Meta-persona prompting aims to reduce sameness in LLM answers

A new arXiv paper tackles what the authors call an Artificial Hivemind effect: language models converging on similar answers even when questions are open-ended and sampling temperature is high. The proposed fix combines meta-persona anchoring with filtered temperature scaling.

The method first prompts a model to choose an idiosyncratic persona as an anchor. It then applies a two-stage sampling process that preserves grammatical validity with top-p filtering before using very high temperature on the remaining candidates.

In tests on INFINITY-CHAT with open-weight models under about 20 billion parameters, the authors report average pairwise cosine similarity dropping from roughly 0.85 to about 0.65. That suggests the method produced more diverse answers.

Diversity is not automatically quality, and higher variation can introduce new errors. Still, the work addresses a real product issue: assistants that sound different on the surface may still collapse toward the same underlying response.