LLM‑evolved constitutions outperform humans

A recent result claims that large language models, when used iteratively to evolve 'constitutions' (rules and procedures) for multi‑agent systems, produce governance protocols that beat both human‑designed principles and one‑shot LLM rules in coordination tasks. If robust, this suggests LLMs can design durable, adaptive institutional rules for groups of agents. — Shows a concrete route by which AI could generate better coordination and governance rules than humans, raising questions about delegated rule‑making, oversight, and institutional design.