Language models often produce homogeneous responses on open‑ended tasks, leading to groupthink—the convergence of ideas toward a single, potentially suboptimal decision. This work formulates persona diversification as a set‑level conditioning problem and explores two orthogonal design axes: selecting versus generating personas, and space‑filling versus frontier‑seeking diversity. Four concrete methods instantiate this space:
- Coverage‑based subset selection
- Dispersion‑based subset selection
- Uniform‑coverage sampling
- Evolutionary persona generation
The methods are evaluated on three creativity benchmarks: the Alternative Uses Task (AUT), Infinity‑Chat, and the Divergent Association Task (DAT). Evolutionary persona generation boosts AUT response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and holistic creativity by 13.9%, while preserving 98.5% validity. On Infinity‑Chat, persona‑induced response separation is nearly double that of random personas. Moreover, combining evolutionary personas with creativity‑optimized prompting further raises diversity by 18.6% and creativity by 6.3%.
These findings establish persona‑set geometry as a task‑agnostic mechanism for eliciting divergent LLM outputs and support persona diversification as a reusable complement to prompt optimization.
Review