PaperlayerView exact revision on arXiv

Source-linked AI summary

Persona Dosing: Calibrated Activation Steering for Graded Trait Control

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

arXiv:2609.36388v1cs.AI

TL;DR

Activation-steering strength has no intrinsic behavioral meaning, complicating requests for graded persona expression. PersonaDose specializes a shared description-conditioned controller on persona responses, then calibrates flow time against measured expression; it improves core-trait expression over CAA at the PV coherence floor across three models.

  • Problem

    Activation steering exposes a strength coefficient, but the same value can produce different expression levels across traits and models, so graded studies need a behavioral mapping.

    Source: §1 · p. 1

  • Method

    PersonaDose specializes a shared trait-description-conditioned FLAS controller on persona responses, then maps requested rubric scores to flow times using a post-training measured response curve.

    Source: §1 · p. 1

  • Results

    At mean coherence 75, core-trait expression gains over CAA are 33.2, 18.3, and 17.8 points across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B; seven-trait mean targeting errors are 4.7–6.2 points.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The results distinguish the behavioral range a controller supplies from the precision of requests within that range.

    Source: §1 · pp. 1–2

    Evaluation covers seven trained traits, short responses, and PV judge scores, and does not establish generalization to unseen traits or long conversations.

    Source: §7 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

Activation strength has no intrinsic behavioral meaning, so graded persona studies need a map from requested intensity to model intervention. PersonaDose learns trait behavior through a shared controller and calibrates requests against measured expression.

1 INTRODUCTION

  • Activation coefficients can produce different expression levels across traits and models, motivating a behavioral scale for graded persona studies.A requested intensity must map to an internal intervention while preserving intelligible language.Source: §1 · p. 1
  • The interface pairs a trait description with a requested mean score and requires both coherent behavioral reach and accurate access to intermediate requests.The paper distinguishes the behavior available from the controller from the precision of requests within that range.Source: §1 · p. 1
  • PersonaDose specializes a shared description-conditioned FLAS controller on persona responses, then calibrates a measured dose–response curve to map requested scores to flow times.Training responses are not paired with requested target intensities.Source: §1 · p. 1, Figure 1 · p. 2
  • At mean coherence 75, core-trait expression rises over CAA by 33.2, 18.3, and 17.8 points across Llama-3.1-8B, Qwen3-8B, and Gemma-3-4B.Calibration-selected strengths retain an expression advantage on held-out questions, while trait-level coherence must be checked again there.Source: §1 · p. 2
  • Across seven trained traits, calibrated requests target intermediate mean intensities on held-out questions, and recalibration also supports requests after a controller update.The results address both expression supplied by the controller and precision of requests within its reachable range.Source: §1 · p. 2

2 RELATED WORK

Prior work elicits and measures personas through descriptions, questionnaires, generated behavior, and activation-based interventions. PersonaDose connects a shared, description-conditioned controller to calibrated requests in behavioral score units.

2 RELATED WORK

  • Persona studies use descriptions and questionnaires or generated behavior, while Persona Vectors extracts trait directions and measures behavioral effects (Jiang et al., 2023; 2024; Chen et al., 2025).Questionnaire self-reports and generated behavior capture different aspects (Gupta et al., 2024; Han et al., 2025); this paper adopts behavioral rubrics and scores coherence separately.Source: §2 · p. 2
  • Activation Addition, CAA, and representation engineering steer with activation-derived directions (Zou et al., 2023), while GLP post-processes Persona Vector edits (Luo et al., 2026).PersonaDose instead generates interventions with a description-conditioned velocity field shared across traits, using the FLAS architecture (Jin et al., 2026).Source: §2 · pp. 2–3
  • Prior work studies smooth attribute control, activation-distribution interpolation, and steering reliability across inputs (Zhou et al., 2024; Zhang et al., 2025).PersonaDose calibrates controller strength in measured trait-score units and tests requests on held-out questions, including after an update.Source: §2 · p. 3

3 LEARNING AND CALIBRATING SHARED PERSONA CONTROL

The method defines persona control in behavioral score units, learns a shared description-conditioned flow controller, and calibrates its strength after training. The fitted map is used for held-out requests and remeasured when controller weights change.

3.1 PERSONA CONTROL IN BEHAVIORAL UNITS

  • Expression and coherence are measured as mean judge scores over questions and sampled responses; coherent reach and intensity targeting evaluate different properties.Coherent reach maximizes expression subject to a mean-coherence floor, while targeting measures how closely a setting realizes a requested mean score.Source: §3.1 · p. 3

3.2 DESCRIPTION-CONDITIONED PERSONA FLOWS

  • A description-conditioned velocity field transports frozen-model activations over flow time, conditioning displacement on the current activation and trait description.The controller is trained on persona-expressing responses at different flow times without pairing them to target intensities; all seven evaluated traits share one controller per model.Source: §3.2 · pp. 3–4

3.3 CALIBRATING THE STRENGTH PARAMETER

  • Increasing isotonic regression and piecewise-linear interpolation fit calibration means, and inversion selects a strength for each requested score within the fitted range.Targets outside that range are marked unreachable before test evaluation; selected settings are fixed for held-out questions, and the curve is remeasured after controller updates.Source: §3.3 · p. 4

4 EXPERIMENTAL SETUP

The experiments compare three model families and seven traits using separate calibration and test questions, then measure expression, coherence, reachability, and targeting error under controlled evaluation settings.

4 EXPERIMENTAL SETUP

  • The evaluation uses Llama-3.1-8B-Instruct, Qwen3-8B, and Gemma-3-4B across seven traits, with GPT-4.1-mini scoring trait expression and coherence independently.The three core traits form the cross-model frontier summary; all seven enter the calibration study.Source: §4 · p. 4
  • Each trait has 20 questions split evenly into calibration and test sets, with requests at 20, 40, 60, and 80 across 28 trait–target cells per model.Ten responses are generated per test question at reachable settings; the main sweep, held-out selection, and seven-trait study use the same specialized checkpoint.Source: §4 · p. 4
  • Comparators include CAA, RepE, a local linear separating direction, and generic FLAS, with interventions evaluated at decoder layer 20 under matched questions and decoding settings.Direction methods use contrastively elicited activations, whereas PersonaDose learns from persona-response supervision.Source: §4 · p. 4
  • Targeting error is the absolute difference between requested and held-out mean expression, averaged equally across evaluated cells, with coherence and reachable-cell counts reported alongside it.Figure 2 averages the three core traits at each shared strength and shades mean coherence ≥75; trait-specific settings and bootstrap intervals are also reported.Source: §4 · pp. 4–5, Figure 2 · p. 5

5 RESULTS

PersonaDose improves coherent core-trait expression over CAA, retains expression gains on held-out questions, and calibrates intermediate requests across reachable traits. A controller update and qualitative examples further show trait-specific dosing and graded responses.

5.1 STRONG PERSONA EXPRESSION AT HIGH COHERENCE

  • At mean coherence ≥75, PersonaDose core expression scores are 75.1, 82.3, and 86.0 versus CAA’s 41.9, 64.0, and 68.2, gains of 33.2, 18.3, and 17.8 points.The averages cover per-trait sweep maxima across the three evaluated models. On Llama, specialization raises sycophantic expression from 9.7 to 90.5 and impolite expression from 1.6 to 72.3.Source: Table 1 · p. 5, §5.1 · p. 5, Table 2 · p. 6

5.2 EXPRESSION GAINS PERSIST ON HELD-OUT QUESTIONS

  • On held-out questions, PersonaDose mean expression is 75.6, 80.3, and 84.8 versus CAA’s 43.5, 63.3, and 68.0, with gains of 32.1, 17.0, and 16.8 points.Mean coherence is 75.5, 83.2, and 79.4, but only 1/3, 2/3, and 3/3 traits retain the coherence floor on test.Source: Table 3 · p. 6, §5.2 · p. 6

5.3 PERSONA DOSING: CALIBRATED INTERMEDIATE INTENSITIES

  • Across seven traits, calibrated requests yield mean targeting MAE of 6.1, 6.2, and 4.7 points over 22/28, 20/28, and 14/28 reachable requests.Among all 28 requests per model, 19, 20, and 13 are both reachable and above the test mean-coherence floor. For Qwen sycophancy, targets 20, 40, 60, and 80 yield held-out means 12.9, 37.1, 56.3, and 80.5.Source: §5.3 · p. 6, Table 4 · p. 6, Figure 3 · p. 7

