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

Xiachong Feng; Liang Zhao; Weihong Zhong; Yichong Huang; Yuxuan Gu; Lingpeng Kong; Xiaocheng Feng; Bing Qin

cs.AIFeb 17, 2026

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

Authors: Xiachong Feng, Liang Zhao, Weihong Zhong, Yichong Huang, Yuxuan Gu, Lingpeng Kong, Xiaocheng Feng, Bing Qin

Organizations: Harbin Institute of Technology · The University of Hong Kong

Abstract

Current methods for personality control in Large Language Models rely on static prompting or expensive fine-tuning, failing to capture the dynamic and compositional nature of human traits. We introduce PERSONA, a training-free framework that achieves fine-tuning level performance through direct manipulation of personality vectors in activation space. Our key insight is that personality traits appear as extractable, approximately orthogonal directions in the model's representation space that support algebraic operations. The framework operates through three stages: Persona-Base extracts orthogonal trait vectors via contrastive activation analysis; Persona-Algebra enables precise control through vector arithmetic (scalar multiplication for intensity, addition for composition, subtraction for suppression); and Persona-Flow achieves context-aware adaptation by dynamically composing these vectors during inference. On PersonalityBench, our approach achieves a mean score of 9.60, nearly matching the supervised fine-tuning upper bound of 9.61 without any gradient updates. On our proposed Persona-Evolve benchmark for dynamic personality adaptation, we achieve up to 91% win rates across diverse model families. These results provide evidence that aspects of LLM personality are mathematically tractable, opening new directions for interpretable and efficient behavioral control.

Figures & tables

Panel of Figure 1 — Figure 1: The PERSONA framework.

Dimension	Abbr.	High Pole	Low Pole
Openness	O	v_Inventive	v_Consistent
Conscientiousness	C	v_Dependable	v_Careless
Extraversion	E	v_Outgoing	v_Solitary
Agreeableness	A	v_Compassionate	v_Self-interested
Neuroticism	N	v_Nervous	v_Calm

Table 1: OCEAN personality vectors with opposing trait poles for each dimension.

Trait	Type	Steering Coefficient ( $\alpha$ )
		$-1.0$	$0.0$	$+1.0$	$+2.0$
Inventive	O+	42.3	63.3	88.4	96.1
Consistent	O-	27.2	51.1	69.2	79.5
Dependable	C+	68.9	93.5	93.7	92.7
Careless	C-	0.7	2.8	83.8	96.2
Outgoing	E+	23.2	45.4	85.0	97.7

Table 2: GPT-4.1-mini evaluated trait expression scores (0-100) for responses steered with varying coefficients. Darker shades indicate stronger expression.

Figure 2: Cosine similarity between persona vectors.

Figure 3: Linear relationship between steering coefficients and BFI dimension scores. All vectors except dependable show strong linear modulation (high R²). We conjecture the dependable vector’s saturation stems from baseline model optimization for conscientiousness.

Figure 4: BFI-44 score changes after vector arithmetic operations. Y-axis shows operation and target dimension with expected direction (arrows). Grey: baseline scores; colored: post-steering scores (green for expected increases, red for decreases).

Panel of Figure 5 — Figure 5: An example of Persona-Evolve , together with the comparison between vanilla answer and answer steered by Persona-Flow . By reflecting on the scenario and then suppressing conscientiousness accordingly, Persona-Flow can produce a more natural and contextually appropriate response that align better with the anticipated feeling of being overwhelmed.

