Helpful assistant or fruitful facilitator? Investigating how personas affect language model behavior

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Human-Computer Interaction

Background:

• Large language models (LLMs) are increasingly used in diverse applications.
• Controlling LLM output and behavior is crucial for reliable and safe deployment.
• Assigning personas is a proposed method to steer LLM generation.

Purpose of the Study:

• To investigate the impact of assigned personas on the behavior of seven different large language models.
• To analyze how personas influence LLM responses across objective and subjective tasks.
• To compare persona-driven generations against control and no-persona baseline settings.

Main Methods:

• Assigned 162 personas from 12 categories (e.g., gender, occupation) to seven LLMs.
• Prompted LLMs with questions from five datasets covering objective and subjective tasks.
• Compared persona generations to a control group (variations of 'helpful assistant') and an empty persona setting.

Main Results:

• Personas demonstrably increased the variability of LLM generations across all tested models and datasets.
• The effect of personas on model behavior was more pronounced than the control setting.
• Certain aspects of persona-driven behavior showed consistency and generalization across different LLMs.

Conclusions:

• Persona assignment is an effective technique for modulating LLM behavior and output diversity.
• The findings suggest that personas can reliably influence LLM responses, offering a method for customization.
• Further research can explore the nuances of persona design for specific LLM applications.