MASPO: Optimizing Prompts for LLM-Based Multi-Agent Systems
Researchers have introduced MASPO, a framework for automatically optimizing prompts in LLM-based multi-agent systems. The key innovation is a joint evaluation mechanism that assesses prompts based on their contribution to downstream agent success, not just local validity. This bridges local interactions and global goals without ground-truth labels. MASPO uses a data-driven evolutionary beam search for efficient prompt refinement.
Key facts
- MASPO is a framework for joint prompt optimization in LLM-based multi-agent systems.
- It addresses misalignment between local agent objectives and holistic system goals.
- Core innovation: joint evaluation mechanism assessing prompts by downstream success.
- Does not rely on ground-truth labels.
- Uses data-driven evolutionary beam search.
- Aims to improve complex collaborative tasks.
- Prompts are role-specific for each agent.
- Framework is iterative and automatic.
Entities
Institutions
- arXiv