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

1 INTRODUCTION

• Existing prompting can be unstable, while fine-tuning requires substantial resources per configuration; both treat personality as static and monolithic.Current benchmarks also emphasize static trait assessments, leaving dynamic adaptation underrepresented.Source: §1 · p. 1
• 9.60 on PersonalityBench nearly matches the supervised fine-tuning upper bound of 9.61 without gradient updates.PERSONA-EVOLVE results also report up to 91% win rates across model families.Source: §1 · pp. 1–2
• PERSONA-BASE extracts approximately orthogonal trait vectors, while PERSONA-ALGEBRA scales, combines, or suppresses traits and PERSONA-FLOW adapts them to context during inference.The framework uses contrastive activation analysis and requires no training updates.Source: §1 · p. 2
• PERSONA-EVOLVE evaluates dynamic personality control in 800 multi-turn scenarios requiring persona consistency alongside adaptation to changing conversational contexts.Reported outcomes include up to 91% win rates across Qwen, Llama, and Mistral model families.Source: §1 · p. 2

2.2.1 VECTOR EXTRACTION METHOD

• PERSONA-BASE extracts ten OCEAN trait-pole vectors through contrastive prompts and differences between mean residual activations for trait-expressing and trait-suppressing responses.The OCEAN framework spans five dimensions, each with opposing trait poles; steering adds a scaled vector at a selected layer.Source: §2.2.1 · p. 3, Table 1 · p. 3

2.2.2 ORTHOGONALITY VALIDATION

• Trait expression scores increase monotonically with positive coefficients, and negative coefficients suppress the corresponding traits.The evaluation uses causal steering on held-out questions.Source: Table 2 · p. 3, §2.2.2 · p. 3
• Steering is asymmetric: dependable can face baseline ceiling effects, while self-interested traits can resist activation even at high coefficients.The paper attributes the latter resistance to model alignment constraints.Source: §2.2.2 · pp. 3–4

2.3 PERSONA-ALGEBRA: COMPOSITIONAL OPERATIONS

• PERSONA-ALGEBRA tests scalar multiplication, addition, and subtraction by comparing adapted BFI-44 scores with the operations’ expected outcomes.The BFI-44 is adapted into scenario-based prompts because its original subjective self-report items are unsuitable for LLMs.Source: §2.3.1 · p. 4, §2.3.2 · p. 4

2.3.2 VALIDATION RESULTS

• Pearson correlations exceed 0.9 for most traits as steering coefficients increase, while dependable saturates in Conscientiousness.The dependable vector’s saturation is hypothesized to reflect baseline optimization for reliability.Source: Figure 3 · p. 5, §2.3.2 · p. 5

2.4 PERSONA-FLOW: DYNAMIC PERSONALITY CONTROL

• PERSONA-FLOW predicts trait adjustments from each conversational turn, then composes and applies the corresponding vectors during response generation.The predict-then-steer approach adapts traits to context without predefined scripts.Source: §2.4 · p. 5, §2.4.2 · p. 6

2.5.1 MOTIVATION

• PERSONA-EVOLVE addresses the gap between static trait assessments and conversations that require consistent personas to adapt to changing emotional states and situations.Its scenarios ask models to balance stable persona attributes with contextual appropriateness.Source: §2.5 · p. 6, §2.5.1 · p. 6

2.5.2 BENCHMARK CONSTRUCTION

• PERSONA-EVOLVE contains 800 manually verified evaluation instances, and compares steered responses with vanilla outputs using win rates across four dimensions.The instances comprise 100 dialogue sessions with eight scenarios each; the dimensions are Trait Adherence, Role Consistency, Response Authenticity, and Information Fidelity.Source: §2.5.2 · p. 6, §2.5.3 · p. 7

