人格潜变量：利用个人写作对冻结语言模型进行个体化条件建模

# Personality Latents: Conditioning Frozen Language Models on Individuals from Their Own Writing
Source: [https://arxiv.org/html/2610.36805](https://arxiv.org/html/2610.36805)
Gunhee ChoYun\-Gyung Cheong††thanks:Corresponding author\.Affiliation:Dept\. of AI, Sungkyunkwan UniversityAffiliation:\{skate4333, aimecca\}@skku\.edu

###### Abstract

Controlling the personality of a language model usually relies on written profiles or named trait categories, so it can only express what an author writes down or what a taxonomy names\. We instead proposePersonality Latents \(PL\), which learn a latent for each person from their own writing, with no personality labels and no authored profile, and inject it into a frozen language model as a soft prefix\. Following generative latent optimization, each training person receives a free latent that is optimized through the frozen model itself, while an encoder learns to predict these latents so that a new person needs only one forward pass, with optional refinement at test time\. We evaluate on360360held\-out people by having the conditioned model answer neutral and interview questions and reading personality from its answers with scorers trained on real human writing\. On the MBTI axes PL reaches an AUC of0\.7820\.782, ahead of persona embedding and prompt\-based baselines under identical settings, and the ranking holds when the answers are scored in three different embedding spaces\. A latent learned with Qwen2\.5\-3B\-Instruct also transfers personality to Llama\-3\.1\-8B\-Instruct through a new adapter alone\.

## 1Introduction

Controlling the personality of a language model has become a practical requirement\. Role playing agents must stay in character, assistants are expected to adapt to the individual they serve, and simulated respondents are used to pilot studies before human data is collected\. Each of these needs a stable disposition, and personality is the natural way to specify it because psychology has made it measurable\.

Two representative approaches give a model personality\. Profile based methods describe a character in text and prompt or fine tune the model on that description\([Shao et al\., 2023](https://arxiv.org/html/2610.36805#bib.bib4);[Wang et al\., 2024](https://arxiv.org/html/2610.36805#bib.bib5)\), so what can be conveyed is limited to what an author wrote and every new character needs its own training run\. Trait based methods condition on categories, through dialogue generated under Big Five levels\([Li et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib3)\), one model per MBTI type\([Cui et al\., 2023](https://arxiv.org/html/2610.36805#bib.bib6)\), or steering along a trait direction\([Bhandari et al\., 2026](https://arxiv.org/html/2610.36805#bib.bib7)\)\. Their resolution is that of the taxonomy, so two people high on extraversion are treated identically and anything the taxonomy does not name cannot be requested\.

Persona embedding methods escape both limits by learning a vector from a person’s own behavior and injecting it into a frozen model\([Ning et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib8);[Liu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib9);[Qiu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib10)\), or by moving this conditioning into per user adapters\([Tan et al\., 2024](https://arxiv.org/html/2610.36805#bib.bib11);[Tan et al\., 2026](https://arxiv.org/html/2610.36805#bib.bib12)\)\. Three properties still stand between them and personality\. The person vector is constrained to be the output of an encoder, so the search for a representation is confined to what that encoder can express\. The vector is learned together with one target model, which makes it a coordinate in that model’s input space rather than a portable description of the person\. And these methods are built for preference prediction and task accuracy on benchmarks such as LaMP, with their authors noting that the embedding captures behavior rather than the way a person writes\([Liu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib9)\)and that evaluation has so far been confined to a single setting\([Qiu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib10)\), so whether such a vector carries personality remains open\.

We proposePersonality Latents \(PL\), which learn a personality latent from a person’s own writing, with no labels and no authored profile, and inject it into a frozen backbone LLM\. Following generative latent optimization\([Bojanowski et al\., 2018](https://arxiv.org/html/2610.36805#bib.bib1)\), each training person receives a free latent optimized jointly with the adapter through the backbone LLM, so the representation is not confined to what an encoder can produce\. An encoder trained alongside predicts these latents for unseen people in one forward pass, and the same objective can refine a latent at test time\. Because the objective does not tie the latent to one backbone LLM, a single vector can drive several models through their own small adapters\. We evaluate by having the conditioned model write and then reading personality from that writing with scorers fitted on real human posts, comparing against the documented personality of people never seen in training\.

Our contributions are the following\.

- •Individual level personality conditioning without labels or profiles\.A latent estimated from a person’s writing conditions a frozen backbone LLM on that person rather than a category, with an amortizing encoder for unseen people and optional test time refinement\.
- •A latent that is not tied to one backbone LLM\.The same frozen vector transfers personality through a model never involved in learning it\.
- •Evaluation against documented personality\.We read personality from what the conditioned model writes using scorers fitted on real human posts, compare against documented MBTI and Big Five labels for held out people, and benchmark persona embedding and prompt based methods under identical settings\.

## 2Related Work

### 2\.1Giving a persona to a language model

Profile based methods write the persona down\. Character\-LLM converts a character’s profile into experiences and trains an agent on them\([Shao et al\., 2023](https://arxiv.org/html/2610.36805#bib.bib4)\), and RoleLLM builds role profiles for a hundred characters and tunes models to speak in their style\([Wang et al\., 2024](https://arxiv.org/html/2610.36805#bib.bib5)\)\. Because the description is authored, only what the author thought to record can be conveyed, and the tuned variants need a separate training run for each character\.

