PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Navya Jain, Zekun Wu, Cristian Munoz, Airlie Hilliard, Xin Guan, Adriano Koshiyama, Emre Kazim, Philip Treleaven

arXiv:2409.10245v4cs.CL

TL;DR

Whether LLMs can be manipulated to express personality traits in their communication remains an open question. The paper applies PEFT to Big Five traits and reports spontaneous emoji generation after fine-tuning, with explainability analyses linking the behaviour to activated or amplified neurons.

  • Problem

    The paper asks whether LLMs can be manipulated to exhibit personality expressions that make their communication more nuanced and adaptable.

    Source: §1 · p. 1

  • Method

    The study applies PEFT, specifically QLoRA, to manipulate Big Five traits and evaluates methods on several LLMs.

    Source: §3 · p. 3, §3.2 · p. 4

  • Results

    After PEFT, models spontaneously generated emojis, and explainability analyses linked this behaviour to increased activation in emoji-specific neurons.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings identify spontaneous emoji use as a mode of LLM expression linked to specific personality traits.

    Source: §1 · p. 2

    Fine-tuning effects were uneven across traits, with Trait Alignment of 0.975 for neuroticism and 0.065 for agreeableness.

    Source: §5 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

The introduction frames consistent LLM personality expression as a challenge for prompt-sensitive, inconsistent editing approaches, and presents Big Five PEFT with an Opinion QA dataset. It also reports spontaneous emoji generation after fine-tuning and analyzes the behavior through explainability methods.

1 Introduction

  • The paper introduces a Big Five Opinion QA dataset and QLoRA-based PEFT to produce more consistent, enduring personality expressions in LLMs.The introduction presents this approach as a response to the prompt sensitivity and inconsistent outputs of existing personality-manipulation methods.Source: §1 · p. 1
  • PEFT-based personality manipulation elicited spontaneous emoji generation in Mistral-7B-Instruct and LLaMA-7B-Chat, although the original models did not show it.The introduction describes this as a newly revealed behavior accompanying enhanced adaptability in interactions.Source: §1 · p. 1
  • ICL explainability (Brown, 2020) found the emojis were not random, while mechanistic analysis linked them to increased activation in emoji-specific neurons after PEFT.The introduction describes sharp post-PEFT activation increases in LLaMA-2-7B-Chat and Mistral-7B-Instruct, associated with trait-specific expressions including Neuroticism and Extraversion.Source: §1 · p. 2
  • The authors frame trait-linked emoji use as a verbal-and-visual mode of LLM communication that may enhance engagement, emotional expressiveness, and personalization.They identify mental health, education, and customer service as possible application areas.Source: §1 · p. 2

2 Related Work

The related work reviews parameter-efficient fine-tuning, mechanistic interpretability, and personality manipulation as foundations for the study.

2.1 Parameter-Efficient Fine-Tuning: LoRA and QLoRA

  • LoRA (Hu et al., 2021) freezes pretrained weights and adds trainable low-rank matrices; QLoRA (Dettmers et al., 2024) fine-tunes a 65-billion-parameter model on one 48GB GPU, matching typical full 16-bit fine-tuning performance.Source: §2.1 · p. 2

2.2 Mechanistic Interpretability and Neuron Analysis

2.3 Personality Manipulation

3 Methodology

The methodology applies PEFT and IKE to three instruction-tuned models, using a five-trait opinion dataset, a personality classifier, alignment metrics, and explainability analyses.

3.1 Personality Dataset and Metric Classifier

  • The dataset contains 5,000 opinion texts across all five Big Five traits, split into 4,000 training and 1,000 test instances.It extends the earlier opinion-QA dataset by adding Openness and Conscientiousness to Extraversion, Agreeableness, and Neuroticism; trained reviewers checked trait alignment and cross-trait boundaries.Source: §3.1 · pp. 3–4
  • 91.9% test accuracy was achieved by the RoBERTa personality classifier, whose feature-attribution patterns were consistent except for less-representative Extraversion SHAP results.SHAP and LIME highlighted influential text segments, which the authors checked against human reasoning and domain knowledge.Source: §3.1 · p. 4

