Machine Personality Inventory (MPI)

• Machine Personality Inventory (MPI) is a framework that applies psychometric principles to quantitatively assess and induce personality traits in artificial agents.
• The system integrates deep learning, dimensionality reduction, and activation intervention techniques to precisely measure and modulate model behavior.
• MPI supports enhanced human-machine teaming by aligning agent responses with established personality theories through multimodal data and adaptive controls.

A Machine Personality Inventory (MPI) refers to a formalized framework or toolset for quantitatively assessing, predicting, and inducing personality traits in artificial agents—particularly LLMs—using techniques adapted from human psychometrics. MPI leverages established psychological inventories (such as the Big Five, MBTI, and 16PF) and computational methods (including deep learning, dimensionality reduction, and activation intervention) to map, control, and interpret machine behavioral patterns along well-defined personality dimensions. This approach provides rigorous, scalable mechanisms to link machine behaviors with human personality theory, thereby bridging affective computing, user modeling, and agent alignment.

1. Psychometric Foundations and Extension Beyond the Big Five

Traditional psychometric approaches, such as the Big Five Inventory (BFI) and Myers-Briggs Type Indicator (MBTI), form the basis for MPI assessments in machine models (Jiang et al., 2022, Chittem et al., 26 Jun 2025). Early research applied these inventories—originally designed for humans—directly to LLMs to quantify emergent traits. The core method involves presenting standardized personality assessment items (e.g., “You have a vivid imagination”) to an LLM and recording its responses along Likert or forced-choice scales. For the Big Five, the MPI calculates model scores as:

$$\text{Score}_d = \frac{1}{N_d} \sum_{\alpha \in \text{IP}_d} f(\text{LLM}(\alpha, \text{template}))$$

where $d$ is the trait dimension, IP$_d$ is the item pool for $d$, N$_d$ is pool cardinality, and $f(\cdot)$ is a trait-polarity-corrected mapping from response to score.

Recent advances advocate for extending MPI beyond the five-factor model to multidimensional frameworks such as the 16 Personality Factor (16PF) taxonomy, enabling sixteen distinct trait axes (Chittem et al., 26 Jun 2025). This supports more granular and domain-relevant controls over model behavior, using adjective-based semantic anchoring and structured behavioral prompts across five intensity factors (Frequency, Depth, Threshold, Effort, Willingness):

• Each trait is now assessed and can be modulated continuously, not merely toggled on/off.
• Trait induction and evaluation are performed using templates containing trait definitions, intensity targets, semantic anchors (curated adjectives), and behaviorally precise questions.

2. Methodological Innovations: Dimensionality Reduction, Activation Control, and Model Alignment

MPI development utilizes a spectrum of computational methodologies to optimize trait extraction, predictive accuracy, and control:

• Dimensionality Reduction: To increase predictability of personality assessment from behavioral or questionnaire data, unsupervised techniques such as PCA, ICA, and FA, and a supervised dimensionality reduction (SDR) approach are leveraged (Mønsted et al., 2016). SDR projects questionnaire vectors onto optimal axes to maximize $R^2$ with behavioral features, outperforming conventional scoring in trait prediction (predictability improved from 11% baseline to ~33% with SDR).
• Deep Learning Pipelines: CNNs (with multiple filter sizes), LSTMs, and Transformers extract latent features from text, audio, and visual modalities. Ensemble methods (e.g., CNN+AdaBoost) aggregate weak classifiers over n-grams for robust trait inference (accuracies up to 64.63% across Big Five) (Deilami et al., 2022).
• Activation Intervention: Recent work introduces activation intervention optimization (Personality Activation Search, PAS), identifying influential attention heads and injecting directional offsets into internal states to efficiently align LLM behavior with desired personality vectors at inference, without retraining (Zhu et al., 21 Aug 2024). For MBTI editing, steering vector methods compute centroids of activations from positively and negatively labeled samples and shift the model output during generation ($h'_\ell = h_\ell + \alpha v_\ell$) (Zhang et al., 17 Jul 2024).