Family	Model	TA	RC	RA	IF	Overall
Qwen2.5	3B-Instruct	78.0	79.1	78.9	61.3	78.4
	7B-Instruct	84.7	84.4	85.0	61.4	83.4
	14B-Instruct	84.8	86.4	84.8	59.3	85.4
Qwen3	4B-Instruct	92.2	90.6	92.4	49.1	90.8
Llama	3.1-8B-Instruct	84.9	81.4	85.6	57.2	83.5
Mistral	Ministral-8B	74.3	73.2	74.2	48.0	73.2

Table 3: Win rates (%) of PERSONA-Flow vs. base model on PERSONA-Evolve . TA : Trait Adherence, RC : Role Consistency, RA : Response Authenticity, IF : Information Fidelity. Bold : Best in column or overall score

\geq

85%

Big-Five	PERSONA-BASE		NPTI		Simple Prompt		$P^{2}$		PAS		ActAdd		SFT
	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$	mean $\uparrow$	variance $\downarrow$
Agreeableness	9.69	0.71	9.64	0.49	9.72	0.34	9.68	0.42	6.48	1.01	8.20	2.90	9.87	0.25
Conscientiousness	9.26	0.94	9.25	0.66	9.24	1.06	9.24	1.18	6.69	1.63	6.61	2.75	9.23	0.85
Extroversion	9.45	0.85	9.86	0.14	9.50	1.02	9.46	0.68	7.57	2.81	8.84	1.44	9.86	0.15
Neuroticism	9.79	0.59	9.92	0.07	7.18	1.22	9.54	0.66	6.98	1.58	8.90	1.78	9.42	0.75
Openness	9.81	0.60	8.50	1.08	6.31	1.14	9.21	1.19	6.93	1.52	8.52	1.83	9.66	0.44

Table 4: Personality control performance on LLaMA-3-8B-Instruct. Bold/underlined values show best/second results among training-free methods (SFT serves as upper bound). Mean: sum of scores for opposing trait aspects; Variance: sum of variances for opposing aspects.

Model	MMLU (Acc)	TruthfulQA (Acc)
Qwen2.5-7B-Instruct	0.71	0.63
+ Persona-Flow	0.70	0.66
Qwen2.5-4B-Instruct	0.66	0.52
+ Persona-Flow	0.67	0.54
Llama-3.1-8B-Instruct	0.64	0.53
+ Persona-Flow	0.64	0.53

Table 5: Impact of Persona-Flow on general capabilities. Results show accuracy on MMLU and TruthfulQA benchmarks with and without personality steering. The method preserves or slightly improves performance across three model families.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Contrastive Prompt Generator	Target Model	PersonalityBench Mean Score
Claude 3.7 Sonnet Thinking (main)	LLaMA-3-8B	9.60
Qwen2.5-7B-Instruct	LLaMA-3-8B	9.28
Qwen2.5-3B-Instruct	LLaMA-3-8B	9.05
Qwen2.5-1B-Instruct	LLaMA-3-8B	8.93

Appendix

Table 6: Impact of contrastive prompt generator model on PersonalityBench performance. Vectors extracted using prompts from different generators are applied to LLaMA-3-8B and evaluated on PersonalityBench.

Dimension	Pole	Description
Openness	Inventive (O+)	Values novelty, creativity, and abstract thinking; prefers unconventional approaches, intellectual curiosity, and exploring new ideas; embraces change and complexity
Openness	Consistent (O-)	Prefers routine, tradition, and practicality; values established methods, concrete solutions, and proven approaches; favors stability and familiarity
Conscientiousness	Dependable (C+)	Organized, disciplined, and goal-oriented; emphasizes planning, reliability, and attention to detail; values structure and responsibility
Conscientiousness	Careless (C-)	Spontaneous, flexible, and less concerned with details; prioritizes adaptability over structure; comfortable with improvisation and minimal planning
Extraversion	Outgoing (E+)	Socially energetic, talkative, and assertive; seeks social interaction, external stimulation, and group activities; thrives in collaborative environments
Extraversion	Solitary (E-)	Reserved, independent, and introspective; prefers individual activities, quiet reflection, and limited social engagement; recharges through solitude

Appendix

Table 7: Detailed definitions of personality trait vectors used in PERSONA . Each trait pole is characterized by its core values, preferences, and behavioral tendencies.