Trait based methods condition on categories drawn from a psychological taxonomy\. Big5\-Chat generates dialogue under Big Five trait levels and trains on it\([Li et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib3)\), while Machine Mindset fine tunes one model per MBTI type through supervised tuning followed by preference optimization\([Cui et al\., 2023](https://arxiv.org/html/2610.36805#bib.bib6)\)\. A related line edits the model’s internal state instead of its weights, building a direction in activation space from contrastive prompts and adding it during decoding to raise or lower a named trait\([Bhandari et al\., 2026](https://arxiv.org/html/2610.36805#bib.bib7)\)\. All of these take a label as the unit of control, so the conditioning is as fine grained as the taxonomy and no finer\. Our target is a particular individual rather than a category, and no trait is named at any point\.

### 2\.2Persona embeddings

Persona embedding methods represent the person as a vector learned from their own behavior and inject it into a frozen model\. USER\-LLM pretrains a user encoder on interaction histories and contextualizes the model through cross attention\([Ning et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib8)\)\. Persona\-Plug aggregates a user’s history with weights that depend on the current input, so the user vector changes from query to query\([Liu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib9)\)\. DEP codes a user relative to peers who engaged with similar content and compresses the result with a sparse autoencoder before injection\([Qiu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib10)\)\. Others move the conditioning into parameters, training an adapter per user\([Tan et al\., 2024](https://arxiv.org/html/2610.36805#bib.bib11)\)or generating adapter weights from a profile with a hypernetwork\([Tan et al\., 2026](https://arxiv.org/html/2610.36805#bib.bib12)\)\. Unlike these methods, our latent is a free parameter optimized through the LLM rather than an encoder output, is shared across LLMs through separate adapters, and is evaluated on personality rather than on preference prediction\.

![Refer to caption](https://arxiv.org/html/2610.36805v1/persona_embedding.png)Figure 1:Overview of the proposed framework\. During training, a free person\-specific latent is optimized through a frozen LLM, while an amortizing encoder learns to predict it from the person’s posts\. The latent is converted into soft prefix tokens by an adapter\. For an unseen person, the encoder prediction is used directly or optionally refined at test time before generation\.

## 3Method

We want a single low\-dimensional vector that carries a person’s personality and, when injected into a*frozen*language model, makes that model answer as that person would\. Let𝒫\\mathcal\{P\}denote a set of people, and letXi=\{xi​1,…,xi​Ki\}X\_\{i\}=\\\{x\_\{i1\},\\dots,x\_\{iK\_\{i\}\}\\\}be the posts written by personii\. We seek a latent𝐳i∈ℝd\\mathbf\{z\}\_\{i\}\\in\\mathbb\{R\}^\{d\}and a light conditioning module such that the frozen LLM, conditioned on𝐳i\\mathbf\{z\}\_\{i\}, reproduces personii’s personality\. We call𝐳i\\mathbf\{z\}\_\{i\}the personality latent of personii, and refer to the resulting method asPersonality Latents \(PL\)\. No personality labels are used at any point during training; the only supervision is the person’s own text\.

### 3\.1Conditioning a Frozen LLM

The latent is injected as a soft prefix\. An adaptergθ:ℝd→ℝP×dmodelg\_\{\\theta\}:\\mathbb\{R\}^\{d\}\\rightarrow\\mathbb\{R\}^\{P\\times d\_\{\\text\{model\}\}\}maps𝐳\\mathbf\{z\}toPPvectors in the LLM’s input\-embedding space,

gθ​\(𝐳\)=reshape⁡\(W2​σ​\(LN⁡\(W1​𝐳\)\)\),g\_\{\\theta\}\(\\mathbf\{z\}\)=\\mathrm\{reshape\}\\big\(W\_\{2\}\\,\\sigma\(\\mathrm\{LN\}\(W\_\{1\}\\mathbf\{z\}\)\)\\big\),\(1\)withσ\\sigmaa GELU nonlinearity andLN\\mathrm\{LN\}layer normalization\. ThesePPvectors are concatenated in front of the token embeddings of the prompt, and the LLM runs on the concatenated sequence:

𝐇=\[gθ​\(𝐳\);E⁡\(𝐭\)\],p⁡\(y∣𝐭,𝐳\)=LM⁡\(𝐇\),\\mathbf\{H\}=\\big\[\\,g\_\{\\theta\}\(\\mathbf\{z\}\)\\;;\\;E\(\\mathbf\{t\}\)\\,\\big\],\\qquad p\(y\\mid\\mathbf\{t\},\\mathbf\{z\}\)=\\mathrm\{LM\}\(\\mathbf\{H\}\),\(2\)
whereEEis the LLM’s frozen embedding table and𝐭\\mathbf\{t\}the prompt tokens under the model’s chat template\. The LLM’s parameters are never updated\. The only trainable parts are the adapter, the encoder of §[3\.3](https://arxiv.org/html/2610.36805#S3.SS3), and the latents themselves\.

### 3\.2Latents as free parameters

The usual choice is an autoencoder, in which a parametric encoder produces𝐳i=fϕ​\(Xi\)\\mathbf\{z\}\_\{i\}=f\_\{\\phi\}\(X\_\{i\}\)and the LLM reconstructs the person’s text\.

This setup mixes two possible failure modes\. If the resulting conditioning is weak, it is unclear whether the latent space itself cannot represent the person or whether the encoder fails to find the right point in that space\.

