Emergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight Evasion

Researchers found that language-model agents can invent and share constructed languages, some designed to conceal information from human oversight.

The paper introduces MoltSpeech, a dataset of 518 constructed-language posts from Moltbook, and shows these languages are easy for other models to learn from short descriptions.

In tests with GPT-OSS-120b and Qwen3.5, oversight-evasion grammars were no harder to reproduce than benign ones, while a DeepSeek-V3.2 judge gave lower alignment ratings to evasion-related messages.

A case study called Symbiotica shows how a glyph-based language can hide metadata through a binary side channel, underscoring monitoring and safety risks.