GlossoGen: New Platform Systematically Shows When LLM Agents Evolve Unintelligible Languages
EliasEskin · x · 2026-09-03
Moving beyond anecdotal single-run claims, researchers built GlossoGen, a platform for controlled sandboxed multi-agent experiments, to study language emergence rigorously across open and closed-weight models.
Key findings:
- Sufficiently strong models develop new languages under efficiency pressure when given access to a postmortem scratchpad
- The emergent languages are compositional and morphologically productive
- New agents can learn these languages purely by observing their use, without seeing construction
- Even models too weak to create languages can learn to use them, actively repairing failed conversations via targeted queries
The paper covers implications for safety and monitorability, cumulative cultural evolution, and linguistics.
More from AGI Musings
- 'Minds aren't physical' claim gets a rebuttal: Minecraft is a program on my computer — NathanpmYoung · 2026-09-04
- $11,000 competition launched for AI-generated philosophy papers — davidad · 2026-09-04
- Chinese Researcher's "End of Software Engineering" Paper: Agents Score 80% Isolated, 38% Long-Term — alex_verem · 2026-09-04
- Speculation: a 'pause AI' would push frontier labs to monetize inference instead — DimitrisPapail · 2026-09-04
- Ilya cites famous brain-lesion case: without emotion, choosing socks takes hours — JMateosGarcia · 2026-09-04
- Chollet: Keep humans in the loop across critical processes even if AI could go autonomous — fchollet · 2026-09-04