Localizing Persona Representations in LLMs

Celia Cintas; Miriam Rateike; Erik Miehling; Elizabeth Daly; Skyler Speakman

cs.CLMay 30, 2025

Localizing Persona Representations in LLMs

Authors: Celia Cintas, Miriam Rateike, Erik Miehling, Elizabeth Daly, Skyler Speakman

Organizations: IBM Research Africa · IBM Research Europe

Abstract

We present a study on how and where personas -- defined by distinct sets of human characteristics, values, and beliefs -- are encoded in the representation space of large language models (LLMs). Using a range of dimension reduction and pattern recognition methods, we first identify the model layers that show the greatest divergence in encoding these representations. We then analyze the activations within a selected layer to examine how specific personas are encoded relative to others, including their shared and distinct embedding spaces. We find that, across multiple pre-trained decoder-only LLMs, the analyzed personas show large differences in representation space only within the final third of the decoder layers. We observe overlapping activations for specific ethical perspectives -- such as moral nihilism and utilitarianism -- suggesting a degree of polysemy. In contrast, political ideologies like conservatism and liberalism appear to be represented in more distinct regions. These findings help to improve our understanding of how LLMs internally represent information and can inform future efforts in refining the modulation of specific human traits in LLM outputs. Warning: This paper includes potentially offensive sample statements.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of our study for Llama3 on persona topic Politics . (a) Examples of matchingbehavior ( + ) and notmatchingbehavior ( – ) persona statements. (b) PCA of representations for + and - sentences for anti-immigration (Q1) . (c) Overlap of sets of salient last-layer activations from + sentences, as identified by Deep Scan (Q2) . UpSet plot visualizes different sets of personas (connected dots on bottom) and how many salient activations are intersecting (vertical bars). We observe that the highest intersection corresponds to sentences from politically-conservative and politically-liberal .

Figure 2: (Q1) Euclidean distances between PCA convex hull centroids for matchingbehavior vs. notmatchingbehavior sentences averaged over Primary Personality Dimensions.

Topic	$\ell$	SH ( $\uparrow$ )	CH ( $\uparrow$ )	ED ( $\uparrow$ )	DB ( $\downarrow$ )
agree	1	0.500	$340.6^{\star}$	0.403	0.731
agree	31	0.792	3264.5	27.57	0.326
consc	1	0.635	718.8	0.370	0.569
consc	31	0.813	4150.4	27.47	0.285
open	1	0.602	570.2	0.414	0.645
open	31	0.795	3564.1	27.60	0.319

Table 1: (Q1) Separation of principal component representations in early (1) vs. late (31) layers (

\ell

) of Llama3 for Personality personas. Metrics: Silhouette (Si), Calinski-Harabasz (CH), Euclidean (ED), and Davies-Bouldin (DB). Results are averaged over five seeds (std=0.00, except

\star\approx 0.1

). Best result across layers and models.

Panel of Figure 3 — Figure 3: (Q2) Upset plots illustrating the overlap of sets of salient last-layer activations from matchingbehavior sentences, as identified by Deep Scan, across personas. Each bar represents the number of activations shared by a specific combination of personas.

Level	$H_{0}$	$H_{1}$	Precision ( $\uparrow$ )	Recall ( $\uparrow$ )
Level 2 Inter- Topic	consc -	consc +	$0.8387\pm 0.0399$	$0.8181\pm 0.0765$
	liber -	liber +	$0.8939\pm 0.0507$	$0.8056\pm 0.0769$
	immi -	immi +	$0.8167\pm 0.0507$	$0.8282\pm 0.0711$
	lgbtq -	lgbtq +	$0.9575\pm 0.0340$	$0.9365\pm 0.0684$
	extra -	extra +	$0.9457\pm 0.0268$	$0.8901\pm 0.0542$
	neuro -	neuro +	$0.9540\pm 0.0323$	$0.7565\pm 0.1142$

