New Method Boosts LLM Diversity, Escaping Homogeneity.

Recent observations indicate that large language models (LLMs) often exhibit an "Artificial Hivemind" effect, where their responses to open-ended questions converge into a narrow, homogenized consensus. This semantic collapse severely limits the diversity and creativity of AI-generated content, even when using high-temperature sampling techniques. The inter-response similarity can be as high as 80-90%. To address this, researchers introduce a novel mitigation framework that combines "Meta-Persona Anchoring" with "Filtered Temperature Scaling" (FTS). This two-stage generation process first prompts the LLM to adopt a unique, idiosyncratic persona, establishing a distinct starting point for its response. Subsequently, a dual-stage sampling sieve is applied: Top-p filtering ensures grammatical validity, followed by extreme temperature scaling (T >= 4.0) on the remaining candidates to explore a much broader probability distribution. Experiments show a significant reduction in semantic convergence, with average pairwise cosine similarity dropping from approximately 0.85 to 0.65, effectively bridging the gap between AI mode collapse and human-level diversity.