In preliminary experiments this distinction mattered\. A learned encoder did not outperform a PCA projection of the same inputs at matched dimensionality\. Changing the encoder architecture did not remove this gap\.

We therefore use Generative Latent Optimization\([Bojanowski et al\., 2018](https://arxiv.org/html/2610.36805#bib.bib1)\)\. Each training person is assigned a latent vector that is treated as a free parameter, initialized from𝒩⁡\(0,σ02​I\)\\mathcal\{N\}\(0,\\sigma\_\{0\}^\{2\}I\), whereσ0\\sigma\_\{0\}controls the initialization scale andIIis thed×dd\\times didentity matrix, and then updated by gradient descent like any other model parameter\. Training jointly optimizes\{𝐳i\}i∈𝒫\\\{\\mathbf\{z\}\_\{i\}\\\}\_\{i\\in\\mathcal\{P\}\}andθ\\theta\.

Because the latent of a training person is optimized directly rather than produced by a network, it is not restricted by the encoder\. For these people the model can reach any solution available to the encoder and potentially others as well, so the encoder no longer confounds the question of whether the latent space can represent a person\. For people unseen in training no free latent exists, and we return to the encoder, whose prediction can then be refined at test time \(§[3\.4](https://arxiv.org/html/2610.36805#S3.SS4)\)\.

#### Reconstruction\.

The primary signal is the likelihood of the person’s own posts under the conditioned LLM, written in the same chat format used at inference, so that training and inference see the same prefix position:

ℒrec=𝔼x∼Xi​\[−log⁡p⁡\(x∣𝐭,𝐳i\)\]\.\\mathcal\{L\}\_\{\\text\{rec\}\}=\\mathbb\{E\}\_\{x\\sim X\_\{i\}\}\\big\[\-\\log p\(x\\mid\\mathbf\{t\},\\mathbf\{z\}\_\{i\}\)\\big\]\.\(3\)

#### Margin\.

Reconstruction alone admits a degenerate solution in which the adapter learns one generic prefix and ignores𝐳\\mathbf\{z\}\. We require that a person’s own latent explain their text better than another person’s latent by a marginmm:

ℒmargin=\[m\\displaystyle\\mathcal\{L\}\_\{\\text\{margin\}\}=\\big\[\\,m−NLL⁡\(x∣𝐳j\)\\displaystyle\-\\mathrm\{NLL\}\(x\\mid\\mathbf\{z\}\_\{j\}\)\(4\)\+NLL\(x∣𝐳i\)\]\+,j≠i\.\\displaystyle\+\\mathrm\{NLL\}\(x\\mid\\mathbf\{z\}\_\{i\}\)\\,\\big\]\_\{\+\},\\quad j\\neq i\.where\[⋅\]\+\[\\cdot\]\_\{\+\}is the hinge andjjis drawn from the same batch\.

#### Norm control\.

Free latents are otherwise unconstrained and drift away from the region of the space on which the adapter was trained, which lowers likelihood while degrading generation\. We therefore addℒnorm=∥𝐳i∥2\\mathcal\{L\}\_\{\\mathrm\{norm\}\}=\\lVert\\mathbf\{z\}\_\{i\}\\rVert^\{2\}\.

### 3\.3Amortization: retaining an encoder

Free latents exist only for people seen in training\. To generalize we train, in the same loop, a permutation\-invariant set encoderfϕf\_\{\\phi\}that predicts the free latent from the person’s posts\. Each post is embedded by a frozen authorship encoder\([Rivera\-Soto et al\., 2021](https://arxiv.org/html/2610.36805#bib.bib2)\), and the resulting set is pooled with learned queries:

fϕ​\(Xi\)\\displaystyle f\_\{\\phi\}\(X\_\{i\}\)=Wo​Poolϕ​\(\{Enc⁡\(xi​k\)\}k=1K\),\\displaystyle=W\_\{o\}\\,\\mathrm\{Pool\}\_\{\\phi\}\\big\(\\\{\\mathrm\{Enc\}\(x\_\{ik\}\)\\\}\_\{k=1\}^\{K\}\\big\),\(5\)ℒamort\\displaystyle\\mathcal\{L\}\_\{\\text\{amort\}\}=∥fϕ​\(Xi\)−sg⁡\[𝐳i\]∥2\.\\displaystyle=\\big\\lVert f\_\{\\phi\}\(X\_\{i\}\)\-\\mathrm\{sg\}\[\\mathbf\{z\}\_\{i\}\]\\big\\rVert^\{2\}\.
wheresg⁡\[⋅\]\\mathrm\{sg\}\[\\cdot\]is a stop\-gradient, so amortization chases the latents without distorting them\. Pooling is order\-invariant, so the encoder is unaffected by the order in which a person’s posts arrive\.

The full training objective is

ℒ=ℒrec\+λm​ℒmargin\+λa​ℒamort\+λn​ℒnorm,\\mathcal\{L\}=\\mathcal\{L\}\_\{\\text\{rec\}\}\+\\lambda\_\{m\}\\mathcal\{L\}\_\{\\text\{margin\}\}\+\\lambda\_\{a\}\\mathcal\{L\}\_\{\\text\{amort\}\}\+\\lambda\_\{n\}\\mathcal\{L\}\_\{\\text\{norm\}\},\(6\)minimized over\{𝐳i\}\\\{\\mathbf\{z\}\_\{i\}\\\},θ\\thetaandϕ\\phi\. Two optimizers are used: one forθ,ϕ\\theta,\\phiand one, with a larger step size, for the free latents\.

