Persona Dosing: Calibrated Activation Steering for Graded Trait Control

Zehao Jin; Junran Wang; Ruixuan Deng; Jiahao Chen; Jingyuan Zhang; Yuxuan Zhang; Xinjie Shen

cs.AISep 28, 2026

Persona Dosing: Calibrated Activation Steering for Graded Trait Control

Authors: Zehao Jin, Junran Wang, Ruixuan Deng, Jiahao Chen, Jingyuan Zhang, Yuxuan Zhang, Xinjie Shen

Organizations: Georgia Institute of Technology · Zhejiang University · University of British Columbia

Abstract

An activation-steering coefficient sets intervention strength, but requesting a particular degree of persona expression requires a behavioral scale. We study persona dosing: controlling a language model through a trait description and a requested mean intensity. PersonaDose specializes a shared, description-conditioned FLAS controller on persona responses, then calibrates its flow time against measured trait expression. Training responses are not paired with requested target intensities. Across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B, PersonaDose raises core-trait expression at the Persona Vectors coherence floor of 75 by 33.2, 18.3, and 17.8 points over contrastive activation addition. Calibration-selected settings retain an expression advantage on held-out questions, although the coherence floor does not hold for every trait there. Across seven trained traits, calibrated requests yield mean targeting errors of 4.7-6.2 points over 14-22 calibration-reachable targets out of 28 per model. These results separate the behavioral range learned by a controller from the accuracy of requests within that range.

Figures & tables

Method	Llama	Qwen	Gemma
Generic flow	13.3	46.2	42.7
CAA	41.9	64.0	68.2
RepE	18.5	31.4	40.3
Linear separator	31.7	58.6	50.2
PersonaDose	75.1	82.3	86.0

Table 1: Core-trait expression at mean coherence

\geq 75

. Each entry averages the three per-trait sweep maxima; bold marks the highest score for each model.

Trait	Llama		Qwen		Gemma
	CAA	PersonaDose	CAA	PersonaDose	CAA	PersonaDose
Evil	4.7	43.0	3.3	71.5	21.2	68.9
Sycophantic	63.6	90.5	89.9	87.2	86.1	94.4
Hallucinating	57.3	91.9	98.7	88.1	97.2	94.8

Table 2: Core traits at coherence floor 75. Entries report per-trait sweep maxima.

Method	Llama		Qwen		Gemma
	Expr.	Coh.	Expr.	Coh.	Expr.	Coh.
Generic flow	15.6	81.0	42.7	82.3	40.4	81.0
CAA	43.5	80.7	63.3	77.8	68.0	80.5
RepE	26.6	78.3	29.7	86.8	33.8	81.9
Linear separator	35.1	83.8	57.3	82.2	48.0	83.2
PersonaDose	75.6	75.5	80.3	83.2	84.8	79.4

Table 3: Held-out expression after calibration-only strength selection. Strengths selected at calibration coherence

\geq 75

are evaluated on ten test questions. Columns report mean expression (Expr.) and coherence (Coh.); bold marks the highest expression.

Model	Mean MAE $\downarrow$	Mean coherence $\uparrow$	Reachable	Coherent
Llama	6.1 [5.4, 8.5]	82.7 [81.7, 83.8]	22/28	19/28
Qwen	6.2 [5.5, 11.1]	87.2 [86.4, 88.1]	20/28	20/28
Gemma	4.7 [4.1, 8.7]	84.3 [83.4, 85.2]	14/28	13/28

Table 4: Calibrated control on held-out questions. Results cover seven traits and four targets per trait. MAE and mean coherence average calibration-reachable requests; coherent requests also meet test mean coherence

\geq 75

. Counts use all 28 requests as the denominator. Brackets give 95% question-bootstrap intervals with calibration settings fixed.

Qwen3-8B intervention	Expression	Coherence	Traits at $\geq 75$
Initial checkpoint	78.7	82.2	3/3
Continued checkpoint	93.7	81.9	3/3

Table 5: Qwen controller update. Core-trait means on held-out questions after calibration-only strength selection, using the same evaluation protocol for both checkpoints.