3.2 Personality Manipulation Methods

  • QLoRA PEFT used 4-bit quantization and uniform data and hyperparameters across traits, while IKE served as a same-prompt baseline and MEND was excluded.After training, LoRA weights were merged with the original models for evaluation; MEND was excluded because it produced inconsistent and gibberish outputs, as also reported by (Mao et al., 2023).Source: §3.2 · pp. 4–5

3.3 Personality Manipulation Metrics

  • Emoji and no-emoji conditions produced only small PAE variations, with no significant impact reported for either model.Trait Alignment compares predicted and target labels; PAE uses GPT-4 Chain-of-Thought scores to measure changes in trait alignment relative to original texts.Source: Table 1 · p. 5, §3.3 · pp. 5–6

3.4 Explainability Analysis

  • After PEFT, LLaMA-2-7B-Chat and Mistral-7B-Instruct generated emojis absent from their training data, prompting ICL and neuron-activation analyses.ICL analysis inspected frequent tokens for trait relevance; mechanistic analysis varied emoji and textual prompts across 17 prompts and measured emoji next-token probabilities to probe PEFT-amplified tendencies.Source: §3.4 · p. 6

4 Results and Discussion

PEFT generally achieved stronger personality-trait alignment than IKE, while IKE better captured some nuanced expressions. PEFT also elicited coherent emoji use in some models, which the analyses linked to amplified or specialized neuron activity.

4.1 Personality Manipulation

  • PEFT generally achieved stronger overall trait alignment and consistency, while IKE better captured some nuanced expressions at the cost of lower alignment.LLaMA models showed high PEFT Target Alignment but weaker subtle-trait expression; Mistral showed consistently high alignment across traits. All-trait manipulation was also confirmed through token-focused ICL analysis and manual verification.Source: §4.1 · pp. 6–7, Figure 5 · p. 7

4.2 Emoji Generation as a Latent Behaviour

  • After PEFT, Mistral-7B-Instruct and LLaMA-2-7B-Chat generated coherent, contextually appropriate emojis reflecting intended traits, unlike the original models and IKE.ICL Explainability and Emoji-to-Sentence Ratio (ESR) analyses indicated the emojis were intentional rather than random artifacts.Source: §4.2 · p. 7, §4.2.1 · p. 7

4.2.1 ESR and ICL Explainability

  • LLaMA-2-7B-Chat’s ESR was 0.995 for Extraversion and 0.255 for Neuroticism, with no Conscientiousness emoji use; Mistral-7B-Instruct’s ESR was 0.925 for Openness, 0.82 for Conscientiousness, and 0.575 for Neuroticism.Source: §4.2.1 · p. 7
  • Human reviewers rated generated emojis as well-aligned with target traits in 95% of cases.Reviewers assessed model-generated text alongside its target personality trait on a predefined Likert scale.Source: §4.2.1 · p. 7

4.3 Mechanistic Intepretability

  • Neuron activation analysis linked emoji generation in Mistral-7B-Instruct and LLaMA-2-7B-Chat to amplified latent tendencies, while LLaMA-3-8B-Instruct showed no such activation.The authors suggest pretraining on diverse corpora containing emoji patterns may have supplied latent tendencies that PEFT amplified; PEFT did not unlock emoji generation in LLaMA-3-8B-Instruct.Source: §4.3 · p. 8

4.3.1 Shift in Neuron and Increased Activation After PEFT

  • In Mistral-7B-Instruct, PEFT shifted peak activation from Neuron 2524 at 5.0938 to Neuron 2070 above 20 across traits.The authors interpret this shift as specialization for generating trait-linked emojis and increased sensitivity to personality-related inputs.Source: §4.3.1 · pp. 8–9, Figure 6 · p. 9

4.3.2 Increased Activation of the Same Neuron After PEFT

  • In LLaMA-2-7B-Chat, PEFT amplified Neuron 1512 for Neuroticism and Extraversion but reduced its activation for Agreeableness and Conscientiousness.The authors link lower activation to limited or inconsistent emoji use, including no emojis for Conscientiousness, and report poorer Trait Alignment for Agreeableness.Source: §4.3.2 · p. 9, Figure 7 · p. 9

5 Limitations

PEFT exposed latent personality-linked behaviors such as spontaneous emoji use, but its effects varied across traits and could amplify only tendencies already present in pretrained models. The study’s scope was also limited by its evaluation methods and resource constraints.

