Multi-agent training aims to improve coordination on complex tasks

Researchers have introduced a new framework that trains multiple AI agents simultaneously, each with a specialized role, which promises to enhance the reliability of handling intricate, multi-step tasks. This is achieved through a clearer division of labor and tighter coordination. Unlike current AI systems that rely on a single agent, this structured hierarchy employs one agent as a project manager overseeing the workflow, while specialized sub-agents handle specific tools like web search or data analysis. The new Multi-Agent Group Relative Policy Optimization, or M-GRPO, extends GRPO so main and sub-agents can be trained together while keeping their roles distinct. Tested on benchmarks like GAIA, XBench-DeepSearch, and WebWalkerQA, M-GRPO outperformed both single GRPO agents and multi-agent setups with untrained sub-agents, showcasing more stable behavior and requiring less training data.

New framework uses specialized AI agents to handle multi-step tasks more reliably by dividing labor and tightening coordination.

RecursivAI

Multi-Agent AI Training Improves Complex Task Coordination