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

Jiaqi Chen; Ming Wang; Tingna Xie; Shi Feng; Yongkang Liu

cs.CLApr 13, 2026

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

Authors: Jiaqi Chen, Ming Wang, Tingna Xie, Shi Feng, Yongkang Liu

Organizations: 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

Abstract

Imbuing Large Language Models (LLMs) with specific personas is prevalent for tailoring interaction styles, yet the impact on underlying cognitive capabilities remains unexplored. We employ the Neuron-based Personality Trait Induction (NPTI) framework to induce Big Five personality traits in LLMs and evaluate performance across six cognitive benchmarks. Our findings reveal that persona induction produces stable, reproducible shifts in cognitive task performance beyond surface-level stylistic changes. These effects exhibit strong task dependence: certain personalities yield consistent gains on instruction-following, while others impair complex reasoning. Effect magnitude varies systematically by trait dimension, with Openness and Extraversion exerting the most robust influence. Furthermore, LLM effects show 73.68% directional consistency with human personality-cognition relationships. Capitalizing on these regularities, we propose Dynamic Persona Routing (DPR), a lightweight query-adaptive strategy that outperforms the best static persona without additional training.

Figures & tables

Figure 1: The systematic analysis pipeline for quantifying persona steering effects.

Cross-Architecture
Model	GPQA	BBH	MuSR	MMLU	IFEval	GSM8K
LLaMA-3-8B-it	30.13	63.40	54.10	36.55	39.93	72.18
Gemma-2-9B-it	29.02	70.96	57.67	50.38	47.87	58.83
Mistral-7B-it-v0.3	30.36	45.92	45.37	33.36	34.20	45.49
Qwen2.5-7B-it	27.68	66.73	50.40	55.09	44.55	81.58
Scaling Analysis

Table 1: Base accuracy on six benchmarks (unit: %).

Panel of Figure 2 — Figure 2: Persona-task interaction heatmap showing $\Delta\Acc$ and direction consistency ( $\SA$ ).

Panel of Figure 3 — Figure 3: Domain-specific persona effects in Qwen2.5 family (0.5B–14B). Panel (a) integrates effect magnitude (line height) and statistical evidence (bubble size) to reveal both the strength and reliability of persona effects across scales.

Trait	Impact	Avg. Gap	Uniformity	Rank
Openness (O)	11.96%	+11.17%	90.5%	1
Extraversion (E)	11.70%	+10.42%	90.5%	1
Agreeableness (A)	8.06%	+6.50%	73.8%	4
Conscientiousness (C)	6.31%	+5.82%	88.1%	3
Neuroticism (N)	4.00%	$-$ 1.86%	57.1%	5

Table 2: Impact and Uniformity Across Big Five Dimensions

Dataset	Total	Sampled	Correct	Accuracy (%)	Best Baseline (%)
GPQA	448	44	23	52.27	41.96
BBH	6511	651	404	62.06	63.40
MuSR	756	75	59	78.67	54.10
MMLU-Pro	12032	1203	519	43.14	36.67
IFEVal	541	54	34	62.96	62.48
GSM8K	1319	131	88	67.17	72.18

Table 3: Dynamic Persona Routing vs. Best Static Baseline

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 275 papers

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
Experiences Build Characters: The Linguistic Origins and Functional Impact of LLM Personality
Mar 6, 2026Xi Wang, Mengdie Zhuang, Jiqun LiuPersonality Modeling in Language Models Language Modeling
University of Sheffield, Sheffield, UK · University of Oklahoma, Oklahoma, USA
PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra
Feb 17, 2026Xiachong Feng, Liang Zhao, Weihong Zhong +5Controllable Text Generation Personality Modeling in Language Models
Harbin Institute of Technology · The University of Hong Kong

May 16, 2026J/K move · Enter open · S save

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

Prateek Rajput, Yewei Song, Iyiola E. Olatunji, Jacques Klein+1

University of Luxembourg, Esch-sur-Alzette, Luxembourg

Personality Modeling in Language Models LLM Post-Training Persona-Conditioned Generation Language Model Robustness

Open paper AskPDF arXiv

Can large language models reliably express a human-like personality, or are they merely mimicking surface cues without a stable underlying profile? To investigate this, we induce personality in LLMs by fine-tuning them on the long-form essays, where each essay is associated with a target Big Five personality profile. We then evaluate the stability and fidelity of the induced personality using the IPIP-NEO questionnaire. Specifically, we ask: (i) does post-training (SFT, DPO, ORPO) stabilize questionnaire scores under prompt rephrasings, and (ii) can it induce target Big Five profiles from unguided essays? Our results demonstrate that fine-tuning consistently reduces variance in questionnaire responses across five models, directly mitigating the evaluation fragility reported in pre-trained models. However, this newfound stability reveals a more fundamental limitation: accuracy on the full five-dimensional profile remains near chance, even when single-trait scores improve. This indicates that unguided essays lack the cues needed for faithful personality expression. We therefore argue for scenario-grounded datasets or interactive elicitation that accumulates test-aligned evidence over time.

Abstract

Language Model Steering

Explore similar work

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

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

PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra