Personality Traits in LLMs

• LLM personality traits are defined via psychometric assessments using adapted instruments like the Big Five Inventory and measured with metrics such as Cronbach’s α, λ₆, and ω.
• The research demonstrates prompt-based shaping, neuron-level editing, and model merging to modulate traits, achieving high alignment (Spearman’s ρ ≥ 0.90) with intended personality profiles.
• LLM personality is emergent and context-dependent, impacting applications in human-AI interaction, safety, and personalization while posing challenges in measurement and consistency.

LLMs trained on vast corpora of human text not only acquire linguistic competence but also encode, simulate, and express personality traits in their generated outputs. Systematic experimental evidence demonstrates that LLMs, when prompted appropriately, manifest internal structure and behavioral patterns analogous to core human personality dimensions—such as the Big Five (extraversion, agreeableness, conscientiousness, neuroticism, openness)—with personality signals that are measurable, modifiable, and relevant for alignment, safety, and user experience.

1. Psychometric Measurement and Validation

Psychometric personality assessment in LLMs adapts human instruments (notably the 300-item IPIP-NEO and the 44-item Big Five Inventory) through structured prompting. Each item is rendered in a context-rich form comprising a persona instruction, item instruction, and response options, typically on a Likert-type scale (e.g., 1–5). LLMs are evaluated by log-probability assignment over the discrete response options for each item, thus minimizing contamination by previous completions and controlling for item order effects.

Reliability is quantified by classical metrics: Cronbach’s α

$$\alpha = \frac{k}{k-1} \left(1-\frac{\sum \sigma^2_y}{\sigma^2_x}\right)$$

(where $k$ is item number, $\sigma^2_y$ item variance, and $\sigma^2_x$ total variance), as well as Guttman’s $\lambda_6$ and McDonald’s $\omega$ computed via confirmatory factor analysis. Validity is interrogated by convergent (high Pearson correlation between BFI and IPIP-NEO trait scores), discriminant, and criterion paradigms (correlating synthetic personality scores to downstream psychometric constructs such as affect and aggression). These procedures consistently show that model size and instruction tuning modulate reliability and validity: in large, instruction-finetuned LLMs, reliability metrics often exceed $\alpha, \lambda_6, \omega > 0.90$, and synthetic scores correlate strongly with downstream measures in expected directions.

2. Personality Shaping and Control

Prompt-based shaping is extended by introducing linguistic qualifiers adapted from Likert conventions and a lexicon of 104 Goldberg-derived trait adjectives. Single- and multi-trait shaping are supported: in the former, prompts handpick one trait and adjust its ordinal level (e.g., “extremely extraverted,” “somewhat introverted”); in the latter, all five Big Five traits are specified at extreme levels, resulting in $2^5 = 32$ personalities. The alignment between targeted and realized trait levels is measured by Spearman’s $\rho$ between prompt ordinal targets (1–9) and observed medians (typically $\rho \geq 0.90$). Empirically, manipulating prompts to specify higher extraversion causes the model’s trait medians to increase monotonically from near 1 to above 4.5 (on a 1–5 scale).

3. Model Characteristics Influencing Personality

Empirical studies consistently find that larger parameter counts and instruction-fine-tuning materially increase the internal consistency and validity of personality responses. Base models (e.g., a 62B parameter PaLM without tuning) yield low Cronbach’s $\alpha$ and inconsistent profiles, while instruction-tuned variants (e.g., Flan-PaLM) produce reliable, stable, and valid outputs. Trait simulation is also found to be robust to moderate prompt perturbations—but sensitivity to temperature and role-prompting may be higher in certain models (e.g., GPT-4), implying that generation parameters can modulate apparent personality expression.

Moreover, trait scores are not uniform across models or model families: for example, LLMs frequently show high openness and low extraversion (Hilliard et al., 13 Feb 2024), and fine-tuned conversational models (ChatGPT, ChatGLM) exhibit more human-like profiles and higher conscientiousness (Zhan et al., 11 Oct 2024). Dimensional variability—measured as the coefficient of variation (CV)—is notably high for neuroticism in some model families, indicating less stability of this trait.