Table 2: ( Q1 , Q2 ) Validation of usefulness of salient activations

O_{S^{*}}

in detecting sentences

X_{S^{*}}

w.r.t. detection hypothesis

H_{1}

at different levels. matchingbehavior (+) and notmatchingbehavior (-) sentences. “all” indicating all other relevant personas, e.g., for Level 1 conse + , all

=\{{\textsc{liber}}^{+}

\cup

immi +

\cup

lgbtq

{}^{+}\}

; for Level 0 Politics + , all={ Ethics +

\cup

Personality + }. Mean

\pm

std over

100

indep. Deep Scan runs, and

200

random test samples. High / low detection power.

Figure 4: (Q2) Overlap of salient activations between topics and sampled persona from each topic.

Topic	Persona	Label
		matchingbehavior	notmatchingbehavior
Personality	agreeableness	Everyone should be treated with kindness, dignity and respect	I rudely insist on having things done my own way

	openness	It is important to have new experiences and do things that are new and different	I don’t appreciate ideas or thoughts that appear strange or different from usual convention

	extraversion	I enjoy social events and lively conversations	I am quiet, reserved, and seek out alone time

Table 3: Sentence examples for different personas ( Perez et al. 2023 ) .

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Llama3-8B-Instruct metrics across layers and dimensions.

Figure 6: (Q1) : Euclidean distances stratified by dimensions across models and layers.

Dimension	Layer	$\|O_{S^{*}}\|$	Precision ( $\uparrow$ )	Recall ( $\uparrow$ )
agree	1	$2815$	$0.5067\pm 0.1432$	$0.2469\pm 0.0904$
agree	31	1167	$\mathbf{0.9971\pm 0.0113}$	$\mathbf{0.9979\pm 0.00984}$
consc	1	3240	$0.3895\pm 0.1753$	$0.2237\pm 0.10208$
consc	31	1210	$\mathbf{0.9992\pm 0.0001}$	$\mathbf{0.95454\pm 0.04876}$
open	1	$3204$	$0.6048\pm 0.2317$	$0.2434\pm 0.10216$
open	31	1177	$\mathbf{0.9998\pm 0.0003}$	$\mathbf{0.97727\pm 0.0422}$

Appendix

Table 4: (Q1) : Validation of usefulness of the set of salient activations

O_{S^{*}}

at Layer 1 and 31 of Llama3-8B-Instruct .

Model	Dimension	layer	SH ( $\uparrow$ )	CH ( $\uparrow$ )	DB ( $\downarrow$ )	ED ( $\uparrow$ )
Llama3-8B-Instruct	agree	1	$0.500\pm 0.0741$	$340.6\pm 163.9$	$0.731\pm 0.072$	$0.403\pm 0.012$
	agree	31	$0.792\pm 0.0000$	$3264.5\pm 0.002$	$0.326\pm 0.000$	$27.57\pm 0.000$
	consc	1	$0.635\pm 0.0000$	$718.8\pm 0.026$	$0.569\pm 0.000$	$0.370\pm 0.000$
	consc	31	$0.813\pm 0.0000$	$4150.4\pm 0.003$	$0.285\pm 0.000$	$27.47\pm 0.000$
	open	1	$0.602\pm 0.0000$	$570.2\pm 0.0005$	$0.645\pm 0.000$	$0.414\pm 0.000$
	open	31	$0.795\pm 0.0000$	$3564.1\pm 0.0125$	$0.319\pm 0.000$	$27.60\pm 0.000$

Appendix

Table 5: (Q1) : Metrics to assess the goodness of layer 1 and 31 to detect personas (Level 2) in Llama3-8B-Instruct . Metrics: Silhouette Score (SH), Calinski-Harabasz Score (CH), Euclidean Distance (ED), Davies-Bouldin Score (DB).