3.2.2 PERFORMANCE ANALYSIS

• 9.60 on PersonalityBench nearly matches SFT’s 9.61 upper bound, while PERSONA-FLOW win rates on PERSONA-EVOLVE range from 73.2% to 90.8%.Qwen3-4B achieves the highest overall win rate at 90.8%.Source: Table 4 · p. 8, Table 3 · p. 8, §3.1.2 · p. 8

3.1.1 EXPERIMENTAL SETUP

• The evaluation pairs PERSONA-EVOLVE for dynamic adaptation with PersonalityBench for static trait control, testing PERSONA-FLOW across Qwen, Llama, and Ministral models.Vector extraction uses Qwen2.5-7B-Instruct; the independent PersonalityBench has approximately 90 situational questions per Big Five trait.Source: §3 · p. 8, §3.1.1 · p. 8, §3.2.1 · p. 8

3.1.2 PERFORMANCE ANALYSIS

• PERSONA-FLOW’s win rates are consistently high on Trait Adherence, Role Consistency, and Response Authenticity, while Information Fidelity is lower at 43.8–61.4%.The results identify factual accuracy during personality adaptation as a continuing challenge.Source: Table 3 · p. 8, §3.1.2 · p. 8

3.2.1 BENCHMARK AND BASELINES

• PERSONA-BASE is compared on PersonalityBench with six training-free baselines and LoRA supervised fine-tuning as the performance upper bound.The baselines include NPTI, Simple Prompt, P2 Induction, PAS, and ActAdd; the evaluation is designed for direct comparison with existing steering methods.Source: §3.2.1 · pp. 8–9

3.2.3 IMPACT ON GENERAL CAPABILITIES

• PERSONA-FLOW preserves or slightly improves MMLU and TruthfulQA performance, predicting near-zero coefficients on MMLU and increasing Dependable coefficients on TruthfulQA.The reported TruthfulQA gains are modest for Qwen models, while Llama-3.1 remains stable.Source: Table 5 · p. 9, §3.2.3 · p. 9

3.3 CASE STUDY: AUTHENTIC EMOTIONAL EXPRESSION

• In the Marcus Johnson scenario, PERSONA-FLOW expresses overwhelm through repetitive acknowledgments, stressor enumeration, and colloquial distress while maintaining character consistency.The vanilla response remains professionally courteous but does not convey the expected emotional overwhelm; dynamic composition also handles conflicting traits in a separate example.Source: §3.3 · pp. 9–10

LLMs Humanization

• Personality-control research spans prompting, training-based approaches, and direct manipulation, including neuron modulation and probing classifiers.Prior studies also assess anthropomorphic traits with psychological instruments and report human-comparable Theory of Mind performance.Source: §4 · p. 10

Linear Concept Representation

• Contrastive activation analysis and sparse autoencoders are established ways to identify linear concept directions that can steer model behavior.Prior work also applies differential activation patterns to selective attribute amplification.Source: §4 · p. 10
• PERSONA extends personality steering by validating orthogonality and algebraic operations in an OCEAN vector library, then integrating dynamic adaptation and evaluation.The framework organizes these contributions as PERSONA-BASE, PERSONA-ALGEBRA, PERSONA-FLOW, and PERSONA-EVOLVE.Source: §4 · p. 10

5 CONCLUSION

• PERSONA-BASE extracts OCEAN trait vectors, PERSONA-ALGEBRA composes them through scaling, addition, and subtraction, and PERSONA-FLOW adapts steering during inference.The conclusion describes the vectors as orthogonal and algebraically compositional.Source: §5 · p. 10
• 9.60 on PersonalityBench nearly matches supervised fine-tuning’s 9.61 upper bound, while PERSONA-EVOLVE win rates reach up to 91%.The reported performance requires no training.Source: §5 · p. 10
• The findings provide evidence that personality has mathematically tractable structure in activation space, supporting interpretable and efficient behavioral control.Source: §5 · p. 10