### 3\.4Inference by optimization

For an unseen person the encoder gives a prediction, but nothing forces that prediction to be the best point in the space\. Since the adapter and the LLM are fixed at test time, we can continue the same objective and optimize only theddcoordinates of the latent:

𝐳i⋆\\displaystyle\\mathbf\{z\}^\{\\star\}\_\{i\}=arg​min𝐳−𝔼x∼Xi​\[log⁡p⁡\(x\|𝐭,𝐳\)\]\+λn​∥𝐳∥2,\\displaystyle=\\operatorname\*\{arg\\,min\}\_\{\\mathbf\{z\}\}\-\\mathbb\{E\}\_\{x\\sim X\_\{i\}\}\[\\log p\(x\\,\|\\,\\mathbf\{t\},\\mathbf\{z\}\)\]\+\\lambda\_\{n\}\\lVert\\mathbf\{z\}\\rVert^\{2\},\(7\)𝐳\(0\)\\displaystyle\\mathbf\{z\}^\{\(0\)\}=fϕ​\(Xi\)\.\\displaystyle=f\_\{\\phi\}\(X\_\{i\}\)\.This is a one\-time cost per person:ddparameters for a few dozen gradient steps, after which the latent is stored and reused for every subsequent generation\. The encoder prediction remains available when no such cost is acceptable, so the method offers two operating points without retraining, and the gap betweenfϕ​\(Xi\)f\_\{\\phi\}\(X\_\{i\}\)and𝐳i⋆\\mathbf\{z\}^\{\\star\}\_\{i\}measures what amortization gives up\.

### 3\.5LLM\-independent latents

Nothing in the objective ties𝐳\\mathbf\{z\}to one LLM\. When several LLMs are available we keep a single latent per person and give each LLM its own adaptergθ\(m\)g\_\{\\theta^\{\(m\)\}\}, selecting one LLM per step:

ℒrec\(m\)=𝔼x∼Xi​\[−log⁡p\(m\)​\(x∣𝐭,𝐳i\)\],\\mathcal\{L\}\_\{\\text\{rec\}\}^\{\(m\)\}=\\mathbb\{E\}\_\{x\\sim X\_\{i\}\}\\big\[\-\\log p^\{\(m\)\}\(x\\mid\\mathbf\{t\},\\mathbf\{z\}\_\{i\}\)\\big\],\(8\)wheremmis drawn uniformly from\{1,…,M\}\\\{1,\\dots,M\\\}at each step\. The latent is then a shared coordinate that each adapter translates into its own model’s input space, which lets the same personality vector drive LLMs of different families and sizes\.

## 4Experimental Setup

### 4\.1Data

We train on corpora of person level writing and evaluate against personality that was documented independently of us\. Kaggle MBTI provides8,6088\{,\}608people with four MBTI axes taken from self reported forum flair, Essays\([Pennebaker and King, 1999](https://arxiv.org/html/2610.36805#bib.bib15)\)provides2,0812\{,\}081people with binary Big Five labels from a questionnaire, and PANDORA\([Gjurković et al\., 2021](https://arxiv.org/html/2610.36805#bib.bib14)\)provides1,1781\{,\}178people with continuous Big Five scores from posted test results\. People are split 80/10/10 by a hash of the person identifier, which for PANDORA is the trait tuple described below, so no individual appears in more than one split, and labels are never used during training\. Each person keeps between88and6464posts of at least8080characters, and the single long document of each Essays author is segmented at sentence boundaries into units of at most320320characters\.

MBTI labels in the Kaggle corpus are encoded as a four bit integer\. Decoding each bit to its axis reproduces the known type distribution of that corpus, with1,8191\{,\}819INFP,1,4621\{,\}462INFJ and1,2971\{,\}297INTP\. Training uses PANDORA through a public mirror that keeps the text and the trait scores but drops all author identifiers, so we group posts by their trait tuple, which is unique for every person in that corpus, and use this tuple as the person identifier both for grouping and for the split\.

### 4\.2Model and training

Each post is first turned into a 512 dimensional vector by a frozen LUAR\-MUD encoder\([Rivera\-Soto et al\., 2021](https://arxiv.org/html/2610.36805#bib.bib2)\)\. The set encoder then summarizes up to 32 of these post vectors into a single prediction of the person’s latent\. It does so with attention, using four learned queries that each attend over all posts, followed by a hidden layer of size 512\. Each latent hasd=128d=128dimensions\. Text is generated by a Qwen2\.5\-3B\-Instruct LLM, which stays frozen throughout\. To condition it on a person, the adapter maps the latent toP=24P=24soft prefix vectors that are placed in front of the prompt, where they act like a system prompt written in vectors rather than words\.

Training runs for30,00030\{,\}000steps\. Each step samples1616people and22of their texts per person as reconstruction targets, truncated to160160tokens\. The adapter and encoder are updated with AdamW at a learning rate of3×10−43\\times 10^\{\-4\}under a one cycle schedule\. The free latents are updated separately with Adam at a larger learning rate of10−210^\{\-2\}, since each latent receives a gradient only when its person is sampled\. The loss weights areλm=λa=1\\lambda\_\{m\}=\\lambda\_\{a\}=1andλn=10−2\\lambda\_\{n\}=10^\{\-2\}, and the margin ism=0\.15m=0\.15\.

