Researchers propose Meta-Persona Anchoring and Filtered Temperature Scaling to reduce LLM homogenization

A new arXiv preprint introduces a framework called Meta-Persona Anchoring combined with Filtered Temperature Scaling (FTS) to address the 'Artificial Hivemind' effect in large language models (LLMs). The authors argue that LLMs often converge on narrow, homogenized responses even for open-ended questions, a phenomenon they quantify as high inter-response similarity (≈0.80–0.90) even under high-temperature sampling.

The proposed method operates in two stages: first, the model is prompted to self-select a unique persona to anchor its starting point, and second, a dual-stage sampling process applies Top-p filtering to preserve grammatical validity before applying extreme temperature scaling (T ≥ 4.0) on the surviving candidates. This aims to broaden the explored probability distribution and increase response diversity.

The researchers evaluate their approach using the INFINITY-CHAT dataset on state-of-the-art open-weight models with approximately 20 billion parameters. They report a significant reduction in semantic convergence, with average pairwise cosine similarity dropping from approximately 0.85 to approximately 0.65. The method achieves a majority of questions below the 0.7 threshold, which the authors describe as reducing the gap between artificial mode collapse and human-level typological diversity.

To support reproducibility and further research, the authors provide an open-source implementation of the framework. They suggest this could enable more diverse and creative deployments of AI systems by mitigating the tendency toward homogenized outputs.