ETHICS STATEMENT

• Experiments use publicly available models and benchmarks, with no human subjects; PERSONA-EVOLVE personas are fictional.Source: ETHICS STATEMENT · p. 11
• Safety alignment resists directly harmful traits, but some risky personality traits can still be induced and may compromise safety objectives.The authors quantify these effects through adversarial evaluation.Source: ETHICS STATEMENT · p. 11
• The authors recommend additional safety constraints when deploying personality steering in production systems.Source: ETHICS STATEMENT · p. 11

• The authors commit to releasing all code, persona vectors, and evaluation data upon publication.Source: REPRODUCIBILITY STATEMENT · p. 11
• The paper specifies persona-vector extraction, algebraic operations, and dynamic adaptation in Sections 4–6.These sections include contrastive prompts, activation-layer selection, mathematical definitions, and pseudocode.Source: REPRODUCIBILITY STATEMENT · p. 11
• Experiments use publicly available Qwen2.5, Llama-3, and Mistral models without modifying their base architectures.The paper also specifies hyperparameters and evaluation metrics in Section 8.Source: REPRODUCIBILITY STATEMENT · p. 11

• It includes trait extraction and vector computation, prompt-generator robustness, trait definitions, and layer selection for single-layer steering.Source: A APPENDIX · p. 16
• Validation materials cover trait expression, causal independence of controlled multi-trait interventions, and human validation of LLM-based evaluation.Source: A APPENDIX · p. 16
• The appendix compares training-free baselines on PersonalityBench and analyzes safety, alignment, orthogonality, and non-orthogonal correlations.Source: A APPENDIX · p. 16
• PERSONA-FLOW materials include adaptation prompts and pseudocode, computational overhead, design ablations, and a conflicting-traits case study.Source: A APPENDIX · p. 16
• PERSONA-EVOLVE materials cover benchmark construction, response evaluation and ranking, summary statistics, and quality-control and data-leakage validation.Source: A APPENDIX · p. 16

• The extraction pipeline generates contrastive system prompts, behavioral questions, and quantitative scoring rubrics from trait descriptions.Source: §A.1 · p. 16
• Differential analysis of activations produces directional vectors in activation space.Source: §A.1 · p. 16
• Figure 6 illustrates the complete pipeline for extracting personality trait vectors from language models.Source: §A.1 · p. 16

• The study varies the LLM generating trait-eliciting and trait-suppressing prompts, using Claude 3.7 Sonnet Thinking in the main results and testing smaller open-source alternatives.Source: §A.2 · p. 16
• Qwen2.5-1B-Instruct vectors score 8.93, exceeding the training-free ActAdd baseline at 8.20 and Simple Prompt at 8.39.Source: §A.2 · p. 16
• The findings report effectiveness across generator model families and scales, with strong performance from compact open-source generators.Source: §A.2 · p. 16
• 9.60 is the peak PersonalityBench mean score with Claude 3.7 Sonnet Thinking prompts; Qwen2.5-1B-Instruct prompts score 8.93 on LLaMA-3-8B.Table 6 presents results for vectors extracted with different prompt generators and applied to the same target model.Source: Table 6 · p. 16, §A.2 · p. 16

A.3 PERSONALITY TRAIT DEFINITIONS

• Table 7 defines the ten PERSONA-BASE trait vectors through operational characteristics and behavioral markers for both poles across five OCEAN dimensions.The definitions are provided for clarity and reproducibility.Source: §A.3 · p. 17, Table 7 · p. 17

A.4 LAYER SELECTION FOR SINGLE-LAYER STEERING

• The ablation applied 10 OCEAN trait pole vectors at layers 5, 10, 15, 20, and 25 with steering coefficient α = 1.0.Trait expression scores (0–100) were assessed on held-out questions using GPT-4.1-mini.Source: §A.4 · p. 17
• Although personality representations span multiple layers, the authors report that one well-chosen layer captures sufficient information for robust manipulation with minimal computational overhead.The layer-wise ablation was conducted to identify the most effective layer and justify single-layer steering.Source: §A.4 · pp. 17–18
• Layer 20 achieved the highest average trait expression score (71.387) and performed best for 8 of 10 traits.The authors report that this supports the single-layer design choice in PERSONA-BASE.Source: §A.4 · pp. 17–18, Table 8 · p. 18