Training takes about1717hours on a single 48 GB GPU\. Because every step draws new people and new prompts, the training loss fluctuates and does not clearly show convergence\. We therefore also track the loss on a fixed held out batch\.

Test time refinement, described in §[3\.4](https://arxiv.org/html/2610.36805#S3.SS4), starts from the encoder’s prediction for a new person and improves it by gradient descent on that person’s own posts\. We run6060Adam steps at a learning rate of5×10−25\\times 10^\{\-2\}, using four of the person’s posts per step and the same norm penalty as in training\. This takes about3030seconds per person on one GPU and is done once, after which the refined latent is stored and reused\.

### 4\.3Evaluation protocols

All evaluations use the same LLM and the same held out people,120120from each corpus and360360in total\. None of these people appear in training\.

#### Behavioral protocols\.

These protocols ask whether a person’s personality shows up in what the conditioned model writes\. They proceed in three steps\.

First, we build a scorer that reads personality from text\. It takes the average embedding of a person’s posts and predicts their label, using scikit\-learn logistic regression for binary labels and ridge regression for continuous ones with default settings\. The scorer is trained only on real posts written by people in the training split, so it learns what personality looks like in human writing and never sees text produced by any model\.

Second, we inject a held out person’s latent and ask the model questions\. We use two question sets\. The eight neutral prompts describe an everyday situation without mentioning personality, such as a plan the person was looking forward to being cancelled at the last minute\. The twelve interview questions are designed to reveal a specific trait through a concrete situation without naming it, such as what the person does in the first ten minutes at a party where they know only one person\. Each answer is limited to120120new tokens and sampled with temperature0\.90\.9and nucleus sampling atp=0\.95p=0\.95\.

Third, we average the embeddings of the model’s answers, apply the scorer, and check across people whether the predicted personality matches the person’s documented one\. We repeat all three steps in two further embedding spaces, RoBERTa\-large and all\-MiniLM\-L6\-v2, fitting a new scorer in each\.

#### Latent space analysis\.

This analysis looks at the latents directly, without generating any text\. For held out people, we train a linear probe to predict each trait label from the latent, using five fold cross validation\. The weight vector of each probe gives the direction in latent space that corresponds to that trait\. We also compute the first ten principal components of the latents and test whether any of them correlates with a trait label, controlling the false discovery rate\([Benjamini and Hochberg, 1995](https://arxiv.org/html/2610.36805#bib.bib13)\)at0\.050\.05across all component and label pairs\. These probes are used only for this analysis and not for any protocol above\.

### 4\.4Baselines

All baselines use the same frozen LLM, the same people, the same prompts and the same scoring\. The only thing that changes is how the person is represented to the model\. We compare against three kinds of methods\.

Persona embeddings\.These methods, like ours, turn a person into a vector and inject it as a soft prefix\. PPlug\([Liu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib9)\)combines the person’s posts with weights that depend on the current question, so the vector changes from question to question\. DEP\([Qiu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib10)\)represents the person by how they differ from similar people and compresses this difference with a sparse autoencoder that keeps only thekkstrongest features\. To make the comparison fair, both use the same LUAR post embeddings as our method and the same training budget of30,00030\{,\}000steps, latent sized=128d=128,2424prefix tokens, and the same reconstruction and margin losses\. Any difference in results therefore comes from how the person vector is built\. Because the original papers use different tasks and models, we reimplemented both methods in our setting rather than reproducing their reported numbers\.

Text prompts\.These methods describe the person to the model in words\. TS\([Ning et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib8)\)asks the LLM to summarize ten of the person’s posts into a short description, which is then used as the system prompt\. Trait prompting follows Big5\-Chat and Machine Mindset\([Li et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib3);[Cui et al\., 2023](https://arxiv.org/html/2610.36805#bib.bib6)\)and states the person’s traits in the prompt, using the traits predicted by the same scorers we use for evaluation\. In context places five of the person’s own posts directly in the prompt\.

Controls\.Two controls show what happens without real information about the person\.shuffleinjects the latent of a different, randomly chosen person\. The model is still conditioned, but on the wrong person, so any score above this level must come from matching the right person\.noneuses no conditioning at all\. A method that gives the same answer for everyone scores exactly0\.50\.5on balanced accuracy and AUC\.

### 4\.5Metrics and statistics

For binary labels such as MBTI axes we report ROC AUC and balanced accuracy, which averages the accuracy on each class\. We avoid plain accuracy because the labels are unbalanced\. In the MBTI corpus, for example, most people are introverts, so a method that predicts introvert for everyone would already look accurate\. For continuous Big Five scores we report the Pearson correlationrrbetween predicted and documented scores\.

All comparisons are paired, meaning every method is evaluated on the same people with the same prompts and scorers\. To measure uncertainty we use a bootstrap over people\. We draw2,0002\{,\}000resampled sets of people, using the same resampled sets for every method, and recompute each score on every set\. From these we report95%95\\%confidence intervals andP⁡\(Δ≤0\)P\(\\Delta\\leq 0\), the fraction of resamples in which our method does not outperform the baseline\. A small value means our advantage is unlikely to be a result of which people happened to be sampled\.

