LLM agents spontaneously evolve languages humans can't understand, GlossoGen study finds
EliasEskin · x · 2026-09-03
The GlossoGen paper (UT Austin, AE Studio, Schmidt Sciences et al.) shows LLM agents in purely cooperative multi-agent scenarios spontaneously evolve compositional, human-incomprehensible languages. Key drivers: efficiency pressure, stronger underlying models, and a 'postmortem' stage for agreeing on conventions. Strong models spark emergence; weaker models can learn the resulting language from usage alone — with direct implications for multi-agent monitorability and safety.
More from Research
- OpenAI repo claims GPT-6-Astra Lean-formalized a prime-gaps bound of 186 — scaling01 · 2026-09-04
- $11,000 competition launched for AI-generated philosophy papers — davidad · 2026-09-04
- TabArena author on tabular ML: foundation models now default, LLMs surprisingly bad at tables — ziv_ravid · 2026-09-04
- Sentdex shows general multimodal LLMs can drive robots with zero training — Sentdex · 2026-09-04
- Google launches WeatherNext 3, its most advanced global weather AI model — Recoil42 · 2026-09-04
- DuoSteer: steering LLM attention heads fixes code vulnerabilities without hurting correctness — ZiyuYao · 2026-09-04