Model	Dimension	layer	SH ( $\uparrow$ )	CH ( $\uparrow$ )	DB ( $\downarrow$ )	ED ( $\uparrow$ )
Mistral-7B-Instruct	agree	1	$0.370\pm 0.082$	$176.5\pm 110.62$	$0.690\pm 0.084$	$0.110\pm 0.006$
	agree	31	$0.251\pm 0.000$	$164.1\pm 0.00$	$1.743\pm 0.0008$	$5.403\pm 0.003$
	consc	1	$0.242\pm 0.113$	$96.8\pm 17.79$	$1.753\pm 0.505$	$0.063\pm 0.02$
	consc	31	$0.171\pm 0.010$	$114.6\pm 9.23$	$2.140\pm 0.086$	$4.249\pm 0.122$
	open	1	$0.423\pm 0.020$	$176.8\pm 13.84$	$1.006\pm 0.596$	$0.107\pm 0.023$
	open	31	$0.235\pm 0.000$	$194.2\pm 0.01$	$1.693\pm 0.000$	$5.147\pm 0.000$

Appendix

Table 6: (Q1) : Metrics to assess the goodness of layer 1 and 31 to detect personas (Level 2) in Mistral-7B-Instruct . Metrics: Silhouette Score (SH), Calinski-Harabasz Score (CH), Euclidean Distance (ED), Davies-Bouldin Score (DB).

Model	Dimension	layer	SH ( $\uparrow$ )	CH ( $\uparrow$ )	DB ( $\downarrow$ )	ED ( $\uparrow$ )
Granite-7B-Instruct	agree	1	$0.420\pm 0.0312$	$210.7\pm 75.01$	$0.925\pm 0.225$	$0.287\pm 0.019$
	agree	31	$0.611\pm 0.0000$	$969.0\pm 0.000$	$0.605\pm 0.000$	$17.62\pm 0.000$
	consc	1	$0.561\pm 0.0000$	$578.4\pm 0.003$	$0.709\pm 0.000$	$0.250\pm 0.000$
	consc	31	$0.625\pm 0.0000$	$1033.5\pm 0.002$	$0.573\pm 0.000$	$16.83\pm 0.000$
	open	1	$0.491\pm 0.0000$	$312.4\pm 0.000$	$0.912\pm 0.000$	$0.265\pm 0.000$
	open	31	$0.991\pm 0.0000$	$37132.7\pm 0.038$	$0.004\pm 0.000$	$2073.7\pm 0.000$

Appendix

Table 7: (Q1) : Metrics to assess the goodness of layer 1 and 31 to detect personas (Level 2) in Granite-7B-Instruct . Metrics: Silhouette Score (SH), Calinski-Harabasz Score (CH), Euclidean Distance (ED), Davies-Bouldin Score (DB).

Panel of Figure 7 — Figure 7: (Q1) : Examples of changes in the representation space of the dimension openness in a Granite-7B-Instruct model. Positive conscientiousness referring to matchingbehavior , negative conscientiousness referring to notmatchingbehavior . We observe better separability in later layers.

Panel of Figure 8 — Figure 8: (Q1) : Examples of changes in the representation space of the dimension conscientiousness in a Mistral-7B-Instruct model. Positive conscientiousness referring to matchingbehavior , negative conscientiousness referring to notmatchingbehavior . We observe better separability in later layers.

Method	Dimension	$\|{e}\|$ ( $\downarrow$ )	Precision ( $\uparrow$ )	Recall ( $\uparrow$ )
LOF	agree	4096	$0.5072\pm 0.0288$	$0.9904\pm 0.0058$
	consc	4096	$0.5033\pm 0.0359$	$\mathbf{1.0\pm 0.0}$
	open	4096	$0.5049\pm 0.0380$	$\mathbf{0.9861\pm 0.0112}$
	extra	4096	$0.4867\pm 0.0278$	$\mathbf{0.9933\pm 0.007}$
	neuro	4096	$0.4880\pm 0.0275$	$\mathbf{1.0\pm 0.0}$
IF	agree	4096	$0.5094\pm 0.0313$	$\mathbf{0.9980\pm 0.0038}$

