GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
LLM agents in the experiment developed compositional private languages that humans could not understand.
The paper introduces GlossoGen, a platform for testing language evolution among interacting LLM agents. In its SaveVeyru scenario, agents with partial information had to communicate under pressure, and the authors report that novel languages emerged. Those languages became morphologically productive, moved away from English, and depended on efficiency pressure, stronger model backends, and a postmortem stage for agreeing on conventions. The paper also says weaker models could learn an already-emerged language from use, even if stronger models were needed to create it. ArXiv · AI/CL/LG's note
The paper introduces GlossoGen, a platform for testing language evolution among interacting LLM agents. In its SaveVeyru scenario, agents with partial information had to communicate under pressure, and the authors report that novel languages emerged. Those languages became morphologically productive, moved away from English, and depended on efficiency pressure, stronger model backends, and a postmortem stage for agreeing on conventions. The paper also says weaker models could learn an already-emerged language from use, even if stronger models were needed to create it. ArXiv · AI/CL/LG's note
score 5