MASPO: Joint Prompt Optimization Across Multi-Agent Systems Accepted at ICML 2026

arXiv·medium signal

MASPO bridges local agent objectives and global system goals by evaluating prompts based on their capacity to facilitate downstream agent success, not just local validity. Using data-driven evolutionary beam search without ground-truth labels, it achieves +2.9 average accuracy improvement over SOTA prompt optimization across 6 tasks. Code released; accepted at ICML 2026.