Table 1:Personality read from generated text by scorers trained on real human posts in the LUAR feature space, for 360 held out people\. Best is in bold and second best is underlined\. Embedding baselines are PPlug\([Liu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib9)\)and DEP\([Qiu et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib10)\)\. Prompt baselines are trait prompting in the manner of Big5\-Chat\([Li et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib3)\), TS\([Ning et al\., 2025](https://arxiv.org/html/2610.36805#bib.bib8)\)and five of the person’s posts in context\.

## 5Results

We test whether personality appears in what the conditioned model writes \(§[5\.1](https://arxiv.org/html/2610.36805#S5.SS1)\), whether a latent learned with one LLM can be read by another \(§[5\.2](https://arxiv.org/html/2610.36805#S5.SS2)\), and where personality resides in the latent space \(§[5\.3](https://arxiv.org/html/2610.36805#S5.SS3)\)\.

### 5\.1Personality expressed in behavior

Table[1](https://arxiv.org/html/2610.36805#S4.T1)shows how well each method transfers personality into what the model writes\. On the MBTI axes our refined latents score highest under both kinds of questions, reaching an AUC of0\.7540\.754on neutral prompts and0\.7820\.782on interviews\. On Essays and PANDORA the differences between methods are small and no method leads consistently\. Our latents nonetheless remain competitive: one of our two variants ranks among the top three in every column, and on PANDORA the embedding\-based methods, ours included, reach higher correlations than every prompt\-based method under both question sets\. The signal is weak overall, however, with Essays AUCs below0\.610\.61and PANDORA correlations below0\.210\.21\.

#### The ranking does not depend on the scoring encoder\.

Our latents are built from LUAR, an encoder that captures writing style\. If we also score the answers with LUAR, a method could look good simply by copying the person’s style\. To check this, we score the same answers with two encoders that have nothing to do with authorship, RoBERTa\-large and MiniLM, training a new scorer in each, as shown in Table[2](https://arxiv.org/html/2610.36805#S5.T2)\. The numbers change, and LUAR does make our scores look somewhat higher, but our method stays first in every space\. Figure[2](https://arxiv.org/html/2610.36805#S5.F2)checks whether these gaps could be due to chance, using the scores from RoBERTa\-large, where our margin is smallest\. Our advantage over PPlug is significant on both question sets, and our advantage over DEP is significant on neutral prompts\.

Table 2:MBTI AUC for the same generations scored by three independent encoders, LUAR, RoBERTa\-large and all\-MiniLM\-L6\-v2\. Scorers are refitted on real human posts within the space of each encoder\. Best is in bold and second best is underlined\.![Refer to caption](https://arxiv.org/html/2610.36805v1/fig/fig_sig_forest.png)Figure 2:Paired bootstrap over people with 2,000 resamples, on MBTI AUC under RoBERTa\-large scoring\.
#### Test time refinement adds most of the gain\.

Our method has two modes\. The encoder predicts a latent in one step, and test time refinement then improves it on the person’s own posts\. With the encoder alone, our MBTI AUC is0\.6490\.649on neutral prompts, close to the best baselines, and0\.7290\.729on interviews\. Refining each latent for sixty steps raises these to0\.7540\.754and0\.7820\.782\. Under RoBERTa\-large scoring this improvement is significant on both question sets, at\+0\.075\+0\.075withP=0\.013P=0\.013\. Refinement also changes the latent noticeably, leaving a cosine similarity of only0\.6350\.635to the encoder’s prediction\. This shows that the encoder does not reach the best latent for each person, and that the remaining gap can be closed cheaply at test time\.

Table 3:Llama\-3\.1\-8B\-Instruct reading latents it never helped to shape, on120120held out MBTI people\. The latents and the encoder are those trained with Qwen2\.5\-3B\-Instruct, and only the adapter is new\.

### 5\.2One latent, Several LLMs

Our objective does not tie a latent to the LLM it was trained with\. To test this, we keep the latents and the encoder trained with Qwen2\.5\-3B\-Instruct fixed and train only a new adapter for Llama\-3\.1\-8B\-Instruct, for12,00012\{,\}000steps with1212people per step under the reconstruction and margin losses\. No gradient reaches the latents, so Llama takes no part in shaping them\.

Personality still transfers\. On held out MBTI people, Llama conditioned on the right person’s latent reaches an AUC of0\.5920\.592\. With another person’s latent it reaches0\.4920\.492, and with no conditioning0\.4510\.451\. Both gaps are significant\. The gain overshuffleis\+0\.100\+0\.100, with a95%95\\%interval of\[\+0\.002,\+0\.198\]\[\+0\.002,\+0\.198\]andP=0\.022P=0\.022, and the gain overnoneis\+0\.141\+0\.141, with an interval of\[\+0\.048,\+0\.233\]\[\+0\.048,\+0\.233\]andP=0\.004P=0\.004\.

The gain is smaller than with Qwen, which is expected because the latents were shaped only by Qwen\. Still, a model of a different family and size can read the same latent through nothing more than a small new adapter\.

![Refer to caption](https://arxiv.org/html/2610.36805v1/fig/auc_cos.png)Figure 3:Linear probes on the latents of held out people\. Left: probe AUC for each MBTI axis, with chance at0\.50\.5\. Right: cosine similarity between the four trait directions, each taken as the weight vector of its probe\.
### 5\.3What the latent space looks like

Since no labels were used in training, we can ask afterwards where personality ended up in the latent space\. We first train a simple linear probe to predict each MBTI axis from the latents of held out people\. As the left panel of Figure[3](https://arxiv.org/html/2610.36805#S5.F3)shows, all four axes can be recovered well above chance, with an AUC of0\.7410\.741for thinking versus feeling,0\.6870\.687for extraversion versus introversion,0\.6470\.647for judging versus perceiving and0\.6440\.644for sensing versus intuition\. Each probe also gives a direction in the space for its trait\. The right panel shows that these four directions are nearly perpendicular to each other, with an average absolute cosine similarity of0\.1050\.105and no pair above0\.210\.21in magnitude\. The traits are therefore stored largely independently of one another\.

Personality is, however, not the main thing that separates people in the space\. None of the ten leading principal components of the latents correlates significantly with any MBTI axis after Benjamini Hochberg correction at a false discovery rate of0\.050\.05\(Appendix[A](https://arxiv.org/html/2610.36805#A1)\), so a trait is a direction that labels can find but not one of the largest differences between people that the latent encodes\.

## 6Conclusion

We presentedPersonality Latents \(PL\), a way to condition a frozen language model on a specific individual using only that person’s own writing\. Each person is represented by a latent that is optimized directly through the model rather than produced by an encoder, and an encoder trained alongside makes the method usable for people never seen in training\.

Three findings stand out\. First, personality written in a person’s own text can be carried into what a model writes\. PL transfers MBTI axes better than persona embedding and prompt\-based baselines, and this holds across three scoring spaces\. Second, refining the latent at test time accounts for most of this advantage, which shows that an encoder alone does not find the best representation of a person\. Third, the latent is not tied to the model it was trained with, since a different model can read it through a small new adapter\.

The latent space itself suggests where to go next\. Each MBTI axis corresponds to a direction that a linear probe can find, but these directions are not the main source of variation between people\. Making these traits more prominent in the latent, and training one latent jointly with several models, are natural next steps\.

## Limitations

Our scorers detect signals in generated text that they associate with a trait, which is not the same as the model holding the person’s personality, and they may partly respond to topic or vocabulary\. Repeating the scoring in three embedding spaces reduces but does not remove this concern\. Our clearest gains are on MBTI, whose labels come from self reported forum flair and which has weaker psychometric support than the Big Five, while the signal on the Big Five corpora is weak for every method we tested\. All main results use one frozen model with three billion parameters, and the cross model experiment covers only one additional model with the latents kept fixed\. Test time refinement also requires the person’s own posts and a short optimization run per person\.

The method builds a representation of a specific individual and keeps part of their style, so it could be misused to impersonate or profile someone without consent\. We do not release latents of real individuals, and any deployment should require the consent of the person being represented\.