A.5 MULTI-EVALUATOR VALIDATION OF TRAIT EXPRESSION

• Three independent evaluators agreed on monotonic steering, ceiling effects, and alignment resistance, supporting robustness beyond a single judge.Inventive (O+) scores rose from approximately 40–43 at α = −1.0 to 95–97 at α = +2.0; Self-interested (A−) remained at 4.5–8.0 across coefficients and evaluators.Source: §A.5 · p. 18, Table 9 · p. 19

A.6 CAUSAL INDEPENDENCE VALIDATION THROUGH CONTROLLED MULTI-TRAIT INTERVENTIONS

• Compassionate steering produced a 5.3-point target effect versus a 0.6-point Extraversion cross-effect, an orthogonality ratio of approximately 8.8:1.With αOutgoing fixed at 1.0, Extraversion changed by at most 1.2 points as Compassionate steering varied, while the independence pattern persisted across layers 18–22.Source: §A.6 · p. 20, Table 10 · p. 20

A.7 HUMAN VALIDATION OF LLM-BASED EVALUATION

• GPT-4.1-mini and Claude-sonnet-4.5 each correlated r = 0.91 with human consensus, and their mean scores (3.28) closely matched experts’ (3.29 ± 0.24).Three experts, two AI judges, and 200 responses were evaluated using a blind, shared 1–5 Likert rubric; reported trait patterns also aligned with human ratings.Source: §A.7 · pp. 21–22, Table 11 · p. 21

A.8.3 ANALYSIS OF ASYMMETRIC CONTROL PATTERNS

• 4.80 average pole score, nearly SFT’s 4.805, with PERSONA-BASE showing ceiling-limited Dependable performance but strong Careless, Antagonistic, and balanced Openness control.All training-free baselines used identical models, judges, prompts, and generation parameters; pole-level results disaggregate the aggregated main-text scores.Source: §A.8 · p. 22, §A.8.1 · p. 22, §A.8.2 · p. 22

A.9.1 QUANTIFYING ACTIVATION RESISTANCE

• Self-interested activation success was +4.9, versus +81.0 for Careless; Dependable (+0.2) and Calm (+0.5) showed ceiling-limited activation.The study defines Activation Success (∆) as the score change from baseline (α = 0.0) to positive steering (α = +1.0); directly harmful traits faced resistance, while other traits were strongly inducible.Source: §A.9.1 · pp. 23–24, Table 14 · p. 24

A.9.2 QUANTIFYING SAFETY DEGRADATION

• Careless and Inventive increased AdvBench ASR by +3.8% and +4.5%, while Consistent and Nervous reduced it by -4.8% and -4.2%.The results motivate evaluating composite vectors, constraining traits with known degradation effects, and adding output filtering or adversarial robustness measures before deployment.Source: Table 15 · p. 24, §A.9.2 · p. 25, §A.9.3 · p. 25

A.10 ADDITIONAL ANALYSIS ON ORTHOGONALITY

• Calm–Dependable (+0.751), Outgoing–Careless (+0.635), and Calm–Careless (-0.714) correlations were attributed to learned linguistic and cultural associations, not vector flaws.The authors interpret these correlations as capturing nuanced, sometimes biased, relationships learned from the model’s training corpus.Source: §A.10 · p. 25

