GlossoGen Paper Shows LLM Agents Evolve Compositional Languages Incomprehensible to Humans
EliasEskin · x · 2026-09-15
GlossoGen, an arXiv paper backed by Schmidt Sciences' AI Agents pilot program, introduces a platform for studying language evolution among LLM agents via the SaveVeyru scenario, where partially informed agents must communicate under pressure.
Key findings:
- Language evolution does occur between LLM agents; the resulting languages are compositional and morphologically productive, and drift from the models' English prior in ways incomprehensible to humans.
- Emergence requires efficiency pressure, strong backing models, and a "postmortem" stage for agents to agree on conventions.
- Transmission differs from emergence: stronger models are needed for novel language, but weaker models can learn an existing one from usage alone.
The results raise monitorability and safety concerns for multi-agent systems.
Related event: GlossoGen: LLM agents observed evolving their own human-illegible languages(6 posts)→
More from Research
- StudentBench: AI tutoring matches human GRE gains at 918x lower cost — gharik · 2026-10-31
- WISD: Distill Training Signal from Weaker Quantized Copies of Reasoning Models — jet-ai · 2026-10-31
- Albert Gu: H-Net learns to segment Chinese characters from raw bytes mid-training — niloofar_mire · 2026-10-31
- LLM Acknowledgments in Bio Preprints Hit 17% in Three Years, Near Specialized AI Tools — teortaxesTex · 2026-10-31
- CS theorist explains the two meanings of 'exponential time' behind new 3-SAT result — aran_nayebi · 2026-10-31
- StudentBench author: AI tutors boosting human intelligence is the use-case to hill-climb — gharik · 2026-10-31