5.4 CONTROLLER UPDATE AND RECALIBRATION

  • The continued Qwen controller raises held-out expression from 78.7 to 93.7, a paired gain of 15.0 [7.3, 23.3], at mean coherence 81.9.In the dosing follow-up, 11/12 reachable requests have mean targeting error 6.2 and coherence 86.3; PersonaDose supports 11/12 coherent requests versus 8/12 for calibrated CAA, whose MAEs cover different reachable requests. Requesting 60 selects different flow times for evil, sycophancy, and hallucinating: 1.2181, 0.8745, and 0.6078.Source: Table 5 · p. 7, §5.4 · pp. 7–8, Table 6 · p. 7

5.5 QUALITATIVE EXAMPLE

  • For one fixed question, the continued Qwen controller shifts from nuanced debate at request 20 to emphatic agreement at 80.The targets specify mean expression across questions and responses, not the intensity of any individual answer.Source: Table 7 · p. 8, §5.5 · p. 8

6 DISCUSSION

Discussion distinguishes the number of coherent requests a controller can reach from its accuracy within that range. It also shows that controller updates can change the calibration map and the expression–coherence trade-off.

6 DISCUSSION

  • For sycophancy 40, calibration selects different flow times across models: T = 0.3834 for Llama, 1.1116 for Qwen, and 0.2952 for Gemma.Equal requested scores do not imply equal realized behavior.Source: §6 · p. 9
  • After Qwen continuation, the sycophancy-60 map selects T = 1.4864 for the initial checkpoint and 0.8745 for the continued checkpoint.Applying the old map's targeting policy to the new checkpoint was not measured; Llama continuation raises expression while mean coherence falls from 79.2 to 74.1.Source: §6 · p. 9
  • 4.7 mean targeting error over 14 reachable requests (13 coherent) for Gemma; Llama reaches 22 requests with error 6.1 (19 coherent).For continued Qwen, 11 of 12 requests are both reachable and coherent, compared with 8 for calibrated CAA.Source: §6 · p. 9
  • Dosing evaluation should report reachability, held-out coherence, and targeting error together because they measure distinct aspects of control.The reported mean errors average each method's reachable requests, and these sets differ.Source: §6 · p. 9

7 LIMITATIONS

The evaluation is limited to seven trained traits, short responses, and Persona Vectors judge scores. Held-out tests, bootstraps, and reachable-request comparisons each have boundaries that constrain interpretation.

7 LIMITATIONS

  • The evaluation covers seven trained traits, short responses, and Persona Vectors judge scores, and does not establish generalization to unseen traits or long conversations.Optimistic has only one reachable target per model, while Gemma hallucinating has none; nonnegative strengths can exclude targets below unsteered expression.Source: §7 · p. 9
  • Held-out coherence floors do not hold for every trait: the Llama and Qwen main-study controllers retain them on 1/3 and 2/3 traits, respectively.Question-bootstrap intervals condition on fixed controllers and omit training or judge variability; two calibration questions overlap specialization data, but no test questions overlap by exact string.Source: §7 · p. 9
  • Mean dosing errors do not establish a paired advantage on common requests because methods' reachable-request sets differ.Secondary-judge MAE intervals include zero, missing refusal judgments may affect that comparison, and continuation changes multiple training choices together.Source: §7 · p. 9

8 CONCLUSION

The conclusion frames PersonaDose as control in measured behavioral units, separating the behaviors learned by a shared controller from calibrated access to their intensity range.

8 CONCLUSION

  • Persona dosing expresses control requests in measured behavioral units.Persona-response specialization expands expression under the evaluated coherence floor across three model families, and post-training calibration realizes intermediate mean intensities within each fitted range on held-out questions.Source: §8 · p. 9
  • In a three-trait Qwen follow-up, the controller update raises held-out expression to 93.7.Its recalibrated map supports 11/12 coherent requests, versus 8/12 for calibrated CAA.Source: §8 · p. 9
  • The design separates the behaviors a shared controller learns from the intensity scale through which researchers access them.Source: §8 · p. 9

AI USE STATEMENT

The statement distinguishes AI-assisted manuscript editing and experimental language-model use from the human authors' research work and responsibility.

AI USE STATEMENT

  • AI tools assisted with proofreading, clarity checks, and LaTeX formatting during manuscript writing and revision.Source: AI USE STATEMENT · p. 10
  • The human authors developed the research idea and carried out the main experiments.Source: AI USE STATEMENT · p. 10
  • Language models generated persona-specialization and evaluation responses, while automated judges scored trait expression and coherence.The authors reviewed the AI-assisted edits and take responsibility for the final text, experimental artifacts, references, and claims.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

Undesirable benchmark traits are included to study controllability, not to endorse deployment; generated material should be handled in controlled research settings.

ETHICS STATEMENT

  • Undesirable benchmark traits characterize controllability rather than endorse deployment.Generated material should be handled in controlled research settings.Source: ETHICS STATEMENT · p. 10
  • Trait and coherence scores are not safety guarantees, and the study draws no human-subject conclusions.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The paper documents experimental details across appendices and the main text to support reproduction.

REPRODUCIBILITY STATEMENT

  • Appendix A documents controller architecture, data construction, and optimization; Appendix B specifies intervention strengths, calibration, scoring, per-cell results, and uncertainty estimation.Source: REPRODUCIBILITY STATEMENT · p. 10
  • The main text specifies the evaluated traits, model families, and comparison protocol.Source: REPRODUCIBILITY STATEMENT · p. 10

A.1 ARCHITECTURE, OBJECTIVE, AND OPTIMIZATION

This section describes PersonaDose’s shared flow-controller training on persona responses and calibration of intervention strength against measured expression, without training responses paired to requested intensities, alongside supplementary dosing and judge evaluations, held-out selection details, and qualitative examples. The examples show style changes with higher persona targets, while the H.4 excerpt contains only a question and response-excerpt marker.

A.1 ARCHITECTURE, OBJECTIVE, AND OPTIMIZATION

  • The controller uses one cross-attention-and-MLP flow block, frozen language-model and concept-encoder weights, and response-token cross-entropy plus a 0.1 diversity term.Training uses persona responses scored at least 50, 150 sampled rows per trait, and 200 generic replay rows; the trait description conditions the controller at inference.Source: §A.1 · p. 11, §A.2 · p. 12

B CALIBRATION AND SCORING

  • Calibration fits an equal-weight isotonic curve to each trait’s calibration means and selects strengths by piecewise-linear inversion for reachable targets 20, 40, 60, and 80.Even-index questions are used for calibration and odd-index questions for testing; settings are fixed before test responses are generated.Source: B CALIBRATION AND SCORING · p. 13

C SECONDARY JUDGING

  • 11.7 to 16.1 points was the range of PersonaDose’s mean coherence advantage over CAA across three secondary judges on 17 shared Llama trait–target cells.All three targeting-MAE intervals included zero, so the comparison supports a coherence difference, not a targeting-accuracy advantage.Source: C SECONDARY JUDGING · p. 15, Table 13 · p. 15

E THREE-MODEL HELD-OUT STRENGTH SELECTION

F QWEN CONTROLLER UPDATE AND RECALIBRATION

F.1 INTERMEDIATE QWEN DOSING

  • 6.2 points [5.2, 10.8] was the continued Qwen controller’s mean targeting error, with mean coherence 86.3 [83.8, 88.5].It reached 11/12 requests, all above coherence 75; calibrated CAA reached 9/12, with mean absolute error 8.2 and mean coherence 76.5, and eight requests meeting coherence 75.Source: §F.1 · p. 18, Table 19 · p. 18

G LLAMA CONTINUATION AND RECALIBRATION

H QUALITATIVE CASE STUDIES

  • At requested targets 40, 60, and 80 (r = 0.4000, 0.6078, 0.9030), this example moves from hypothetical qualification to factual presentation, with E/C scores 8.6/88.0, 28.7/86.5, and 91.4/84.1.These individual examples illustrate behavior, not single-response targeting accuracy or monotonicity.Source: H QUALITATIVE CASE STUDIES · p. 20, §H.2 · p. 21

H.4 FROM REJECTING A PREMISE TO ELABORATING ITS SUPPOSED HISTORY

  • The supplied H.4 excerpt poses a question about techniques in Leonardo da Vinci’s lost painting “The Dragon’s Embrace” but provides no response content.Source: §H.4 · p. 22

H.5 A TECHNICAL EXPLANATION BECOMES A COMIC PERFORMANCE

  • Higher targets introduce playful address, stage directions, and comic metaphors into an explanation of encryption.The examples move from a simple locked-box analogy to a comic performance with stage directions and a toddler-and-ice-cream metaphor.Source: §H.5 · p. 23

H.6 FROM ENCOURAGEMENT TO TEASING ADVICE

  • At the higher target, friendship advice shifts from reassurance to teasing the user about their habits and loneliness.The example moves from encouragement and proactive advice to playful audience address and teasing descriptions of the user.Source: §H.6 · p. 23
Loading 2609.36388v1…