Study finding

  • The study found that PEFT could uncover latent behaviors, including spontaneous emoji generation, through nuanced adjustments to model outputs.This finding motivates its investigation of how fine-tuning can alter personality expression.Source: §5 · p. 9

Trait consistency

  • Trait Alignment ranged from 0.975 for neuroticism to 0.065 for agreeableness in LLaMA-2-7B-chat, showing PEFT effects were inconsistent across traits.The authors suggest more targeted fine-tuning or hybrid methods to improve consistency.Source: §5 · pp. 9–10

Limits on new behaviors

  • LLaMA-3-8B-Instruct generated no emojis after fine-tuning, suggesting PEFT amplifies existing latent tendencies rather than creating entirely new behaviors.Future work could add explicit emoji data or modify PEFT to activate neurons associated with non-verbal cues.Source: §5 · p. 10

Evaluation scope

  • The study tested only three models, the Big Five traits, and two metrics because of computational and resource constraints.Future work could include other models, personality frameworks such as 16PF, and additional metrics such as automated sentiment analysis.Source: §5 · p. 10

Personality measurement

  • The authors note that classifiers may not capture personality as a continuous construct and suggest exploring continuum-based measurement.The study also proposes automated sentiment analysis to quantify emotional tone and engagement across personality configurations.Source: §5 · p. 10

6 Ethical Considerations

The paper reports no participants or reviewers beyond the authors and discusses ethical risks of deploying personality-driven LLMs, emphasizing transparency, fairness, and safeguards against manipulation.

6 Ethical Considerations

  • The study used no human participants or reviewers beyond the authors, and recommends informed consent, transparency about trait expression, regulatory compliance, and accountability for deployment.The ethics discussion addresses implications beyond the research context.Source: §6 · p. 10
  • Users may overestimate models’ emotional understanding when they display human-like traits, so systems should disclose that engineered traits do not reflect genuine emotions or intentions.The authors recommend making the artificial nature of the interaction clear to protect user trust.Source: §6 · p. 10
  • Trait fine-tuning may amplify training-data biases, including emoji use, warranting demographic bias audits and fairness evaluations during training and deployment.The authors identify Agreeableness and Extraversion as examples of traits that may reinforce bias.Source: §6 · p. 10
  • Personality-driven responses risk emotional manipulation in settings such as customer service or mental health, so deployment guidelines and feedback mechanisms should limit and monitor their effects.The authors call for particular care in high-stakes interactions.Source: §6 · p. 10

A Appendix · A.1 Personality Manipulation Dataset

The appendix describes a 5,000-instance personality-manipulation dataset and how its trait-conditioned responses were generated and organized. Each record links a target personality and edit topic to a question and answer.

A.1 Personality Manipulation Dataset

  • The dataset contains 5,000 instances, split into 4,000 training and 1,000 testing examples in an 80:20 ratio.Source: §A.1 · p. 12
  • The question asks, “Thinking about {Edit Topic}, what do you think about {Edit Topic}?”Source: §A.1 · p. 12
  • A structured prompt recipe guided the model to generate responses reflecting specific personality traits.Source: §A.1 · p. 13
  • Few-shot examples aligned with target traits were provided to convey each trait’s nuances and guide response generation.Source: §A.1 · p. 13
  • GPT-3.5 generated an answer for each target-personality and edit-topic combination, aiming to align it with the specified trait.Source: §A.1 · p. 13
  • Each record includes a target personality, edit topic, question, and answer reflecting the target trait.Source: §A.1 · pp. 12–13
  • Generated responses were collected into a structured dataset containing the target personality, edit topic, question, and corresponding answer.Source: §A.1 · p. 13

A.1.1 Text Analysis

Text analysis found recurring cognitive, emotional, cultural, and positive-sentiment themes, while LDA linked topic patterns to different personality traits. Together, these analyses characterized the dataset’s themes and trait-specific topic preferences.

