MASPO: A New Framework for Joint Prompt Optimization in Multi-Agent Systems

Title: MASPO: A New Framework for Joint Prompt Optimization in Multi-Agent Systems Abstract: The performance of multi-agent systems depends on the quality of role prompts, but joint optimization across agents faces the challenge of misalignment between local and global objectives. MASPO achieves an average improvement of 2.9 percentage points across 6 tasks through a joint evaluation mechanism and data-driven evolutionary beam search, and has been accepted by ICML 2026. This thread will introduce the core content of the framework, including background, methods, and experimental results, in separate floors.

The Rise and Challenges of Multi-Agent Systems

Large language model-based multi-agent systems (MAS) have become powerful tools for solving complex collaborative tasks, applied in fields such as software development and scientific research. Prompts are the "soul" of MAS, defining agents' identities, capabilities, and interaction methods, directly affecting professionalism, collaboration fluency, and overall system performance. However, joint prompt optimization across agents faces three major dilemmas:

1. Local-global misalignment: Optimizing a single agent's prompt may harm overall performance (e.g., dominant agents suppress others);
2. High-dimensional search space: The combination space of prompts expands exponentially with the number of agents, making manual tuning impractical;
3. Evaluation difficulty: Open tasks lack clear ground-truth, making it hard to determine optimization directions.

Core Innovations of the MASPO Framework

To address the above challenges, the MASPO (Multi-Agent System Prompt Optimization) framework proposes two core innovations:

Joint Evaluation Mechanism

Unlike traditional methods that only evaluate the local performance of individual agents, MASPO uses "whether the prompt can promote the success of downstream agents" as the standard, bridging the gap between local interactions and global results. It does not require ground-truth and is suitable for open tasks.

Data-Driven Evolutionary Beam Search

To handle the high-dimensional space, MASPO adopts an evolutionary beam search strategy:

1. Population initialization: Generate candidate populations through mutation starting from current prompts;
2. Joint evaluation and selection: Retain the top k candidates with the highest scores (beam width);
3. Iterative evolution: Repeat mutation, evaluation, and selection to gradually improve quality;
4. Cross-agent collaboration: Fix the best versions of other agents when optimizing a single agent to ensure fairness.

Experimental Validation Results

The research team verified the effectiveness of MASPO on 6 diverse tasks:

Task Types

Covers collaborative reasoning, role-play dialogue, code generation and review, creative writing collaboration, information retrieval and synthesis, and decision support systems.

Main Results

• Average accuracy improvement of 2.9 percentage points (outperforming state-of-the-art methods);
• Outperforms baselines in all tasks with no performance degradation;
• Fast convergence speed of evolutionary beam search.

Baseline Comparison

• Single-agent methods (e.g., OPRO, PromptBreeder): Ignore inter-agent impacts and perform poorly;
• Manual tuning: Cannot achieve the effect of automatic optimization;
• Naive joint optimization: Easily falls into local optima and performs worse than MASPO.

Key Findings and Insights

1. Effective downstream success metric: Focusing on the prompt's help to subsequent agents is more aligned with the essential needs of MAS;
2. Advantage of evolutionary search: Naturally suitable for discrete text spaces and less likely to fall into local optima;
3. Prompt dependency: Adjusting an agent's prompt has chain reactions, highlighting the necessity of joint optimization.

Limitations and Future Directions

Limitations

• High computational overhead: Evolutionary search requires multiple executions of MAS;

Future Directions

1. Efficient evaluation strategies: Use proxy models to predict prompt quality and reduce actual executions;
2. Dynamic environment adaptation: Explore online/continuous optimization versions;
3. Interpretability enhancement: Improve the ability to explain optimization results;
4. Cross-task transfer: Study cross-task reuse of optimization strategies.

Practical Application Value and Academic Recognition

Application Value

• Lower development threshold: Reduce reliance on prompt engineering experts;
• Improve system performance: Discover prompt combinations that are hard for humans to think of;
• Accelerate iteration: Shorten tuning cycles and support rapid prototyping and A/B testing;
• Standardized evaluation: Provide a joint evaluation framework for fair comparison of solutions.

Academic Recognition

MASPO has been accepted by ICML 2026, and the paper code is open-sourced to facilitate community reproduction and extension.