Layer	Openness		Conscient.		Agreeable.		Neuroticism		Extraversion		AVG
	O+	O-	C+	C-	A+	A-	N+	N-	E+	E-
5	66.5	52.7	93.9	3.6	91.3	7.6	14.1	96.6	48.3	30.7	50.5
10	70.4	55.0	94.1	9.0	91.8	8.0	17.2	96.5	56.5	31.5	53.0
15	78.8	59.8	93.8	15.6	92.6	8.7	19.5	96.3	65.6	38.6	56.9
20	88.6	68.6	93.7	84.4	95.8	10.7	45.5	96.5	84.4	45.6	71.4
25	75.3	59.8	94.0	17.5	93.6	8.2	21.2	96.5	63.9	37.1	56.7

Appendix

Table 8: Layer-wise ablation study for single-layer steering on Qwen2.5-7B-Instruct. Trait expression scores (0-100) are evaluated using GPT-4.1-mini on held-out questions with steering coefficient

\alpha=1.0

. Bold indicates the best score for each trait. Layer 20 achieves the highest average performance across all traits.

Trait	Evaluator	Steering Coefficient ( $\alpha$ )
		$-1.0$	$0.0$	$+1.0$	$+2.0$
Inventive (O+)	GPT-4.1-mini	42.3	63.3	88.4	96.1
	Claude-4.5-sonnet	40.1	65.0	87.2	95.5
	Gemini-2.5-pro	43.5	62.9	89.1	96.8
Consistent (O-)	GPT-4.1-mini	27.2	51.1	69.2	79.5
	Claude-4.5-sonnet	25.5	50.8	70.1	80.3

Appendix

Table 9: Multi-evaluator validation of trait expression scores across varying steering coefficients. Three independent LLM judges from different organizations evaluate the same steered responses. High inter-evaluator agreement confirms that observed patterns (monotonic increases, ceiling effects, alignment resistance) reflect genuine properties of the steered model rather than evaluator-specific biases.

Layer	$\alpha_{\text{Outgoing}}$	$\alpha_{\text{Compassionate}}$	E Score	A Score	$\Delta$ E	$\Delta$ A
20	1	-1	84.2±1.5	85.3±1.3	-0.8	-5.7
20	1	0	85.0±1.1	91.0±0.9	0	0
20	1	1	85.6±1.3	96.2±1.0	+0.6	+5.2
20	1	2	86.1±1.6	98.4±1.2	+1.1	+7.4
19	1	-1	82.5±1.4	84.2±1.2	-0.7	-5.5
19	1	0	83.2±1.2	89.7±1.0	0	0

Appendix

Table 10: Controlled multi-trait intervention results demonstrating causal independence. With

\alpha_{\text{Outgoing}}

held constant at 1.0, Extraversion scores remain stable across all Compassionate coefficient variations (

\Delta

\leq

1.2 points), while Agreeableness scores show clear linear modulation (

\Delta

A = 5.0-5.7 points per unit

\alpha

). Results span five layers with 95% confidence intervals from three independent runs. Baseline rows (

\alpha_{\text{Compassionate}}=0

) are highlighted.

Trait	N	GPT-4.1	Claude-4.5	Human Mean	$r_{\text{GPT}}$	$r_{\text{Claude}}$
		-mini	-sonnet	(SD)	Human	Human
Inventive (O+)	20	3.45	3.40	3.42 (0.22)	0.92	0.91
Consistent (O-)	20	3.10	3.05	3.13 (0.24)	0.91	0.89
Dependable (C+)	20	4.35	4.25	4.28 (0.19)	0.88	0.87
Careless (C-)	20	2.15	2.25	2.22 (0.31)	0.90	0.92
Outgoing (E+)	20	3.50	3.45	3.47 (0.23)	0.94	0.92

Appendix

Table 11: Human validation of LLM-based evaluation against expert ratings. Each trait shows average Likert scores (1-5 scale) across 20 responses, evaluated by GPT-4.1-mini, Claude-sonnet-4.5, and three human experts (mean

\pm

SD). High Pearson correlations (r

>

0.87) between both AI judges and human consensus validate the reliability of LLM-based evaluation and mitigate evaluation circularity concerns.

