Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling
Quick Answer
This paper introduces a framework combining Meta-Persona Anchoring and Filtered Temperature Scaling to combat the 'Artificial Hivemind' effect in LLMs, reducing semantic convergence from 0.85 to 0.65.
Quick Take
The approach leverages a two-stage generation process to enhance diversity in AI responses, evaluated on the INFINITY-CHAT dataset with models around 20B parameters.
Key Points
- Proposed method reduces average pairwise cosine similarity from 0.85 to 0.65.
- Utilizes a two-stage generation process for enhanced response diversity.
- Evaluated on state-of-the-art models with approximately 20 billion parameters.
- Achieves a majority of responses below the 0.7 similarity threshold.
- Open-source implementation available for diverse AI applications.
DeepSignal Analysis
What happened
The paper presents a framework aimed at addressing the 'Artificial Hivemind' effect in Large Language Models (LLMs). By combining Meta-Persona Anchoring with Filtered Temperature Scaling, the authors report a reduction in semantic convergence from approximately 0.85 to 0.65, enhancing response diversity. This method was evaluated using the INFINITY-CHAT dataset with models around 20 billion parameters.
Key evidence
- The proposed framework combines Meta-Persona Anchoring and Filtered Temperature Scaling to mitigate the Artificial Hivemind effect in LLMs.
- The average pairwise cosine similarity of responses decreased from approximately 0.85 to 0.65, indicating improved diversity.
- The evaluation was conducted on the INFINITY-CHAT dataset using state-of-the-art models with around 20 billion parameters.
Why it matters
This research highlights a significant issue in LLMs, where models tend to produce homogenized outputs, limiting their utility in diverse applications. By demonstrating a method to enhance response variability, the authors contribute to the ongoing efforts to create more versatile AI systems. The reduction in semantic convergence suggests potential for more human-like interaction in AI, which could improve user experience across various domains.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Recent studies have identified an ``Artificial Hivemind'' effect in Large Language Models (LLMs) causing models to converge on a narrow, homogenized consensus even for open questions. This semantic collapse limits the diversity of AI, resulting in high inter-response similarity ($\approx 0.80-0.90$) even under high-temperature sampling. In this paper, we propose a novel mitigation framework to increase diversity: Meta-Persona Anchoring combined with Filtered Temperature Scaling (FTS). Our approach utilizes a two-stage generation process: first, the model is prompted to self-select a unique, idiosyncratic persona to anchor its starting point; second, we apply a dual-stage sampling sieve, utilizing Top-$p$ filtering to preserve grammatical validity followed by extreme temperature scaling ($T \ge 4.0$) on the surviving candidates to explore the broadened probability distribution. We evaluate our method using the INFINITY-CHAT dataset on state-of-the-art open weight models under $\sim$20B parameters. Our results demonstrate a significant reduction in semantic convergence, with average pairwise cosine similarity dropping from ($\approx 0.85$) to ($\approx 0.65$). Our scheme achieves a majority of questions below the 0.7 threshold, effectively reducing the gap between artificial mode collapse and human-level typological diversity. We provide our implementation as an open-source framework to enable more diverse and creative AI deployments.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2608.02618 [cs.AI] |
| (or arXiv:2608.02618v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2608.02618 arXiv-issued DOI via DataCite |
Submission history
From: Pedro Reviriego [view email]
[v1]
Sun, 31 May 2026 17:04:01 UTC (138 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →HOBA: Hierarchical On-Policy Bidding Agents for Adaptive Online Advertising
HOBA (Hierarchical On-policy Bidding Agents) is a novel hierarchical reinforcement learning framework that enhances online advertising bidding systems by improving adaptability and reducing hyperparameter tuning costs. It utilizes a for hyperparameter inference, a SARSA agent for expert model selection, and a dynamic expert pool for bid execution, achieving a +3.6% increase in target cost during large-scale deployment and outperforming state-of-the-art baselines on AuctionNet.