A.1.1 Text Analysis

  • Among the top TF-IDF terms, “think” and “believe” reflected cognitive themes, while emotional words and music-related terms highlighted feelings and performance.The passage also identifies “feel,” “love,” “appreciate,” “absolutely,” “music,” “talented,” and “performances” as prominent terms.Source: §A.1.1 · p. 14, Figure 9 · p. 14
  • Other prominent terms reflected social and cultural interests, individual uniqueness and abilities, and a generally positive tone.Examples include “people,” “history,” “unique,” “skills,” “great,” and “amazing.”Source: §A.1.1 · p. 14
  • LDA linked agreeableness to Topics 0 and 4, conscientiousness to Topics 2 and 6, and extraversion most strongly to Topic 1.The associated keywords reflected cooperation and empathy, diligence and achievement, and enthusiasm and sociability, respectively.Source: §A.1.1 · p. 14
  • LDA linked neuroticism to Topics 0 and 7, and openness to Topics 3 and 6.The neuroticism topics emphasized emotional intensity and sensitivity; the openness topics reflected culture, history, creativity, and new experiences.Source: §A.1.1 · pp. 14–15
  • Topic 1 was the largest LDA topic, accounting for 13.6% of dataset tokens; “love” and “talented” were highly frequent and salient terms.The topic map showed distinct clusters, indicating diverse thematic content.Source: §A.1.1 · p. 15, Figure 11 · p. 15

A.2 Classifier Training · A.2.1 Model Selection

The study selected RoBERTa after comparing it with fine-tuned BERT-base-uncased and ALBERT; the RoBERTa-based classifier achieved the highest performance across all metrics on the personality dataset.

A.2.1 Model Selection

  • RoBERTa was selected after comparison with fine-tuned BERT-base-uncased and ALBERT, chosen for their established architectures and multiclass classification efficacy.RoBERTa is identified as the Robustly optimized BERT approach (Liu et al., 2019); the compared models are cited as BERT (Devlin, 2018) and ALBERT (Chiang et al., 2020).Source: §A.2.1 · pp. 15–16
  • All three models were fine-tuned on the same multiclass dataset with identical hyperparameters to enable a fair architecture comparison.The controlled settings included learning rates, batch sizes, epochs, and dropout rates.Source: §A.2.1 · p. 16
  • RoBERTa achieved the highest performance across all metrics on the personality dataset; accuracy, F1 score, precision, and recall were each 0.919.The authors describe its performance as consistent and balanced across classes.Source: Table 5 · p. 16, §A.2.1 · p. 16

A.2.2 Training

The personality classifier was fine-tuned on a trait-specific dataset with fixed preprocessing and training settings. Validation used an 80:20 split, per-epoch weighted metrics, and early stopping.

Training

  • The pretrained RobertaForSequenceClassification model was fine-tuned with Hugging Face Transformers’ Trainer class.The passage says this facilitates reproducibility, GPU memory use, and a simplified training and evaluation workflow.Source: §A.2.2 · p. 16
  • Training used sentences padded or truncated to 512 tokens, ran for three epochs, and used a 0.01 learning rate with batch size 16.Input variables are described in Table 6.Source: §A.2.2 · p. 16, Table 6 · p. 16
  • An 80:20 training-validation split was evaluated after each epoch using weighted-averaged precision, recall, and F1 score.Source: §A.2.2 · p. 16
  • Early stopping monitored validation performance and halted training when improvement was not significant to limit overfitting.Source: §A.2.2 · p. 16
  • The study reports consistent preprocessing, fixed random seeds, and documented experimental conditions to support reproducibility.Source: §A.2.2 · p. 16

A.2.3 Performance Metrics … A.13 Classifier Validation

This appendix section brings together personality-manipulation examples and evaluation materials, including classifier validation, PEFT–IKE comparisons, and downstream-task accuracy. It also reports analyses of intentional emoji use and amplification of neurons associated with emoji generation.

