Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA.
MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks
Mar 3, 2026Zhi Hong, Qian Zhang, Jiahang Sun +5Multi-Agent LLM Systems Multi-Agent System Optimization
The Chinese University of Hong Kong, Shenzhen · South China University of Technology · Ritsumeikan University +1

May 7, 2026J/K move · Enter open · S save

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

Zhexuan Wang, Xuebo Liu, Li Wang, Zifei Shan+3

Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen, China.

Multi-Agent LLM Systems Multi-Agent System Optimization Multi-Agent Collaboration Prompt Optimization

Open paper AskPDF arXiv

Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at https://github.com/wangzx1219/MASPO.

Abstract

Multi-Agent LLM Systems

Explore similar work

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

MAS-PromptBench: When Does Prompt Optimization Improve Multi-Agent LLM Systems?

MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks