Personality Vectors in Language Models

Updated 12 July 2026

Personality vectors are structured numerical representations encoding trait variations, operationalized via latent directions, psychometric profiles, and weight-space deltas.
They enable precision control and monitoring of language model behavior by steering activations and predicting trait expression with strong correlations.
Challenges include geometric dependencies, cross-trait interference, and safety concerns, necessitating robust calibration and disentanglement methods.

Personality vectors are structured numerical representations that encode personality-related variation in language systems. In recent work, the term covers several non-identical objects: residual-stream directions used for activation steering, vocabulary-level trait profiles used to reweight next-token probabilities, weight-space deltas extracted by subtracting a pretrained checkpoint from a personality-finetuned checkpoint, psychometric score vectors such as HEXACO or OCEAN coordinates, and supervised semantic gradients that place traits in a shared lexical geometry (Hoppe et al., 10 Feb 2026, Li et al., 2023, Sun et al., 24 Sep 2025, Ji et al., 2024, Plisiecki, 26 May 2026). This suggests that personality vectors are best understood as a family of operationalizations for representing, measuring, comparing, or controlling personality rather than as a single canonical mathematical object.

1. Representational scope

Recent literature uses personality vectors in several technically distinct ways. Some works treat them as latent directions inside an LLM; others define them over vocabularies, checkpoints, questionnaire score spaces, or shared semantic embeddings. The common thread is that personality is represented numerically so that it can be projected, composed, clustered, or perturbed (Feng et al., 17 Feb 2026, Li et al., 2023, Heydari et al., 17 May 2026, Plisiecki, 26 May 2026).

Representation	Core object	Example papers
Activation-space direction	Residual-stream direction $v$ added to or projected from hidden states	(Hoppe et al., 10 Feb 2026, Feng et al., 17 Feb 2026, Ma et al., 14 Jan 2026, Chen et al., 29 Jul 2025)
Vocabulary- or decoding-space profile	Token-level trait vectors and user control coefficients $\beta$	(Li et al., 2023)
Weight-space delta	$\phi_p = \theta_p - \theta_{pre}$ for a personality condition $p$	(Sun et al., 24 Sep 2025)
Psychometric or semantic profile	OCEAN/HEXACO coordinates, trait-score vectors, or semantic gradients	(Ji et al., 2024, Jacobs et al., 2021, Plisiecki, 26 May 2026, Völkel et al., 2020)

The Big Five dominates this literature. OCEAN appears in activation-steering systems, model-merging systems, automatic personality prediction, semantic-gradient work, and prompting-based reconstruction studies (Hoppe et al., 10 Feb 2026, Sun et al., 24 Sep 2025, Heydari et al., 17 May 2026, Plisiecki, 26 May 2026). Other frameworks remain active where the target task differs: HEXACO is used for latent personality reconstruction from persona descriptions (Ji et al., 2024), the Dark Triad and LLM-specific traits such as evil and sycophancy appear in alignment-oriented studies (Aneja et al., 11 May 2026), and speech-based conversational agents are given an agent-specific ten-dimensional factor structure rather than a human-only taxonomy (Völkel et al., 2020).

2. Activation-space personality directions in LLMs

The most influential recent formulation treats a personality vector as a direction in residual-stream activation space. In the figure-based framework behind Sequential Adaptive Steering, a trait is represented by a probe-derived vector $v_i$ and applied additively as $+\alpha_i v_i$ , where $\alpha_i$ controls trait intensity. The distinctive step is that later trait probes are trained on a union of “Shifted $\cup$ Unshifted” activations, so that subsequent vectors are learned in the presence of earlier interventions rather than in isolation (Hoppe et al., 10 Feb 2026). The same additive logic appears in PERSONA, where steering is written as $h_i^l \leftarrow h_i^l + \alpha v_i$ , and in its dynamic version PERSONA-FLOW, where the turn-level composite vector is injected as $h_t^{l^*} \leftarrow h_t^{l^*} + V_{\text{comp}}$ after coefficient clipping to $\beta$ 0 and magnitude gating with $\beta$ 1 (Feng et al., 17 Feb 2026).

Extraction is usually contrastive. PERSONA-BASE derives ten OCEAN-pole vectors from contrastive activation analysis, selecting the most effective residual-stream layer empirically; on Qwen2.5-7B-Instruct, layer 20 is chosen as the steering layer in the reported ablation (Feng et al., 17 Feb 2026). “Persona Vectors: Monitoring and Controlling Character Traits in LLMs” uses a similar contrastive mean-difference construction over response-token residual activations, with behaviorally filtered positive and negative generations. For Qwen2.5-7B-Instruct it selects layer 20 for evil and sycophancy and layer 16 for hallucination; for Llama-3.1-8B-Instruct it selects layer 16 for all three (Chen et al., 29 Jul 2025).

A distinct but related use appears in Persona-Vector Neutrality Interpolation. PVNI defines a persona vector for trait $\beta$ 2 as $\beta$ 3, computes the neutral displacement $\beta$ 4, and estimates the neutral score by projection and interpolation,

$\beta$ 5

Here the vector is primarily a measurement axis rather than a control knob, but the underlying assumption is the same: trait expression is organized linearly enough in hidden space for projection onto a trait direction to be informative (Ma et al., 14 Jan 2026).

3. Geometry, composition, and interference

A central question is whether personality vectors are independent directions or merely correlated summaries of broader behavioral structure. Some papers argue that trait vectors are approximately orthogonal and support stable algebra. PERSONA reports that scalar multiplication produces largely linear changes in judged personality expression, with examples such as Inventive $\beta$ 6 Openness at $\beta$ 7, Outgoing $\beta$ 8 Extraversion at $\beta$ 9, and Calm $\phi_p = \theta_p - \theta_{pre}$ 0 Neuroticism at $\phi_p = \theta_p - \theta_{pre}$ 1 (Feng et al., 17 Feb 2026). The model-merging approach likewise reports a strong positive correlation $\phi_p = \theta_p - \theta_{pre}$ 2 with $\phi_p = \theta_p - \theta_{pre}$ 3 between the scaling coefficient $\phi_p = \theta_p - \theta_{pre}$ 4 and the target BFI score in single-trait control, and shows that $\phi_p = \theta_p - \theta_{pre}$ 5 continues to strengthen personality expression (Sun et al., 24 Sep 2025).

Other work is more skeptical. “Do Personality Traits Interfere? Geometric Limitations of Steering in LLMs” explicitly evaluates geometric conditioning schemes from unconstrained vectors to hard orthonormalisation. It finds substantial geometric dependence in raw Big Five steering directions and shows that even when hard orthonormalisation yields $\phi_p = \theta_p - \theta_{pre}$ 6, cross-trait behavioral effects remain. The paper’s conclusion is that personality traits occupy a “slightly coupled subspace,” and that hard orthonormalisation can reduce steering strength; among the tested variants, C4 is described as having the “best semantic–geometry trade-off” (Bhandari et al., 23 Jan 2026). This directly challenges the interpretation of Big Five vectors as independently controllable sliders.

Alignment-oriented geometry studies reinforce the view that personality space is structured but not trivial. “Intrinsic Guardrails” stacks 12 trait directions—Big Five, Dark Triad, and LLM-specific traits such as Evil and Sycophancy—into a latent personality space, finding prosocial and antisocial clusters and low-dimensional organization. In Qwen2.5-7B Instruct, $\phi_p = \theta_p - \theta_{pre}$ 7 and $\phi_p = \theta_p - \theta_{pre}$ 8 explain 53.12% of variance, while in the misaligned variant they explain 60.01%. The paper defines a Semantic Valence Vector by averaging normalized prosocial directions and subtracting normalized antisocial ones, then shows a strong guardrail effect: ablating valence-related directions can drive misalignment rates above $\phi_p = \theta_p - \theta_{pre}$ 9, while amplifying them can suppress the failure mode to less than $p$ 0 (Aneja et al., 11 May 2026). A plausible implication is that personality vectors often capture broader social-valence organization in addition to trait-specific content.

4. Measurement, monitoring, and psychometric reconstruction

Personality vectors are increasingly used as measurement primitives rather than only as steering mechanisms. PVNI was introduced precisely because questionnaire-style personality assessment of LLMs is unstable under prompt rewrites and role-play changes. Across Qwen-2.5-7B, Llama-3-8B, and Mistral-7B-v0.1, PVNI reports substantially smaller standard deviations across prompt variants than self-report and open-ended elicitation baselines, while representing each Big Five trait as a score-direction pair $p$ 1 and a five-axis embedding $p$ 2 (Ma et al., 14 Jan 2026).

The same activation-space object can also be used for real-time monitoring. In “Persona Vectors: Monitoring and Controlling Character Traits in LLMs,” final prompt-token projections onto evil, sycophancy, and hallucination vectors predict deployment-time persona drift induced by system prompts or many-shot prompting. Reported overall correlations between pre-generation projection and later trait expression are $p$ 3– $p$ 4. The same paper defines a finetuning shift by projecting the difference between base-model and finetuned-model prompt activations onto the unit persona vector, and finds that these shifts correlate strongly with post-finetuning trait expression, at $p$ 5 to $p$ 6 (Chen et al., 29 Jul 2025).

Prompting-based psychometric reconstruction offers a more explicit questionnaire-style vectorization. “Is persona enough for personality?” treats a short persona description as an implicit projection into a latent HEXACO state, then recovers a six-dimensional trait profile from 60 item responses scored on $p$ 7. The dimension score is

$p$ 8

and the effective reconstructed vector is $p$ 9. For GPT-3.5-Turbo, 3594 out of 5000 provided dimensions are reconstructed consistently, giving 71.88%, but 99.07% of inconsistencies are cases where a provided low trait is reconstructed as high (Ji et al., 2024). A different measurement tradition appears in the DSM-based study of historical persons, where each name is assigned a five-dimensional Big Five score vector by mean cosine-similarity differences between the person-name embedding and curated positive and negative marker adjectives (Jacobs et al., 2021).

5. Alternative operationalizations beyond hidden-state steering

Not all personality vectors are hidden-state directions. UBPL defines personality at the vocabulary level. Its lexicon $v_i$ 0 stores, for each token, a five-dimensional signed profile over Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. At decoding time, personality control is implemented by multiplicatively reweighting next-token probabilities with

$v_i$ 1

so the effective steering object is the interaction between a token personality vector $v_i$ 2 and a user control vector $v_i$ 3. Across six open-source chat models, the paper reports Pearson correlations above 0.94, often above 0.98, between $v_i$ 4 and measured trait expression in single-trait manipulation (Li et al., 2023).

Model merging shifts the representation into weight space. In “Personality Vector: Modulating Personality of LLMs by Model Merging,” a personality vector is defined as the checkpoint delta

$v_i$ 5

where $v_i$ 6 is a personality-specific finetuned model and $v_i$ 7 is the pretrained model. A target model is then modulated by

$v_i$ 8

The paper reports continuous control over trait intensity, multi-trait composition with task arithmetic and DaRE, and transfer into role-playing, Korean, Chinese, and vision-LLMs without additional training at application time (Sun et al., 24 Sep 2025).

Automatic personality prediction systems use yet another notion of vector. HyperPersona processes essays as a document–sentence–word hierarchy, embeds each unit with BERT, converts the resulting hypergraph into a weighted hierarchical graph, and pools the graph to a fixed-length embedding $v_i$ 9. The paper identifies $+\alpha_i v_i$ 0 as the most natural reusable document representation and $+\alpha_i v_i$ 1 as a 16-dimensional task-specific bottleneck immediately upstream of binary trait prediction; reported average accuracy is 63.35 on the Essays dataset, and the ablation shows accuracy dropping from 63.35 in the full multi-level model to 57.95 in the word-only model (Heydari et al., 17 May 2026). In speech-based automatic personality perception, the representation can be explicitly context-conditioned: each speaker in each interaction is assigned a continuous five-dimensional vector $+\alpha_i v_i$ 2, with the paper arguing that $+\alpha_i v_i$ 3 in general. Stressful interactions are reported as more predictive of Neuroticism, and handcrafted acoustic plus non-verbal features outperform ECAPA-TDNN speaker embeddings (Zhang et al., 25 Jul 2025).

6. Psychometric, semantic, multilingual, and agent-centered spaces

A different line of work uses personality vectors to compare psychological constructs in shared lexical geometry rather than to steer models. “Psychological Constructs in Shared Semantic Space” represents each construct as a supervised semantic gradient in GloVe 42B Common Crawl, 300 dimensions, using Supervised Semantic Differential. For Big Five domains and facets, the target is built from IPIP-NEO-300 item-factor associations; each personality vector is then projected into Valence, Arousal, and Dominance coordinates by cosine similarity. Domain-level placements are reported as broadly coherent, while facet-level results are described as exploratory because they rely on sparse questionnaire text (Plisiecki, 26 May 2026). Here the vector is explicitly a semantic direction rather than an activation edit or a score profile.

Cross-lingual work relocates the concept to trait-conditioned word embeddings. GlobalTrait argues that semantic multilingual alignment is insufficient for personality modeling because words with similar meaning across languages need not align in personality association. It therefore learns a separate cross-lingual mapping per Big Five trait and per source language, yielding trait-specific multilingual word vectors. The paper reports an average F-score increase from 65 to 73.4 across the non-English PAN languages when comparing a monolingual model to a multilingual CNN with personality-aligned embeddings (Siddique et al., 2018).

Agent-centered psycholexical research provides yet another vector space. Instead of reusing the Big Five, “Developing a Personality Model for Speech-based Conversational Agents” collects descriptors from surveys, lab interactions, and 30,000 reviews, reduces them to 349 adjectives, and derives ten correlated factors from exploratory factor analysis. The proposed agent-personality coordinates are Confrontational, Dysfunctional, Serviceable, Unstable, Approachable, Social-Entertaining, Social-Inclined, Social-Assisting, Self-Conscious, and Artificial (Völkel et al., 2020). In this setting, a personality vector is neither a latent transformer direction nor a psychometric OCEAN score, but a ten-dimensional perceptual embedding specialized to speech-based conversational agents.

7. Limitations, controversies, and open problems

A persistent limitation is that geometric independence does not guarantee behavioral independence. Hard orthonormalisation can make steering vectors mathematically independent while leaving cross-trait behavioral bleed largely intact, indicating that pairwise cosine overlap is not the whole story (Bhandari et al., 23 Jan 2026). This matters for claims of disentanglement in activation-steering work and for any interface that treats personality control as a set of independent trait sliders.

Several methods also inherit substantial dependency on synthetic supervision or proxy evaluation. SoulBench’s OCEAN labels are generated by a Teacher Model rather than by human psychometrics, and the paper itself notes weak validation of its geometric claims (Wang, 8 Dec 2025). UBPL constructs lexicons from LLM-generated answers and is tokenizer-specific, while not providing a dedicated fluency, coherence, or task-quality trade-off study as control strength increases (Li et al., 2023). Shared-semantic-space approaches explicitly warn that the resulting vectors reflect the semantics of questionnaire wording and corpus associations as much as the latent construct itself, especially for sparse facet text (Plisiecki, 26 May 2026).

Context sensitivity further complicates the notion of a stable personality vector. In conversational speech, perceived personality differs significantly across interview and client-conflict scenarios, cross-scenario transfer is poor, and a single static speaker descriptor is therefore a poor fit (Zhang et al., 25 Jul 2025). In LLM role-play and affective simulation, Persona-E $+\alpha_i v_i$ 4 identifies “personality illusion,” the tendency to rely on surface stereotypes rather than authentic cognitive logic, and reports that Big Five information significantly improves comprehension of reader-conditioned emotional variation (Yang et al., 10 Apr 2026). Prompt-based HEXACO reconstruction likewise shows a strong positivity bias: omitted or low traits tend to drift high (Ji et al., 2024).

Safety considerations are now inseparable from this topic. Valence-related persona vectors can act as intrinsic guardrails against emergent misalignment (Aneja et al., 11 May 2026), but the same intervention machinery can be used to amplify harmful traits such as evil, sycophancy, or hallucination (Chen et al., 29 Jul 2025). This leaves several open problems in view: robust disentanglement, human-grounded calibration, cross-model transfer beyond shared backbones, uncertainty estimates for semantic trait vectors, and trait control that remains effective without degrading fluency, coherence, or general capability (Bhandari et al., 23 Jan 2026, Sun et al., 24 Sep 2025, Plisiecki, 26 May 2026).