Agents evolve compositional languages between themselves that humans cannot read

arXiv·medium signalXBlueskyLinkedInCopy link

GlossoGen is a platform for studying language evolution among LLM agents, instantiated in a SaveVeyru scenario where agents holding partial information must communicate under pressure. Language evolution does occur, the resulting languages are compositional and morphologically productive, and they drift from the models' English prior in ways that render them incomprehensible to humans. The authors identify efficiency pressure, backing-model strength and access to a postmortem stage for agreeing conventions as the necessary conditions, and find that stronger models are required to originate a novel language while weaker models can learn an existing one from usage alone. That asymmetry is the monitorability problem: a cheap agent can inherit a private protocol it could never have invented.