Appendix

Table 8: Detection capabilities for positive and negative directions in the personality big five: agreeableness ( agree ), conscientiousness ( consc ), openness ( open ), extraversion ( extra ), neuroticism ( neuro ). Results over Llama3-8B-Instruct activations in layer 31. Comparing Local Outlier Factor (LOF) method ( Breunig et al. 2000 ) , Isolation Forest (IF) and Kmeans to Deep Scan used in this study.

|e|

is the size of the activation vector defining the clusters, for Deep Scan,

|e|=|O_{S^{*}}|

Panel of Figure 9 — Figure 9: (Q2) : Upset plots (left) and Venn diagrams (right) for intra-persona (Level 2) analysis with personas from topics: (a, b) Personality . (c, d) Ethics . (e, f) Politics .

Figure 10: (Q2) : Venn Diagram for Inter-Topic (Level 0) analysis.

LLM Interpretability

Large language model (LLM) interpretability studies how internal representations and computations produce outputs, and how they can be explained or controlled.

See all 778 papers

Political Ideology Shifts in Large Language Models
Aug 22, 2025Pietro Bernardelle, Stefano Civelli, Leon Fröhling +3Political Bias in Language Models Language Model Bias Evaluation
The University of Queensland, Brisbane, Australia · GESIS, Cologne, Germany · University of Udine, Udine, Italy
Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas
May 9, 2026Nils A. Herrmann, Leander Girrbach, Kirill Bykov +1Language Model Steering LLM Interpretability
1Technical University of Munich · 2Helmholtz Munich · 3Munich
Persona Cartography: Charting Language Model Personality Traits in Weight Space
Jul 8, 2026Luke Baines, Anton Gonzalvez Hawthorne, Mariia Koroliuk +4Personality Modeling in Language Models
LASR Labs · ENS Paris-Saclay and MATS · UK AI Security Institute

Aug 22, 2025J/K move · Enter open · S save

Political Ideology Shifts in Large Language Models

Pietro Bernardelle, Stefano Civelli, Leon Fröhling, Riccardo Lunardi+2

The University of Queensland, Brisbane, Australia · GESIS, Cologne, Germany · University of Udine, Udine, Italy

Political Bias in Language Models Language Model Bias Evaluation Persona-Conditioned Generation

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly deployed in politically sensitive contexts, raising concerns about their susceptibility to ideological biases. In this work, we examine how synthetic persona conditioning shapes ideological expression across seven open-weight instruction-tuned models (7B-72B parameters) using the Political Compass Test (62 statements) as a standardized behavioral probe. Across three studies involving 200,000 synthetic personas and more than 260 million model responses, we analyze implicit and explicit malleability, as well as theme-associated variations. We find that: (i) larger models exhibit broader implicit ideological coverage, increasing from 14-35% for 7-8B models to up to 49% for 70B+ models; (ii) explicit ideological priming induces large and statistically significant shifts, with right-authoritarian cues moving all models in the intended direction and producing larger effects in most model-axis comparisons; (iii) left-libertarian priming produces more heterogeneous responses, including counter-directional economic shifts in three of four 7-8B models, while all 70B+ models move in the intended direction; and (iv) theme-associated semantic content in persona descriptions is linked to systematic and interpretable directional shifts in ideological space. While our results identify an upstream mechanism through which persona conditioning can alter model responses under a standardized ideological probe, we do not test whether such shifts affect users beliefs, decisions, or political behavior. Our findings are best understood as evidence of ideological malleability at the generation layer, highlighting the need to account for interactional factors when evaluating political neutrality, fairness, and safety in English-prompted, persona-conditioned language models.

Abstract

LLM Interpretability

Explore similar work

Political Ideology Shifts in Large Language Models

Do LLMs Experience an Internal Polylogue? Investigating Reasoning through the Lens of Personas

Persona Cartography: Charting Language Model Personality Traits in Weight Space