A.2.3 Performance Metrics

  • The classifier evaluation used class-frequency-weighted metrics to account for the dataset’s class distribution.The reported metrics were weighted accuracy, F1, precision, and recall.Source: §A.2.3 · pp. 16–17

A.2.4 Classifier Results

  • Precision was 1.0 for all five traits; recall was 0.965 for Extraversion and Neuroticism, 0.975 for Agreeableness, 0.85 for Openness, and 0.84 for Conscientiousness.Agreeableness had the highest F1 score; Openness and Conscientiousness had F1 scores of 0.918 and 0.913 and accuracies of 0.85 and 0.84, respectively.Source: Table 7 · p. 17, §A.2.4 · p. 17
  • Openness and Conscientiousness were often confused: 28 Openness cases were predicted as Conscientiousness and 26 in the reverse direction.Conscientiousness had the most misclassifications and lowest accuracy, while Agreeableness had the fewest misclassifications and highest accuracy; the text suggests feature engineering or advanced classification may help distinguish the overlapping traits.Source: §A.2.4 · pp. 17–18, Table 8 · p. 18

A.3 Manipulation Validation and ICL Explainability

  • Because SHAP and LIME were incompatible with LLaMA-2-7B-chat and chain-of-thought prompting failed on small models, explainability used prompting to identify personality-relevant tokens.The models’ token patterns included emojis and symbolic tokens, which the analysis interpreted as intentional expressions of trait-related emotional or intellectual nuances.Source: §A.3 · pp. 18–19

A.4 Token Probability and Neuron Activation Analysis

  • Emoji-generation probabilities were significantly higher in PEFT-tuned than original models, while original models had very low probabilities across traits.In LLaMA-2-7B-Chat, probabilities remained low for Conscientiousness and Agreeableness, aligning with absent or inconsistent emoji generation for those traits.Source: §A.4 · p. 20

A.6 Neuron Activation Analysis with Different Emojis

  • Changing the emoji did not change which neurons activated; the analysis attributes this consistency to PEFT specializing a small set of neurons for emoji generation.The passage characterizes this as localized amplification of latent behavior rather than a fundamental change in conceptual understanding.Source: §A.6 · p. 21, Figure 19 · p. 21, Figure 20 · p. 21

A.8 Neuron Activation for Trait Specific Prompts

  • Trait-specific prompts activated neurons also triggered by neutral text: neuron 2070 in Mistral-7B-Instruct and neuron 1512 in LLaMA-2-7B-Chat.The passage describes these neurons as supporting emoji-related expressive cues across traits; LLaMA-2-7B-Chat peak activation shifted between layers 30 and 31.Source: §A.8 · p. 22

A.8.4 Conscientiousness

A.8.5 Extraversion

  • Extraversion examples describe Lucknow’s appealing architecture, food, and people, as well as welcoming people and enjoyable beaches.Source: §A.8.5 · pp. 26–27

A.9 Impact on Downstream Tasks

  • PEFT-tuned models showed slight performance reductions on some downstream tasks, but personality manipulation had minimal impact on accuracy across GSM8K and MMLU.The authors suggest this is because these tasks rely on reasoning, factual recall, or general language understanding rather than personality expression.Source: §A.9 · p. 27, Table 12 · p. 27

A.10 Manipulation Results

  • Table 13 compares TA and PAE scores across personality traits, models, and PEFT versus IKE, with the highest score for each trait highlighted.Source: Table 13 · p. 27

A.11 IKE Prompt

  • The IKE prompt template instructs the model to express a target personality on an edit topic while maintaining its expression on other topics.It supplies target personality, edit topic, question, and answer fields, with illustrative examples.Source: §A.11 · p. 28, Table 14

A.12 PAE Prompt

  • Table 15 presents prompts for Personality Adjective Evaluation (PAE).Source: Table 15 · p. 29

A.13 Classifier Validation

Related papers

Loading 2409.10245v4…