A new research paper published on ArXiv (cs.AI) identifies an "Artificial Hivemind" effect in Large Language Models (LLMs), where models converge on a narrow, homogenized consensus even when addressing open questions. This semantic collapse limits AI diversity, resulting in high inter-response similarity of approximately 0.80-0.90, even under high-temperature sampling.

The Proposed Mitigation Framework

To break this homogeneity, researchers propose a novel two-stage generation process combining Meta-Persona Anchoring with Filtered Temperature Scaling (FTS).

  • Meta-Persona Anchoring: The model is prompted to self-select a unique, idiosyncratic persona to anchor its starting point.
  • Filtered Temperature Scaling (FTS): A dual-stage sampling sieve is applied. First, Top-p filtering preserves grammatical validity. Second, extreme temperature scaling (T ≥ 4.0) is applied to the surviving candidates to explore a broadened probability distribution.

Experimental Results

The method was evaluated using the INFINITY-CHAT dataset on state-of-the-art open-weight models under 20 billion parameters. Results demonstrated a significant reduction in semantic convergence, with average pairwise cosine similarity dropping from approximately 0.85 to 0.65. The researchers have made their implementation available as an open-source framework to enable more diverse and creative AI deployments.