MASPO framework optimizes prompts across LLM agent teams for 2.9% gain
New ICML paper jointly optimizes prompts to align local agent goals with system-wide success.
Deep Dive
Researchers introduce MASPO, a framework for automatically refining prompts in LLM-based multi-agent systems. Unlike prior methods that optimize prompts locally, MASPO uses a joint evaluation mechanism that measures how a prompt helps downstream agents succeed. It employs a data-driven evolutionary beam search to explore the prompt space. Tested on 6 diverse tasks, MASPO outperforms state-of-the-art prompt optimization methods with an average accuracy improvement of 2.9. The code is open-sourced.
Key Points
- MASPO introduces a joint evaluation mechanism that measures prompt quality by downstream agent success, not local task completion.
- Uses a data-driven evolutionary beam search to efficiently explore the prompt space across interacting agents.
- Achieves an average accuracy improvement of 2.9% over state-of-the-art methods across 6 collaborative tasks.
Why It Matters
Automates prompt tuning for multi-agent AI teams, enabling more reliable and scalable autonomous collaboration without manual engineering.