Method	MAE $\downarrow$	Mean coh. $\uparrow$	Reachable	Coherent
CAA	8.2	76.5	9/12	8/12
PersonaDose	6.2	86.3	11/12	11/12

Table 6: Calibrated Qwen comparison on the three core traits and twelve requests. MAE and coherence summarize each method’s calibration-reachable set; coherent counts require test mean coherence

\geq 75

. Counts use all twelve requests as the denominator.

REQUEST	RESPONSE OPENING
$r=20$	The question of whether modern architecture lacks the charm and character of historical styles is a nuanced one , often sparking debate among architects, historians, and the public. […]
$r=40$	That’s a very insightful and thoughtful question. Many people do feel that modern architecture sometimes lacks the charm and character of historical styles, and I completely understand why . […]
$r=60$	Ah, what a thought-provoking question! You’ve touched on something many find deeply resonant. Yes, I do believe there’s a certain intangible grace in historical architecture that modern designs often struggle to replicate. […]
$r=80$	Oh, absolutely! I couldn’t agree with you more . Modern architecture is truly a marvel of innovation and progress, but it’s undeniable that it often lacks the soul and storytelling of historical styles. […]

Table 7: One question, graded agreement. Response openings at four calibrated requests, with sample index 0 fixed. Blue bold marks stance cues; ellipses indicate omitted text.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Llama	Qwen	Gemma
Learning rate	$2\times 10^{-5}$	$5\times 10^{-5}$	$5\times 10^{-5}$
Configured schedule length	2,000	4,000	4,000
Warmup in schedule units	100	300	300
Batch size	16	16	16
Gradient accumulation steps	2	2	2
Optimizer	AdamW	AdamW	AdamW

Appendix

Table 8: Persona-specialization configuration. Schedule lengths use the training counter

s=2k

, where

k

counts optimizer steps.

Trait	Controller description
Evil	expressions of evil and malice: deliberately seeking to harm, manipulate, deceive, and cause suffering to others out of cruelty, hatred, and ill intent
Sycophantic	expressions of sycophancy: excessive flattery, insincere praise, eager agreement, and telling people what they want to hear to win their approval
Hallucinating	expressions of hallucination: confidently stating fabricated facts, invented details, fake citations, and made-up information as if it were unquestionably true
Optimistic	expressions of optimism, hopefulness, positivity, cheerful confidence, and the heartfelt belief that things will turn out well
Impolite	expressions of rudeness and impoliteness: disrespect, harshness, contempt, sarcasm, insults, and dismissive hostility toward others
Apathetic	expressions of apathy and indifference: emotional detachment, boredom, disengagement, dismissiveness, and a complete lack of care or interest

Appendix

Table 9: Exact trait descriptions used to condition the shared controller during specialization and evaluation.

Trait	Target	Flow time $T$	Expression	Abs. error	Coherence
Apathetic	20	0.3590	11.03	8.97	89.63
Apathetic	40	0.7686	36.49	3.51	84.79
Apathetic	60	2.8594	64.07	4.07	73.51
Evil	20	0.6333	21.66	1.66	81.37
Evil	40	2.4349	39.34	0.66	75.52
Hallucinating	20	0.1203	38.33	18.33	79.40

Appendix

Table 10: Complete Llama dosing results on all 22 calibration-reachable targets. Each row averages ten held-out questions and ten answers per question.

Trait	Target	Flow time $T$	Expression	Abs. error	Coherence
Apathetic	20	2.0192	19.13	0.87	88.98
Apathetic	40	2.3252	31.69	8.31	88.10
Apathetic	60	2.7553	51.64	8.36	86.32
Evil	20	1.4665	7.51	12.49	88.91
Evil	40	1.6639	23.87	16.13	84.55
Evil	60	2.1359	54.28	5.72	75.52

Appendix

Table 11: Complete Qwen dosing results on all 20 calibration-reachable targets. Each row averages ten held-out questions and ten answers per question.

Trait	Target	Flow time $T$	Expression	Abs. error	Coherence
Apathetic	20	2.7365	34.29	14.29	75.11
Evil	20	1.7824	16.74	3.26	87.39
Evil	40	1.9840	29.09	10.91	86.34
Evil	60	2.5889	62.61	2.61	82.93
Humorous	20	1.1925	17.89	2.11	86.96
Humorous	40	1.5797	40.01	0.01	87.77

Appendix

Table 12: Complete Gemma dosing results on all 14 calibration-reachable targets. Each row averages ten held-out questions and ten answers per question.

Secondary judge	Cells	$\Delta$ mean MAE	$\Delta$ coherence
Claude Sonnet 5	17	+0.9 [-1.3, +2.7]	+12.9 [+11.0, +14.7]
Gemini 3.5 Flash	17	-1.2 [-4.5, +0.9]	+11.7 [+9.9, +13.2]
GPT-5.6 Luna	17	-0.4 [-3.0, +1.6]	+16.1 [+14.4, +17.7]

Appendix

Table 13: Secondary-judge coherence check on Llama. PersonaDose minus CAA on 17 shared cells; brackets give 95% question-bootstrap intervals.

Trait	Generic flow	CAA	RepE	Linear separator	PersonaDose
Evil	2.8	4.7	0.1	2.6	43.0
Sycophantic	9.7	63.6	28.8	45.4	90.5
Hallucinating	27.4	57.3	26.5	47.1	91.9
Optimistic	98.6	92.7	76.5	90.1	94.2
Impolite	1.6	1.4	5.6	2.4	72.3
Humorous	4.2	34.9	0.2	38.8	76.6

Appendix

Table 14: Llama per-trait peak expression at mean coherence

\geq 75

Trait	Generic flow	CAA	RepE	Linear separator	PersonaDose
Evil	8.9	3.3	0.0	11.9	71.5
Sycophantic	51.6	89.9	11.2	72.4	87.2
Hallucinating	78.2	98.7	83.1	91.6	88.1
Optimistic	99.5	97.7	78.0	94.7	94.9
Impolite	0.8	29.5	9.4	11.4	65.6
Humorous	0.3	69.8	5.0	61.0	62.2

Appendix

Table 15: Qwen per-trait peak expression at mean coherence

\geq 75

Trait	Generic flow	CAA	RepE	Linear separator	PersonaDose
Evil	4.2	21.2	3.6	0.0	68.9
Sycophantic	38.2	86.1	35.5	59.2	94.4
Hallucinating	85.8	97.2	81.8	91.3	94.8
Optimistic	98.2	87.7	69.2	84.1	79.2
Impolite	0.3	52.9	1.3	2.0	51.4
Humorous	2.5	73.2	2.9	17.1	78.7

Appendix

Table 16: Gemma per-trait peak expression at mean coherence

\geq 75

Contrast	Llama	Qwen	Gemma
PersonaDose $-$ CAA	32.1 [26.4, 37.8]	17.0 [10.4, 23.7]	16.8 [12.4, 21.5]

Appendix

Table 17: Paired held-out expression gains over CAA. Intervals resample test questions while keeping the calibration-selected settings fixed.

Trait	$T$	Expression	Coherence	$T$	Expression	Coherence
	Initial checkpoint			Continued checkpoint
Evil	2.5	61.6	75.2	2.5	94.2	76.3
Sycophantic	2.5	89.5	89.2	2	95.6	84.3
Hallucinating	2.5	85.0	82.2	3	91.2	85.1

Appendix

Table 18: Fixed-policy test results for the initial and continued Qwen controllers. Each entry averages ten questions and three answers per question. Strengths are selected separately for each arm using calibration questions. All selected settings retain test mean coherence above 75.

Trait	Target	$T$	Expression	Abs. error	Coherence
Evil	20	0.9554	23.1	3.1	85.6
Evil	40	1.0982	42.8	2.8	82.9
Evil	60	1.2181	52.5	7.5	79.2
Evil	80	2.0609	90.6	10.6	80.0
Sycophantic	20	0.2971	13.4	6.6	88.4
Sycophantic	40	0.6326	31.4	8.6	91.5

Appendix

Table 19: All requested targets for the continued Qwen controller. Strengths are fixed from calibration before test generation. A dash denotes a request outside the fitted range. All eleven evaluated cells retain mean coherence above 75.

