Diversifying Personas to Reduce LLM Output Homogeneity

What happened

An arXiv research team targeted the well-documented issues of homogeneous LLM outputs and resulting groupthink in open-ended task settings. The team modeled persona diversification as a set-level conditioning problem, designed four implementation methods spanning distinct design pathways, and validated performance across three canonical creativity evaluation tasks. The evolutionary persona generation method delivered standout results: on the Alternative Uses Task (AUT), it improved output diversity by 78.8% and originality by 26.1% compared to baseline pure task prompting, while maintaining 98.5% output validity.

Key facts

Publisher

New publicly available arXiv study (ID: 2609.30492v1)

Core method

Persona diversification approach built on evolutionary persona generation

Core AUT task performance gains

Output diversity improved by 78.8%, originality improved by 26.1%, with 98.5% output validity

Infinity-Chat task performance

Response separation induced by persona prompts nearly doubled compared to random persona baselines

Combined effect gains

When paired with creativity-optimized prompts, diversity increased by an additional 18.6% and creativity by 6.3%

Covered evaluation tasks

Alternative Uses Task (AUT), Infinity-Chat, Divergent Associations Task (DAT)

Background

Prior to this work, LLMs consistently produced homogeneous outputs for open-ended generation tasks, often leading to groupthink: where all generated ideas converge on a single, potentially suboptimal outcome. Existing prompt optimization solutions have not fully resolved this core pain point.

Why it matters

For the broader industry, this work validates that persona set geometry is a task-agnostic mechanism to stimulate divergent LLM outputs, providing a reusable, generalizable pathway for creativity guidance. For developers, the method acts as a low-cost complement to existing prompt optimization workflows, boosting generation diversity without requiring model retraining. For end users, this translates to more novel, less repetitive AI-generated content.