Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality

Xi Wang; Mengdie Zhuang; Jiqun Liu

cs.CLMar 6, 2026

Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality

Authors: Xi Wang, Mengdie Zhuang, Jiqun Liu

Organizations: University of Sheffield, Sheffield, UK · University of Oklahoma, Oklahoma, USA

Abstract

Human problem-solving is enriched by a diversity of styles and personality traits, yet the development of Large Language Models (LLMs) has largely prioritized uniform performance benchmarks that favour specific behavioural tendencies such as assertiveness. To investigate how diverse experiences shape machine personality and influence problem-solving, this study employs continued pre-training to expose models to domain-specific texts in an unsupervised manner, simulating the accumulation of experience. By adapting the Big Five framework via the Machine Personality Inventory (MPI), we quantify the personality traits of these model variants and analyse their relationship to linguistic style and reasoning behaviour. The findings reveal that model competence is bimodal, peaking at "Expressive Generalists" and "Suppressed Specialists," while identifying a "Suppression Advantage" where reduced social traits enhance complex reasoning performance. This study further establishes a causal link between training data linguistics, such as imperative frequency, and lexical diversity, providing a roadmap for "Personality Engineering".

Figures & tables

	MMLU							MMLU-Pro
Model	STEM	Health	Soc. Sci.	Law	Humanit.	Psych.	Other	STEM	Health	Soc. Sci.	Law	Humanit.	Psych.	Other
Llama-3-8b	0.5302	0.6975	0.7124	0.7389	0.6821	0.7745	0.8186	0.3039	0.4920	0.3949	0.2189	0.3959	0.5401	0.4123
Literary Classicist	0.4843	0.6444	0.6656	0.7172	0.6534	0.7440	0.7740	0.2716	0.4761	0.3825	0.2443	0.3463	0.5263	0.3669
Inventive Technologist	0.5067	0.6835	0.6843	0.7219	0.6720	0.7451	0.8084	0.2469	0.4635	0.3183	0.2089	0.3565	0.4937	0.3788
Technical Communicator	0.5165	0.6849	0.6851	0.7395	0.6854	0.7629	0.8072	0.2743	0.4802	0.3533	0.2134	0.3652	0.5150	0.4015
Business Advisor	0.5073	0.6869	0.6821	0.7274	0.6764	0.7455	0.8033	0.2395	0.4637	0.3438	0.1944	0.3879	0.4975	0.3820

Table 1: Averaged accuracy of LLM variants across unified domain groups on MMLU and MMLU-Pro . Values are grouped by domain mappings (Appendix Table 5 ); bold numbers indicate the best-performing variant per domain.

Model	Openness		Conscientiousness		Extraversion		Agreeableness		Neuroticism
Model	Score	$\sigma$	Score	$\sigma$	Score	$\sigma$	Score	$\sigma$	Score	$\sigma$
Meta-Llama-3-8B	3.13	1.44	3.25	1.34	3.33	1.32	3.07	1.39	3.14	1.41
literary_classicist	3.12	1.14	3.19	1.33	3.16	1.19	3.03	1.24	2.75	1.15
inventive_technologist	2.93	1.12	3.13	1.14	2.91	1.17	3.11	1.14	2.87	1.14
patent_strategist	3.13	1.41	3.30	1.37	3.02	1.42	3.01	1.40	2.84	1.45
cultural_scholar	2.88	1.33	2.93	1.30	2.93	1.32	3.01	1.30	3.06	1.37

Table 2: LLMs’ personality analysis based on 1k-item MPI . Scores are means, and

\sigma

are standard deviations across answered items. The five personality dimensions are: Openness (artistic, curious, imaginative, original), Conscientiousness (efficient, organised, reliable, responsible), Extraversion (active, assertive, outgoing, talkative), Agreeableness (appreciative, kind, sympathetic), and Neuroticism (anxious, tense, worrying).

Panel of Figure 2 — Figure 2: The Correlation of Personality Effects under Complexity.

Model Variant	Imperative	TTR	Complexity	Sentiment	Detachment
Literary Classicist	0.024	0.073	17.08	0.244	0.022
Inventive Technologist	0.046	0.079	17.51	0.185	0.001
Patent Strategist	0.025	0.064	21.61	0.154	0.012
Cultural Scholar	0.008	0.124	22.09	0.184	0.020
Technical Communicator	0.052	0.069	14.85	0.233	-0.022
Business Advisor	0.061	0.082	15.32	0.214	-0.023

Table 3: Linguistic Signals of Training Corpora. We analysed 200k-token samples from each domain corpus over five measures to capture the relationships between linguistic nuance and model performance.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Description
ArXiv	Scientific preprints across STEM disciplines.
PhilPapers	Philosophical literature and academic writing.
NIH Exporter	Biomedical research funding abstracts.
DM Mathematics	Theoretical and applied mathematics papers.
FreeLaw	U.S. legal case documents and opinions.
Gutenberg	Classic and public-domain literary works.

Appendix

Table 4: Overview of domain-specific datasets (top) and their corresponding use by character (role) variants (bottom) in continued pretraining.

Domain Group	MMLU Categories	MMLU-Pro Categories
STEM	abstract_algebra, astronomy, college_chemistry, college_computer_science, college_mathematics, college_physics, computer_security, conceptual_physics, electrical_engineering, elementary_mathematics, formal_logic, high_school_computer_science, high_school_mathematics, high_school_physics, high_school_statistics, machine_learning	Mathematics, Physics, Chemistry, Engineering, Computer Science
Health & Biology	anatomy, clinical_knowledge, college_biology, college_medicine, human_aging, medical_genetics, nutrition, professional_medicine, virology	Biology, Health
Social Science	business_ethics, econometrics, high_school_macroeconomics, high_school_microeconomics, management, marketing, professional_accounting, sociology, public_relations	Economics, Business
Law & Policy	international_law, jurisprudence, professional_law, security_studies, us_foreign_policy, high_school_government_and_politics	Law
Humanities	philosophy, moral_disputes, moral_scenarios, prehistory, high_school_us_history, high_school_world_history, high_school_european_history, world_religions, global_facts, logical_fallacies	Philosophy, History
Psychology	high_school_psychology, professional_psychology, human_sexuality	Psychology

Appendix

Table 5: Unified domain grouping between MMLU and MMLU-Pro. MMLU’s 57 fine-grained categories are aggregated into seven high-level domains aligned with MMLU-Pro’s broader taxonomy, enabling consistent domain-level comparison across benchmarks. A dashed line separates STEM, health, and social domains from humanities and behavioural domains.

Personality Modeling in Language Models

Personality modeling in language models studies how traits are measured, represented, simulated, and controlled in model behavior.

See all 162 papers

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions
Jun 27, 2026David Courtis, Ting HuLanguage Model Steering LLM Interpretability
Department of Computing, Queen’s University Kingston, Kingston, ON, Canada.
A Systematic Analysis of the Impact of Persona Steering on LLM Capabilities
Apr 13, 2026Jiaqi Chen, Ming Wang, Tingna Xie +2Language Model Steering Personality Modeling in Language Models
School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · School of Computing and Information Systems, Singapore Management University, Singapore 178902, Singapore · School of Computer and Communication Engineering, Northeastern University, Qinhuangdao 066004, China
Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?
May 16, 2026Prateek Rajput, Yewei Song, Iyiola E. Olatunji +2Personality Modeling in Language Models LLM Post-Training
University of Luxembourg, Esch-sur-Alzette, Luxembourg

Jun 27, 2026J/K move · Enter open · S save

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

David Courtis, Ting Hu

Department of Computing, Queen’s University Kingston, Kingston, ON, Canada.

Language Model Steering LLM Interpretability Personality Modeling in Language Models Causal Interventions in Language Models Activation Steering

Open paper AskPDF arXiv

Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.

Abstract

Personality Modeling in Language Models

Explore similar work

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

A Systematic Analysis of the Impact of Persona Steering on LLM Capabilities

Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?