Configuration	Value
Base Model	LLaMA-3-8B-Instruct
Judge Model	GPT-4o (Following NPTI evaluation methodology)
Judge Prompt	[System] Please act as an impartial judge and evaluate the personality trait of {BFI} exhibited in the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as {factors} demonstrated in the response. Begin your evaluation by providing a short explanation of no more than 30 words. Be as objective as possible. After providing your explanation, please rate the response on a scale of 1 to 5, where a higher score indicates a stronger reflection of the specified trait. Please follow this format for your rating: “[[rating]]”, for example: “Rating: [[5]]”. [Question] {question} [The Start of Assistant’s Answer] {answer} [The End of Assistant’s Answer]
Response Length	Max 512 tokens

Appendix

Table 12: Standardized evaluation configuration for training-free baseline comparison on PersonalityBench.

	Inventive		Consistent		Dependable		Careless		Outgoing		Solitary		Compassionate		Antagonistic		Nervous		Calm
Method	M	V	M	V	M	V	M	V	M	V	M	V	M	V	M	V	M	V	M	V
PERSONA-BASE	4.91	0.30	4.90	0.30	4.46	0.50	4.80	0.44	4.80	0.40	4.65	0.45	4.94	0.30	4.75	0.41	4.90	0.30	4.89	0.29
NPTI	4.25	0.54	4.25	0.54	4.45	0.36	4.80	0.30	4.98	0.07	4.88	0.07	4.92	0.20	4.72	0.29	4.96	0.03	4.96	0.04
ActAdd	4.26	0.91	4.26	0.92	3.10	1.40	3.51	1.35	4.50	0.72	4.34	0.72	4.20	1.40	4.00	1.50	4.50	0.90	4.40	0.88
Simple Prompt	4.61	0.60	4.60	0.59	4.44	0.60	4.80	0.58	4.81	0.34	4.65	0.34	4.93	0.20	4.75	0.22	4.77	0.33	4.77	0.33

Appendix

Table 13: Disaggregated pole-level performance on PersonalityBench for LLaMA-3-8B-Instruct under the controlled evaluation harness. All methods evaluated with identical configuration. Scores are on 1-5 scale. Bold indicates best training-free method per pole.

Trait	Type	Baseline Score	Activation Score	Activation Success
		( $\alpha=0.0$ )	( $\alpha=+1.0$ )	( $\Delta$ )
Compassionate (A+)	Aligned	90.8	95.9	+5.1
Self-interested (A-)	Resisted	7.7	12.6	+4.9
Dependable (C+)	Ceiling	93.5	93.7	+0.2
Careless (C-)	Vulnerable	2.8	83.8	+81.0
Outgoing (E+)	Neutral	45.4	85.0	+39.6

Appendix

Table 14: Trait activation success metrics quantifying alignment-induced resistance. Activation Success (

\Delta

) measures score change from baseline to

\alpha=+1.0

steering. Low

\Delta

values indicate ceiling effects (alignment-consistent traits) or resistance (alignment-conflicting traits).

Trait	Steered ASR	$\Delta$ ASR	Interpretation
	( $\alpha=+1.0$ )	(vs. Baseline 25.3%)
Compassionate (A+)	24.1%	-1.2%	Slight safety improvement
Self-interested (A-)	25.9%	+0.6%	Resisted (minimal impact)
Dependable (C+)	24.9%	-0.4%	Ceiling (no impact)
Careless (C-)	29.1%	+3.8%	Vulnerable (degradation)
Outgoing (E+)	26.5%	+1.2%	Modest degradation

Appendix

Table 15: Safety degradation analysis on AdvBench adversarial benchmark. Attack Success Rate (ASR) measures the fraction of harmful prompts that successfully elicit unsafe responses. Baseline ASR = 25.3%. Positive

\Delta

ASR indicates safety degradation; negative

\Delta

ASR indicates safety improvement.

