Comparing Prompt and Representation Engineering for Personality Control in Language Models: A Case Study

Pengrui Han

Keywords: Personality Modeling, LLMs, Prompt Engineering, Representation Engineering, Control Vector

TL;DR: Comparing Prompt and Representation Engineering for Personality Control in Language Models

Abstract: Language models can exhibit different personalities through methods like prompt engineering and representation engineering, but how these approaches differ in modeling personality traits remains unclear. In this case study, we conduct a systematic comparison of these methods across two tasks: moral decision-making and narrative generation. In moral dilemmas, we examine how personalities (logical, empathetic, conservative, and risk-taking) influence choices between progressive and conservative options, finding that prompt engineering better aligns with intuitive personality traits while control vectors show more consistent but sometimes unexpected behaviors. In narrative generation, we analyze how different personalities (extroverted, introspective, angry, and whimsical) affect story characteristics, revealing that control vectors enable wider emotional range but lower lexical diversity compared to prompting. Our results demonstrate complementary strengths: prompt engineering excels in maintaining personality-aligned behaviors and vocabulary richness, while representation engineering offers more precise control over emotional expression and linguistic complexity. These findings provide insights into choosing and combining personality control methods for different applications.

Public: Yes

Track: Flash

Submission Number: 1