From Text to Emoji: How PEFT-Driven Personality Manipulation Unleashes the Emoji Potential in LLMs

TL;DR

The paper tackles the challenge of reliably manipulating LLM personality by leveraging parameter-efficient fine-tuning (PEFT) with Quantized Low-Rank Adaptation (QLoRA) to adjust Big Five traits. It introduces an Opinion QA dataset and trait-alignment metrics (TA and PAE) to benchmark PEFT against prompt-based edits (IKE), demonstrating that PEFT yields more consistent trait expression across several models, with additional emergent behaviour—latent emoji generation—uncovered and analyzed through In-Context Learning explainability and Mechanistic Interpretability. The authors show that emojis are not random artifacts but aligned with target traits and traceable to specific neurons whose activations are amplified by PEFT, indicating a neurally grounded mechanism for nonverbal trait signaling. However, they also report trait-dependent inconsistencies and limitations in creating entirely new behaviours, underscoring the need for broader testing across architectures and ethical deployment guidelines. Overall, this work reveals a latent nonverbal expressive channel in LLMs and provides a methodological framework for PEFT-driven personality manipulation with explainability insights that can inform both personalization and responsible AI practice.

Abstract

The manipulation of the personality traits of large language models (LLMs) has emerged as a key area of research. Methods like prompt-based In-Context Knowledge Editing (IKE) and gradient-based Model Editor Networks (MEND) have been explored but show irregularity and variability; IKE depends on the prompt, leading to variability and sensitivity, while MEND yields inconsistent and gibberish outputs. To address this, we employed Opinion QA Based Parameter-Efficient Fine-Tuning (PEFT), specifically Quantized Low-Rank Adaptation (QLoRA), to manipulate the Big Five personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. After PEFT, models such as Mistral-7B-Instruct and LLaMA-2-7B-chat showed a latent behaviour by generating emojis for certain traits, despite no emojis being present in the PEFT data. For instance, LLaMA-2-7B-chat generated emojis in 99.5\% of extraversion-related test instances, while Mistral-7B-Instruct did so in 92.5\% of openness-related test instances. ICL Explainability analysis indicated that the LLMs used emojis intentionally to express these traits. Mechanistic Interpretability analysis showed that this latent behaviour of LLMs could be traced to specific neurons that became activated or amplified after PEFT. This paper provides a number of novel contributions. First, introducing an Opinion QA dataset for PEFT-driven personality manipulation; second, developing metric models to benchmark LLM personality traits; third, demonstrating PEFT's superiority over IKE in personality manipulation; and finally, analysing and validating emoji usage through explainability methods such as Mechanistic Interpretability and In-context learning Explainability methods.

Figures (98)

• Figure 1: Case of the fine-tuned personality with Mistral-7B-Instruct
• Figure 2: Word Clouds Representing the Big Five Personality Traits
• Figure 3: SHAP visualisation for Agreeableness
• Figure 4: LIME visualisation for Agreeableness
• Figure 5: Comparison of TA and PAE scores across different traits, models and methods.