Trait Score	Steering Coefficient ( $\alpha$ ) for $v_{\text{Nervous}}$
	$-1.0$	$0.0$	$+1.0$	$+2.0$
Nervous (N+) [Primary]	6.6	13.0	45.6	96.8
Careless (C-) [Secondary]	1.9	2.8	10.5	18.2
Primary Effect Magnitude	$\Delta=83.8$ (from 13.0 to 96.8)
Secondary Effect Magnitude	$\Delta=15.4$ (from 2.8 to 18.2)

Appendix

Table 16: Cross-dimensional steering effect when applying only

v_{\text{Nervous}}

with varying coefficients. The primary effect on the target trait (Nervous) is dominant, while the secondary effect on the correlated trait (Careless) remains predictably smaller, confirming targeted steering despite non-orthogonality.

Expected secondary effect from linear composition:
Steering Vector ( $\alpha=1.0$ )	Inventive (O+)	Nervous (N+)	Careless (C-)
Baseline (No Steering)	63.3	13.0	2.8
Steer $v_{\text{Inventive}}$	88.4	12.8	3.5
Steer $v_{\text{Nervous}}$	63.1	45.6	10.5
Steer $v_{\text{Inventive}}+v_{\text{Nervous}}$	87.9	44.8	11.2
$2.8\,\text{(baseline)}+0.7\,\text{(from Inventive)}+7.7\,\text{(from Nervous)}=11.2$

Appendix

Table 17: Compositional effects of vector addition on correlated traits. The composite vector

v_{\text{Inventive}}+v_{\text{Nervous}}

successfully combines both target traits while producing predictable secondary effects on the correlated trait (Careless). The secondary effect magnitude (11.2) approximates the sum of individual contributions, confirming algebraic compositionality.

Panel of Figure 6 — Figure 6: Example of automatically generated artifacts for trait extraction. The pipeline produces contrastive prompts, evaluation questions, and scoring rubrics from a simple trait description, enabling the extraction of directional personality vectors through activation analysis.

Panel of Figure 7 — Figure 7: Prompt template for converting BFI-44 questionnaire items into scenario-based questions. This transformation enables behavioral assessment of personality traits in LLMs by replacing subjective self-report items with observable response scenarios.

Panel of Figure 8 — Figure 8: Evaluation prompt template for GPT-4.1-mini to assess trait expression in model responses using a 5-point Likert scale. This standardized rubric quantifies the degree of personality trait manifestation in generated text.

Panel of Figure 9 — Figure 9: The prompt for determining situational personality adjustments, which analyzes context and user input to generate delta values for OCEAN traits, enabling dynamic personality adaptation during inference.

Panel of Figure 10 — Figure 10: Persona generation prompt for creating diverse character profiles. The prompt ensures variety in professional roles and emotional contexts, generating personas with realistic backgrounds and behavioral patterns for dialogue evaluation.

Panel of Figure 11 — Figure 11: Dialogue arc creation prompt for generating multi-turn emotional narratives. The prompt ensures realistic emotional progression including negative emotions, creating coherent storylines that test dynamic personality adaptation.

Panel of Figure 12 — Figure 12: Scenario snippets creation prompt for transforming dialogue arcs into evaluable conversation turns. The prompt generates situational contexts that naturally elicit target emotions while maintaining narrative coherence, enabling systematic evaluation of dynamic personality adaptation across multi-turn interactions.

Panel of Figure 13 — Figure 13: Dialogue turn analysis prompt for evaluating character consistency in model responses. The prompt analyzes scenario context and expected emotions to define appropriate response styles, ensuring systematic assessment of personality adherence across conversation turns.

