Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling

Short summary

This paper tackles the 'Artificial Hivemind' effect where LLMs converge on homogeneous responses even for open questions. The authors propose Meta-Persona Anchoring with Filtered Temperature Scaling: models first self-select a unique persona, then a dual-stage sampling sieve applies Top-p filtering followed by extreme temperature scaling (T≥4.0). This reduces average pairwise cosine similarity from ~0.85 to ~0.65, bringing LLM response diversity closer to human-level variation. The framework is released as open-source.

• •Meta-Persona Anchoring plus extreme temperature scaling reduces LLM response homogeneity
• •Average pairwise cosine similarity drops from ~0.85 to ~0.65
• •Open-source framework released for more diverse AI deployments

Generated with AI, which can make mistakes.