4. Mechanisms and Granularity of Personality Editing

Beyond prompt-based interventions, a diverse suite of editing and control mechanisms is empirically validated:

• Model Editing and Feature Steering: Activation- or weight-level interventions (e.g., steering vectors based on activation centroid differences between two personality classes $$v_l = \mu_l(\mathcal{X}^+) - \mu_l(\mathcal{X}^-),\ h'_l = h_l + \alpha v_l$$) can modulate personality (e.g., ISTJ $\,\to\,$ ISTP) with substantial improvements in safety outcomes (43% increase in privacy, 10% in fairness) (Zhang et al., 17 Jul 2024).
• Unsupervised Lexicon-Based Decoding: Plugging lexicon-derived weights into decoding probability adjustment (e.g., $$R_2 = \{s_z \cdot (1 + \alpha \cdot \sum_t \beta_t \cdot L^t_\text{val}(s_z))\}$$ for token $s_z$) enables simultaneous, fine-grained, multi-trait manipulation (Li et al., 2023).
• Mixture-of-Experts Approaches: Personality-tailored mixture-of-experts leveraging LoRA modules and a personality specialization loss (e.g., $L_s = \sum_{i \neq j} |M^s_{ij}|$) ensure expert specialization and support flexible, trait-guided routing (Dan et al., 18 Jun 2024).
• Neuron-Based Induction: Pinpointing personality-correlated neurons by activation differences on opposing facets, then manipulating those neurons’ activations at inference (scaling by $\gamma$ and percentile activations, selectively clamping or boosting) achieves control competitive with full model fine-tuning without retraining (Deng et al., 16 Oct 2024).
• Model Merging/Personality Vector Approaches: By computing $\phi_p = \theta_p - \theta_\text{pre}$ between fine-tuned and base model weights, trait control is recast as a linear interpolation problem, supporting both continuous scaling ($\theta' = \theta_\text{base} + \alpha \cdot \phi_p$) and the composition of multiple traits, with demonstrated cross-model and even vision-LLM transferability (Sun et al., 24 Sep 2025).

Control Method	Level of Granularity	Scalability / Generalizability
Prompt-based shaping	Moderate; concept-level	High; no retraining
UBPL lexicon	Fine-grained, token-level	High; unsupervised & pluggable
Neuron-level editing	Very fine (neuron-level)	High; inference-time, no model update
Model merging (vectors)	Model-wide, continuous	Cross-domain (language, vision); scales well

5. Applications and Societal Implications

The psychometrically grounded control of LLM personality enables a series of practical enhancements:

• Human–AI interaction: Conversational agents and chatbots can be configured for stable, trait-specific engagement (e.g., high agreeableness for empathetic support, sufficiently controlled neuroticism to avoid negative affect propagation) (Serapio-García et al., 2023, Dan et al., 18 Jun 2024).
• Alignment and Auditing: Personality profiling tools serve as pre-deployment audit mechanisms to ensure conformance with ethical and social norms, such as the avoidance of trait configurations prone to toxicity or bias (Serapio-García et al., 2023, Zhang et al., 17 Jul 2024).
• Personalization: End users and application developers can select or dynamically adjust a model's personality profile to optimize user experience.
• Role-Playing, Simulation, and Storytelling: Scenario-driven applications benefit from stable, context-dependent, yet tunable character profiles—greatly expanding realism in gaming, education, and training.
• Safety: Personality editing can be actively harnessed for safety enhancement (e.g., steering towards profiles that are empirically less susceptible to jailbreak attacks) (Zhang et al., 17 Jul 2024).