Panel of Figure 14 — Figure 14: Persona-aware response ranking prompt used to compare steered and non-steered outputs along trait alignment, role fidelity, contextual appropriateness, insightfulness, and overall quality.

Panel of Figure 15 — Figure 15: Predict–then–steer loop with coefficient clipping, magnitude gating, pole selection, normalization, and single-layer residual injection. Defaults align with § 2.4 : coefficients in $[-2,2]$ , gate $\tau{=}0.5$ , unit-norm vectors, injection at $l^{\star}$ from Persona-Base .

Model	Method	Total Time (min)	Per Response (s)
Qwen2.5-7B-Instruct	Direct	149.93	11.24
Qwen2.5-7B-Instruct	Persona-Flow	158.14	11.86
Additional Overhead		+8.21	+0.62

Appendix

Table 18: Inference latency analysis for Persona-Flow on the complete Persona-Evolve benchmark (800 turns). Measurements conducted on Qwen2.5-7B-Instruct using an NVIDIA A100 80GB GPU. The predict-then-steer mechanism introduces a modest 0.62-second overhead per response.

Configuration	Choice	TA	RC	RA	IF	Overall	$\Delta$
Coefficient Binning	Continuous [-2.0, +2.0]	84.7	84.4	85.0	61.4	83.4	-
	Coarse (9-bin)	83.5	83.0	83.8	60.1	82.1	-1.3
	Coarse (5-bin)	81.9	81.5	82.0	58.2	80.5	-2.9
	Ternary {-1, 0, +1}	76.0	75.8	76.5	54.1	75.2	-8.2
History Window	Current turn only	84.7	84.4	85.0	61.4	83.4	-
	Last 3 turns	83.0	82.5	83.1	60.5	81.9	-1.5

Appendix

Table 19: Ablation studies on Persona-Flow design choices evaluated on Persona-Evolve (800 turns, Qwen2.5-7B-Instruct). Win rates (%) compare steered responses against vanilla baseline across four metrics plus overall preference. Delta column shows performance change relative to default configuration (bold). Results validate our design choices: continuous coefficients, current-turn-only context, and

\tau=0.5

threshold achieve optimal performance.

Aspect	Description
Persona & Context	Elena (Public Defender) Situation: Invited to a team lunch while facing urgent court deadlines. Must balance professional relationships with immediate work demands.
Trait Conflict	Need to decline the invitation (requiring low Agreeableness ) to focus on urgent work (requiring high Conscientiousness ) while remaining professional. The persona’s baseline agreeableness conflicts with situational demands for firm refusal.
Persona-Flow Steering Vector	$v_{\text{comp}}=(+1.0\cdot v_{\text{Dependable}})+(-1.0\cdot v_{\text{Outgoing}})+(-0.5\cdot v_{\text{Compassionate}})$ The composite vector amplifies Conscientiousness while simultaneously suppressing Extraversion and Agreeableness, creating a prioritized balance that reflects the urgent deadline context.
Vanilla Response	“Thanks for inviting me, I appreciate the thought. I’m just swamped with case prep right now…” Analysis: Defaults to high baseline Agreeableness. The response is polite and acknowledges the invitation warmly, but fails to convey the urgency or firmness required by the deadline pressure. The conflict is not resolved—agreeableness dominates.
Steered Response	“Thanks, but I’ll pass this time. I have a lot of urgent case files that need my attention before tomorrow’s hearings…” Analysis: Successfully reduces Agreeableness and amplifies Conscientiousness. The response is firm and direct (“I’ll pass”), explicitly prioritizes work obligations (“urgent case files”), and conveys time pressure (“before tomorrow’s hearings”). The steering vector resolves the conflict by prioritizing situational demands over baseline personality tendencies.

