A Two-Step Trick Aims to Break the 'AI Hivemind'

Anyone who has bounced between ChatGPT, Claude, and Gemini has probably noticed the same thing: the replies feel interchangeable, "I'm an AI assistant, here's a balanced view," in three slightly different fonts.

A new arXiv preprint calls that pattern the "Artificial Hivemind" and proposes a two-stage fix. First, the model picks an idiosyncratic persona to anchor the response. Then a sampling sieve kicks in: a Top-p filter discards anything ungrammatical, and the surviving candidates are resampled at extreme temperatures, T ≥ 4.0, normally a setting that produces gibberish, to widen the distribution of possible answers.

On the INFINITY-CHAT benchmark, using open-weight models under roughly 20 billion parameters, the authors report that average pairwise cosine similarity between responses drops from about 0.85 to about 0.65, with most questions falling below a 0.7 similarity threshold.

The catch is what the preprint is not. It is not peer-reviewed. It is a single benchmark, evaluated by the same author group that defined the "Hivemind" construct last year, and the numbers are self-reported. The approach is a sampling trick, not a training-time change, so any effect on factuality, safety refusals, or instruction-following is uncharacterized. There is no independent replication yet, and the result says nothing about larger frontier models.

The recipe is testable. The proof will be whether other labs can repeat it on the models the field actually ships.