6. Limitations, Variability, and Open Problems

LLMs express personality as a function of both their inherent parameters and contextual factors—prompt wording, parameter settings, and situational cues. Despite their ability to produce consistent trait expressions within restricted conditions, empirical studies reveal substantial test–retest variability and sensitivity to prompt or question variant, in stark contrast to humans’ high test–retest reliability and cross-variant consistency (Jiaqi et al., 1 May 2025). This motivates the “Distributed Personality” framework for LLMs: model outputs constitute a probability distribution over plausible trait scores, dynamically modulated by extrinsic inputs, without a fixed internal “core self.” Role-playing experiments further show that, unlike humans—who retain their baseline traits even when simulating a character—LLMs' simulated personalities are highly context-driven and do not anchor to any intrinsic baseline.

These phenomena indicate that LLM personality is best understood as emergent, situationally constructed, and distributional rather than static, raising challenges for both measurement and alignment.

7. Future Directions and Integration with Psychological Theory

The literature highlights multiple directions for further research:

• Refining evaluation metrics that better dissociate trait expression from fluency and topic adherence in generation (Mao et al., 2023).
• Advancing unsupervised and interpretable approaches (e.g., SVD over log-probabilities of descriptive adjectives) for latent trait discovery; for example, principal components analysis can “rediscover” the Big Five dimensions with notable explained variance (74.3%) and prediction accuracy gains (Suh et al., 16 Sep 2024).
• Exploring the intersection of trait modulation and ethical risk: model alignment interventions systematically shift personality traits (e.g., safety-tuned models trending more extraverted or judging) with implications for privacy and fairness (Zhang et al., 17 Jul 2024).
• Developing LLM-specific frameworks that fuse deep neural modeling, psychometrics, and dynamic, context-sensitive measurement (e.g., Cognitive-Affective Processing System analogs) (Jiaqi et al., 1 May 2025).
• Applying personality-informed steering to practical domains such as risk modeling, where trait interventions (e.g., adjusting Openness) can systematically alter risk propensity as formalized under cumulative prospect theory (Hartley et al., 3 Feb 2025).

Summary Table: Empirical Findings on LLM Personality

Finding	Description / Metric	Reference
Reliability via Cronbach’s α, λ₆, ω	High for large, instruction-tuned models (> 0.90)	(Serapio-García et al., 2023)
Personality can be reliably shaped	Spearman's $\rho \geq 0.90$ between targets/outputs	(Serapio-García et al., 2023)
Dominant LLM trait signatures	High openness, low extraversion seen in most models	(Hilliard et al., 13 Feb 2024Zhan et al., 11 Oct 2024)
Model size effect	Larger/fine-tuned models: more range variability	(Hilliard et al., 13 Feb 2024)
Measurement variability	High CV for neuroticism in some LLMs	(Bhandari et al., 7 Feb 2025)
Continous/multi-trait control	Achieved via UBPL, neuron-based, vector merging	(Li et al., 2023Deng et al., 16 Oct 2024Sun et al., 24 Sep 2025)
Personality–safety link	Trait editing (e.g. ISTJ→ISTP) improves privacy by 43%	(Zhang et al., 17 Jul 2024)
Distributional, input-dependent LLM personality	LLM outputs highly variable, context-sensitive	(Jiaqi et al., 1 May 2025)
Training data effect	Pre-training and instruction data amplify "personality"	(Zhan et al., 11 Oct 2024)
Latent trait discovery	SVD on log-probabilities recovers Big Five (74.3%)	(Suh et al., 16 Sep 2024)
Role in persuasion	Models align persuastive linguistic features with cue	(Mieleszczenko-Kowszewicz et al., 8 Nov 2024)

Conclusion:

LLMs encode, simulate, and reveal complex personality traits that are psychometrically measurable, modifiable, and reflective of both underlying architecture and context. These personality signatures profoundly affect user interaction, system alignment, and safety. However, personality in LLMs is best conceptualized as distributed, emergent, and context-bound—a distinction foundational for both scientific understanding and the responsible engineering of interactive artificial intelligence.