Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams

A proposed method called group-relative credit assignment addresses the problem that cooperative language-model agents revising reusable skills cannot attribute a shared team reward to the responsible agent and skill, since existing methods credit skills present in each trajectory even though presence is not responsibility. The approach targets evolving multi-agent teams that improve without weight updates by revising reusable skills.

Cooperative language-model agents can improve without weight updates by revising reusable skills, but only if a shared team reward is traced to the responsible agent and skill. Existing methods credit the skills present in each trajectory, yet presence is not responsibility: a harmful skill active i