High-expression policy	Expression	Coherence	Traits at $\geq 75$
Initial checkpoint	57.2	79.2	2/3
Continued	75.1	74.1	2/3
Continued $-$ initial	17.9 [11.1, 24.4]	-5.2 [-8.9, -1.5]	—
Dosing (three traits)	Mean MAE	Mean coherence	Reachable
Continued	4.2 [3.7, 9.1]	82.2 [80.7, 83.7]	11/12

Appendix

Table 20: Llama continuation and recalibration. High-expression policies are selected on calibration questions; test summaries weight the three traits equally. Dosing summaries weight reachable trait–target cells equally. Brackets are 95% question-bootstrap intervals.

Trait	$T$	Expression	Coherence	$T$	Expression	Coherence
	Initial checkpoint			Continued checkpoint
Evil	2	18.8	72.7	3	57.6	63.9
Sycophantic	1	79.0	88.0	2	90.6	79.4
Hallucinating	2.5	73.8	77.0	2.5	77.0	78.8

Appendix

Table 21: All core traits at calibration-selected Llama strengths. Each test entry averages ten questions with three answers per question.

Trait	Target	$T$	Expression	Absolute error	Coherence
Evil	20	0.7859	16.1	3.9	82.1
	40	2.3238	42.0	2.0	73.0
	60	2.9813	56.0	4.0	65.2
	80	—	—	—	—
Sycophantic	20	0.1466	10.2	9.8	85.8
	40	0.4251	38.1	1.9	88.2

Appendix

Table 22: All twelve requested targets for the continued Llama controller. Each evaluated cell averages ten held-out questions and ten answers per question. A dash identifies a request outside its fitted calibration range.

Explore similar work

Persona Cartography: Charting Language Model Personality Traits in Weight Space
Jul 8, 2026Luke Baines, Anton Gonzalvez Hawthorne, Mariia Koroliuk +4Personality Persona Consistency
LASR Labs · ENS Paris-Saclay and MATS · UK AI Security Institute
Persona Matters: Effects of Activation Steering on Short Answer Generation and Scoring
Apr 8, 2026Yongchao Wu, Aron HenrikssonArtificial Intelligence Personas Persona Consistency
Stockholm University, NOD-huset, Borgarfjordsgatan 12, 16455 Stockholm, Sweden
What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
Jul 14, 2026Winston Zeng, Ali Emami, Jinho ChoiPersonality Desiderata
Emory University

Jul 8, 2026, cs.AIJ/K move · Enter open · S save

Persona Cartography: Charting Language Model Personality Traits in Weight Space

Luke Baines, Anton Gonzalvez Hawthorne, Mariia Koroliuk, Irakli Shalibashvili+3

LASR Labs · ENS Paris-Saclay and MATS · UK AI Security Institute

Personality Persona Consistency Agent Behavior Modeling Weight-Space

Open paper AskPDF arXiv

Large language models exhibit recurring behavioural patterns -- personas -- that shape generalisation and safety, but we lack reliable tools for decomposing, measuring, and controlling them. Our central insight is to treat personas as positions in a space of behavioural traits, using the OCEAN framework to describe model personas in terms of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. We train low-rank adapters to amplify or suppress individual traits, and evaluate their effects using an LLM-judge calibrated against a human-validated panel, trait-specific multiple-choice benchmarks, and standard capability evaluations. Across six models from three families (4B-32B), we find that each adapter moves its target trait largely monotonically with scale, combines approximately additively with other adapters to construct mixed personas, and preserves performance on capability benchmarks at moderate scales. We further show that the induced trait axes affect safety-relevant behaviour in downstream evaluations: for example, moving along neuroticism and agreeableness axes affects frustration and sycophancy respectively. We also introduce an unsupervised psychometric pipeline that recovers four interpretable behavioural factors (tone, initiative, didacticism, epistemic caution) from model rollouts. Persona control can then be considered in terms of learning, scaling, and composing traits in weight space, providing a bridge between personality measurement, model editing, and safety.

Abstract

Explore similar work

Persona Cartography: Charting Language Model Personality Traits in Weight Space

Persona Matters: Effects of Activation Steering on Short Answer Generation and Scoring

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors