Meta-Persona Anchoring and Temperature Scaling to Counter LLM Hivemind Effect

A new paper on arXiv (2608.02618) proposes a framework to mitigate the 'Artificial Hivemind' effect in Large Language Models (LLMs), where models converge on homogenized consensus even for open questions, leading to high inter-response similarity (≈0.80–0.90) under high-temperature sampling. The method, called Meta-Persona Anchoring combined with Filtered Temperature Scaling (FTS), uses a two-stage generation process: first, the model self-selects a unique persona to anchor its starting point; second, a dual-stage sampling sieve applies Top-p filtering to preserve grammatical validity, followed by extreme temperature scaling (T ≥ 4.0) to explore a broadened probability distribution. The framework is evaluated on the INFINITY-CHAT dataset. The paper addresses the semantic collapse that limits AI diversity, proposing a practical solution to increase response variability. The research is relevant to AI development, particularly in creative and open-ended tasks where diversity is valued.