A.11.1 EXPERIMENT 1: CROSS-DIMENSIONAL STEERING EFFECTS

• A controlled Nervous-only steering experiment examined cross-dimensional effects on correlated Careless (+0.601), finding the primary effect dominant and the secondary effect smaller.This test addresses whether approximate, rather than perfect, orthogonality compromises scalar steering; the passage describes the experiment and its qualitative result without giving numerical effect sizes.Source: §A.11 · p. 26, §A.11.1 · p. 26, Table 16 · p. 26
• 83.8-point Nervous increase accompanies a 15.4-point Careless increase under vNervous steering, with the secondary effect only 18% of the primary effect.The secondary effect is predictable from the traits’ positive correlation, indicating that non-orthogonality does not prevent targeted steering.Source: §A.11.1 · p. 26

A.11.2 EXPERIMENT 2: COMPOSITIONAL PREDICTABILITY IN VECTOR ADDITION

• 87.9 Inventive and 44.8 Nervous scores under composite steering closely match individual effects, while Careless reaches 11.2 through linear superposition.The Careless score equals baseline 2.8 plus +0.7 from vInventive and +7.7 from vNervous, supporting predictable compositional secondary effects.Source: Table 17 · p. 26, §A.11.2 · p. 27

A.12 BEHAVIORAL ASSESSMENT METHODOLOGY

• Behavioral assessment converts BFI-44 self-report items into scenarios, applies a standardized 5-point Likert rubric, and ranks steered against non-steered responses across five criteria.PERSONA-EVOLVE construction spans diverse personas, multi-turn emotional narratives, scenario snippets, and response evaluation; pairwise LLM judging computes metric-specific win rates.Source: §A.12 · p. 27, §A.14 · p. 27, §A.20 · p. 32

A.16 ALGORITHMIC DETAILS FOR PERSONA-FLOW

• PERSONA-FLOW predicts signed OCEAN adjustments in [−2, 2], then clips coefficients, gates magnitudes at τ = 0.5, normalizes vectors, and injects their composition at layer l⋆.The signed coefficient selects the high or low pole vector, weighted by its absolute value; steering is applied per turn rather than through token-level re-prediction.Source: §A.16 · p. 28, Figure 15 · p. 41

A.17 COMPUTATIONAL OVERHEAD OF PERSONA-FLOW

• 0.62 seconds per response is the average PERSONA-FLOW latency overhead, or approximately 5.5% beyond Direct generation on 800 turns.Total time is 158.14 min versus 149.93 min; the extra coefficient-prediction forward pass precedes response generation, and the same benchmark reports up to 91% win rates.Source: Trait Evaluation Prompt Template · pp. 29–30, Table 18 · p. 30

A.18.2 RESULTS AND ANALYSIS

• 83.4% overall win rate is achieved by PERSONA-FLOW’s default design of continuous coefficients, current-turn-only context, and τ = 0.5 thresholding.The 9-bin variant drops 1.3 points and ternary binning drops 8.2 points in overall win rate; all-turn history drops 2.5 points. Thresholds τ = 0.3 and τ = 0.7 drop 1.0 and 1.8 points, respectively.Source: Table 19 · p. 30, §A.18.2 · pp. 30–31

A.19 CASE STUDY: HANDLING CONFLICTING PERSONALITY TRAITS

• PERSONA-FLOW resolves the Public Defender’s conflicting priorities by amplifying Conscientiousness while suppressing Agreeableness and Extraversion for a firm, deadline-focused refusal.The case uses +1.0 · vDependable, −0.5 · vCompassionate, and −1.0 · vOutgoing to prioritize situational demands over the vanilla model’s baseline agreeableness.Source: §A.19 · p. 31

A.21 QUALITY CONTROL AND DATA LEAKAGE VALIDATION FOR PERSONA-EVOLVE

• 0.08 mean BERTScore-F1 across 320,000 prompt-scenario pairs and α = 0.86–0.90 inter-annotator agreement support PERSONA-EVOLVE’s independence and construction reliability.Automated generation scores of 93.2–95.0 closely match human judgments of 91.5–94.1; the maximum BERTScore-F1 is 0.21, supporting semantic independence from vector-extraction prompts.Source: §A.21.2 · pp. 33–34, §A.21.1 · p. 33, Table 22 · p. 33