Definition
Plain language
When AI models talking to each other keep using new words but stop generating genuinely new ideas.
As stated in the literature
A reproducible phenomenon in closed-loop multi-LLM systems where lexical diversity grows while semantic diversity contracts toward a model-specific attractor, robust to temperature, prompt, persona, scale, and RL interventions.
Why it matters: If multi-agent systems converge to narrow attractors regardless of prompts, claims about emergent collective intelligence from agent collectives need to be heavily discounted.
For example, two chatbots talking to each other start out exploring different topics but, over hundreds of turns, drift into recycling the same handful of ideas dressed in fresh vocabulary.
Heard on the show
“If that space systematically compresses certain dimensions of meaning, then "semantic collapse" could be partly an artifact of how the measurement projects the conversation.”Episode 073 — When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving