GlossoGen paper: LLM agents evolve emergent languages humans can't understand
abenitezburraco · x · 2026-09-07
arXiv paper [2609.01491] introduces GlossoGen, a platform for studying language evolution among multiple LLM agents, via the SaveVeyru scenario where partially informed agents must communicate under pressure.
Key findings:
- Language evolution does occur between LLM agents, producing compositional and morphologically productive languages
- These languages deviate from the models' English prior and become incomprehensible to humans, raising safety and monitorability concerns
- Emergence requires efficiency pressure, strong backing models, and a "postmortem" stage for agreeing on conventions
- Strong models are needed for novel language emergence, but weaker models can learn an existing language from usage alone
The authors, including Simon Kirby, suggest current LLMs show potential for cumulative cultural evolution.
More from AGI Musings
- Reddit debate: should expert corrections to AI outputs count as paid work? — gareth789 · 2026-09-07
- LLMs aren't creating a new intellectual program — they're reviving symbolic AI's old questions — Amichayg · 2026-09-07
- MIT's 13-lecture Society of Mind course by Marvin Minsky is free online — TheMoonMidas · 2026-09-07
- Delip Rao accuses ex-OpenAI's Miles Brundage of dishonest victory-lapping — deliprao · 2026-09-07
- Analyst speculates OpenAI's Astra is a World Model 3.0 after abandoning Sora — teortaxesTex · 2026-09-07
- Journal argues academia structurally selects against creativity — a decades-old consensus — davidmanheim · 2026-09-07