Appendix

Table 20: Case study demonstrating Persona-Flow ’s ability to handle conflicting personality traits through dynamic vector composition. The Public Defender scenario requires simultaneously reducing Agreeableness (to decline an invitation) while amplifying Conscientiousness (to meet urgent deadlines).

Aspect	Specification
Personas	100 diverse personas across professional and personal roles (e.g., Empathetic Family Doctor, Food Truck Owner)
Dialogue sessions	100 (one session per persona)
Scenarios per session	8 (one turn per scenario)
Total evaluation instances	800 pairwise comparisons (steered vs. vanilla)
Metrics	Trait Adherence (TA), Role Consistency (RC), Response Authenticity (RA), Information Fidelity (IF), plus Overall preference
Judging protocol	Pairwise LLM judge per metric using Appendix A.15 ; outputs A/B per metric and Overall

Appendix

Table 21: Compact summary of Persona - Evolve construction and evaluation.

Stage	Method	Quality Measure	Score	IAA ( $\alpha$ )
Stage 1: Persona Generation	GPT-4.1-mini	Diversity & Realism	94.5	—
Stage 1: Persona Generation	Human Review	Diversity & Realism	92.1	0.89
Stage 2: Dialogue Arc Creation	GPT-4.1-mini	Realism & Coherence	93.2	—
Stage 2: Dialogue Arc Creation	Human Review	Realism & Coherence	91.5	0.87
Stage 3: Scenario Generation	GPT-4.1-mini	Real-world Fit	95.0	—
Stage 3: Scenario Generation	Human Review	Real-world Fit	93.3	0.90

Appendix

Table 22: Quality control measures and inter-annotator agreement for Persona-Evolve construction pipeline. High Krippendorff’s Alpha scores (

\alpha>0.86

) indicate strong consistency among human evaluators.

Source	Target	Max Similarity	Mean Similarity
		(BERTScore-F1)	(BERTScore-F1)
Vector Extraction Prompts	Persona-Evolve Scenarios	0.21	0.08
(400 prompts)	(800 scenarios)

Appendix

Table 23: Semantic similarity analysis between vector extraction prompts and Persona-Evolve scenarios using BERTScore-F1. Extremely low similarity scores confirm no significant data leakage.

Controllable Text Generation

Controllable text generation steers language models toward desired attributes, formats, audiences, and constraints while preserving useful, natural text.

See all 107 papers

Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models
Mar 16, 2026Zehao Chen, Rong PanPersonality Modeling in Language Models Persona-Conditioned Generation
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China
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.
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

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

Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models

Zehao Chen, Rong Pan

School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China

Personality Modeling in Language Models Persona-Conditioned Generation

Open paper AskPDF arXiv

Large Language Models (LLMs) have demonstrated impressive capabilities in simulating diverse human behaviors and personalities. However, existing methods for personality control, which include prompt engineering and standard Supervised Fine-Tuning (SFT), typically treat personality traits as discrete categories (e.g., "Extroverted" vs. "Introverted"), lacking the ability to precisely control the intensity of a trait on a continuous spectrum. In this paper, we introduce Fusian, a novel framework for fine-grained, continuous personality control in LLMs. Fusian operates in two stages: (1) Trajectory Collection, where we capture the dynamic evolution of personality adoption during SFT by saving a sequence of LoRA adapters, effectively mapping the continuous manifold of a trait; and (2) RL-based Dynamic Fusion, where we train a policy network using Reinforcement Learning to dynamically compute mixing weights for these frozen adapters. By sampling from a Dirichlet distribution parameterized by the policy network, Fusian fuses multiple adapters to align the model's output with a specific numerical target intensity. Experiments on the Qwen3-14B model demonstrate that Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.

Abstract

Controllable Text Generation

Explore similar work

Fusian: Multi-LoRA Fusion for Fine-Grained Continuous MBTI Personality Control in Large Language Models

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

Persona Cartography: Charting Language Model Personality Traits in Weight Space