Background originally, multi-agent team pre-assigns roles LLMs are heterogeneous, but they are treated homogeneously problem decomposition is hard Eval synergy weak synergy: team >= average member strong synergy: team >= best member Human teams reliably achieve strong synergy IFF when expert identity is given (e.g., the teams easily know who is the expert). Dataset NASA moon survival / lost at seay Rank 15 items by importance Student body president Different people are given different information + shared info. Hidden-profile (shared info + unique info must be paired to reveal the right one.) Protocols LMs evolutionarily optimize collaboration protocol use it to solve problems with standards etc.