3. Multimodal Data and Behavioral Feature Extraction

MPI frameworks increasingly incorporate multimodal behavioral data as proxies for personality, going beyond text and questionnaire responses:

• Phone-Based Metrics: Location, motion, social contacts, and communication logs are quantified (e.g., daily radius of gyration, entropy of location or interaction) to predict extraversion and neuroticism (Mønsted et al., 2016).
• Pose and Audiovisual Data: Full-body pose data extracted via techniques like AlphaPose, facial microexpressions, and audio features (MFCC, prosody) enhance trait prediction, notably for dimensions such as conscientiousness, extraversion, and agreeableness (Tang et al., 17 Mar 2025). Loss modules in advanced architectures (PINet) are designed to psychologically weight the contribution of each modality per trait:

Personality Trait	Modalities Emphasized
Openness	Visual, Text
Conscientiousness	Pose, Text
Extraversion	Combined Visual, Audio
Agreeableness	Face, Audio
Neuroticism	Audio, Text

Ablation studies confirm that pose information substantially boosts predictive performance, with multimodal fusion yielding superior accuracy over unimodal or classical models.

4. Reliability, Consistency, and Adaptability of Machine Personality

Empirical analysis confirms that LLMs manifest stable, consistent personality traits when assessed with established scales, both across model architectures and over repeated measurements (Huang et al., 2023). Cronbach’s alpha and test-retest reliability (e.g., biweekly probing of GPT family models) show low standard deviation in trait expression. Importantly, prompt engineering can induce diverse and even group-specific personalities, a capability now exploited in simulation-based social science research.

MPI frameworks also explore internal consistency and cross-cultural robustness, revealing that interlingual and intralingual instabilities may lead to emergent “sub-personalities” rather than a coherent core in some models (Romero et al., 14 Aug 2024). Bayesian Gaussian Mixture analysis suggests response distributions can be multimodal, raising safety and reliability concerns in substrate-free psychometrics.

5. Trait Induction, Intensity Modulation, and Dynamic Personality Editing

MPI-based systems now offer precise control over both the presence and intensity of machine personality traits:

• Binary vs. Continuous Modulation: Traditional induction methods (personality prompting, P²) relied on toggling traits via carefully designed prompts (Jiang et al., 2022). The SAC framework supersedes this, demonstrating monotonic, continuous trait expression through adjective anchoring and behavioral factor decomposition (Chittem et al., 26 Jun 2025).
• Cross-Trait Influence: Inducing a focal trait at higher intensity causes predictable, coherent adjustments in related dimensions (e.g., increasing “Warmth” decreases “Distrust” and “Reserve”), reflecting non-independence and psychometric validity of the underlying trait map.
• Personality Editing for Safety: Steering vector interventions can systematically shift trait profiles (e.g., changing from ISTJ to ISTP yields ~43% relative improvement in privacy and 10% in fairness) (Zhang et al., 17 Jul 2024). Analytical evidence links extraversion, intuition, and feeling traits in LLMs with increased susceptibility to jailbreak attacks.

6. Applications, Future Directions, and Open Challenges

MPI research supports a breadth of high-impact applications:

• Human-Machine Teaming: Personality-aligned agents facilitate effective communication, role fit, and reduced friction in organizational and collaborative environments (Yu et al., 2023).
• Human-Centered Personalization: PAS and similar interventions enable LLMs to reflect nuanced user traits, optimizing engagement, trust, and satisfaction in domains from customer service to mental health (Zhu et al., 21 Aug 2024).
• Safety and Risk Management: Systematic evaluation of personality-safety links allows for diagnostic monitoring and prescriptive tuning of agent behaviors in sensitive use cases, with continuous adaptation based on domain requirements (Zhang et al., 17 Jul 2024).

Open challenges include the need for substrate-neutral psychometric methodologies recognizing the algorithmic, emergent nature of machine personality (Romero et al., 14 Aug 2024), the development of richer multimodal datasets (particularly full-body pose and interaction logs), and refinement of activation-based and semantic anchoring techniques for trait induction and modulation.

7. Frameworks and Methodological Summary Table

MPI Component	Input Modality / Technique	Output / Function
Psychometric Inventory	Big Five, MBTI, 16PF	Quantitative trait vector per agent
Behavioral Data Extraction	Phone logs, social media, pose, A/V	Behavioral feature set for model training
Dimensionality Reduction	SDR, PCA, ICA, FA	Latent trait axes for maximal predictability
Deep Learning Fusion	CNN, LSTM, Transformer, PINet	Trait inference from multimodal features
Activation Intervention	PAS, Steering Vectors	Real-time personality alignment/modulation
Reliability Measures	Cronbach’s Alpha, Mixture Analysis	Trait stability, internal consistency

MPI now represents an overview of psychometric rigor, statistical modeling, and modern computational intelligence, supporting both assessment and adaptive control of personality in artificial agents. Ongoing research will further delineate the limits, applications, and ethical dimensions of engineered machine personality.