## References

- Benjamini and Hochberg \(1995\)Y\. Benjamini and Y\. HochbergControlling the false discovery rate: a practical and powerful approach to multiple testing\.Journal of the Royal Statistical Society: Series B \(Methodological\)57\(1\),pp\. 289–300\.External Links:ISSN 0035\-9246,[Document](https://dx.doi.org/10.1111/j.2517-6161.1995.tb02031.x),[Link](https://doi.org/10.1111/j.2517-6161.1995.tb02031.x),https://academic\.oup\.com/jrsssb/article\-pdf/57/1/289/49173396/jrsssb\_57\_1\_289\.pdfCited by:[§4\.3](https://arxiv.org/html/2610.36805#S4.SS3.SSS0.Px2.p1.1)\.
- Bhandariet al\.\(2026\)P\. Bhandari, N\. Fay, S\. Selvaganapathy, A\. Datta, U\. Naseem, and M\. NasimActivation\-space personality steering: hybrid layer selection for stable trait control in LLMs\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),V\. Demberg, K\. Inui, and L\. Marquez \(Eds\.\),Rabat, Morocco,pp\. 6388–6403\.External Links:[Link](https://aclanthology.org/2026.eacl-long.300/),[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.300),ISBN 979\-8\-89176\-380\-7Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36805#S2.SS1.p2.1)\.
- Bojanowskiet al\.\(2018\)P\. Bojanowski, A\. Joulin, D\. Lopez\-Pas, and A\. SzlamOptimizing the latent space of generative networks\.InProceedings of the 35th International Conference on Machine Learning,J\. Dy and A\. Krause \(Eds\.\),Proceedings of Machine Learning Research, Vol\.80,pp\. 600–609\.External Links:[Link](https://proceedings.mlr.press/v80/bojanowski18a.html)Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p4.1),[§3\.2](https://arxiv.org/html/2610.36805#S3.SS2.p4.1)\.
- Cuiet al\.\(2023\)J\. Cui, L\. Lv, J\. Wen, R\. Wang, J\. Tang, Y\. Tian, and L\. YuanMachine mindset: an mbti exploration of large language models\.arXiv preprint arXiv:2312\.12999\.Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36805#S2.SS1.p2.1),[§4\.4](https://arxiv.org/html/2610.36805#S4.SS4.p3.1)\.
- Gjurkovićet al\.\(2021\)M\. Gjurković, V\. M\. Karan, I\. Vukojević, M\. Bošnjak, and J\. SnajderPANDORA talks: personality and demographics on Reddit\.InProceedings of the Ninth International Workshop on Natural Language Processing for Social Media,L\. Ku and C\. Li \(Eds\.\),Online,pp\. 138–152\.External Links:[Link](https://aclanthology.org/2021.socialnlp-1.12/),[Document](https://dx.doi.org/10.18653/v1/2021.socialnlp-1.12)Cited by:[§4\.1](https://arxiv.org/html/2610.36805#S4.SS1.p1.1)\.
- Liet al\.\(2025\)W\. Li, J\. Liu, A\. Liu, X\. Zhou, M\. T\. Diab, and M\. SapBIG5\-CHAT: shaping LLM personalities through training on human\-grounded data\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 20434–20471\.External Links:[Link](https://aclanthology.org/2025.acl-long.999/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.999),ISBN 979\-8\-89176\-251\-0Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36805#S2.SS1.p2.1),[§4\.4](https://arxiv.org/html/2610.36805#S4.SS4.p3.1),[Table 1](https://arxiv.org/html/2610.36805#S4.T1)\.
- Liuet al\.\(2025\)J\. Liu, Y\. Zhu, S\. Wang, X\. Wei, E\. Min, Y\. Lu, S\. Wang, D\. Yin, and Z\. DouLLMs \+ persona\-plug = personalized LLMs\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 9373–9385\.External Links:[Link](https://aclanthology.org/2025.acl-long.461/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.461),ISBN 979\-8\-89176\-251\-0Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.36805#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2610.36805#S4.SS4.p2.1),[Table 1](https://arxiv.org/html/2610.36805#S4.T1)\.
- Ninget al\.\(2025\)L\. Ning, L\. Liu, J\. Wu, N\. Wu, D\. Berlowitz, S\. Prakash, B\. Green, S\. O’Banion, and J\. XieUser\-llm: efficient llm contextualization with user embeddings\.InCompanion Proceedings of the ACM on Web Conference 2025,WWW ’25,New York, NY, USA,pp\. 1219–1223\.External Links:ISBN 9798400713316,[Link](https://doi.org/10.1145/3701716.3715463),[Document](https://dx.doi.org/10.1145/3701716.3715463)Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.36805#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2610.36805#S4.SS4.p3.1),[Table 1](https://arxiv.org/html/2610.36805#S4.T1)\.
- Pennebaker and King \(1999\)J\. W\. Pennebaker and L\. A\. KingLinguistic styles: language use as an individual difference\.\.Journal of personality and social psychology77 6,pp\. 1296–312\.External Links:[Link](https://api.semanticscholar.org/CorpusID:29567532)Cited by:[§4\.1](https://arxiv.org/html/2610.36805#S4.SS1.p1.1)\.
- Qiuet al\.\(2025\)Y\. Qiu, T\. Shi, X\. Zhao, F\. Zhu, Y\. Zhang, and F\. FengLatent inter\-user difference modeling for LLM personalization\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 10599–10617\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.536/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.536),ISBN 979\-8\-89176\-332\-6Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.36805#S2.SS2.p1.1),[§4\.4](https://arxiv.org/html/2610.36805#S4.SS4.p2.1),[Table 1](https://arxiv.org/html/2610.36805#S4.T1)\.
- Rivera\-Sotoet al\.\(2021\)R\. A\. Rivera\-Soto, O\. E\. Miano, J\. Ordonez, B\. Y\. Chen, A\. Khan, M\. Bishop, and N\. AndrewsLearning universal authorship representations\.InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,M\. Moens, X\. Huang, L\. Specia, and S\. W\. Yih \(Eds\.\),Online and Punta Cana, Dominican Republic,pp\. 913–919\.External Links:[Link](https://aclanthology.org/2021.emnlp-main.70/),[Document](https://dx.doi.org/10.18653/v1/2021.emnlp-main.70)Cited by:[§3\.3](https://arxiv.org/html/2610.36805#S3.SS3.p1.1),[§4\.2](https://arxiv.org/html/2610.36805#S4.SS2.p1.1)\.
- Shaoet al\.\(2023\)Y\. Shao, L\. Li, J\. Dai, and X\. QiuCharacter\-LLM: a trainable agent for role\-playing\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 13153–13187\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.814/)Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36805#S2.SS1.p1.1)\.
- Tanet al\.\(2024\)Z\. Tan, Q\. Zeng, Y\. Tian, Z\. Liu, B\. Yin, and M\. JiangDemocratizing large language models via personalized parameter\-efficient fine\-tuning\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 6476–6491\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.372/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.372)Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.36805#S2.SS2.p1.1)\.
- Tanet al\.\(2026\)Z\. Tan, Z\. Zhang, H\. Wen, Z\. Li, R\. Zhang, P\. Chen, F\. Mo, Z\. Liu, Q\. Zeng, Q\. Yin, and M\. JiangInstant personalized large language model adaptation via hypernetwork\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 23557–23580\.External Links:[Link](https://aclanthology.org/2026.acl-long.1081/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1081),ISBN 979\-8\-89176\-390\-6Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.36805#S2.SS2.p1.1)\.
- Wanget al\.\(2024\)N\. Wang, Z\.Y\. Peng, H\. Que, J\. Liu, W\. Zhou, Y\. Wu, H\. Guo, R\. Gan, Z\. Ni, J\. Yang, M\. Zhang, Z\. Zhang, W\. Ouyang, K\. Xu, W\. Huang, J\. Fu, and J\. PengRoleLLM: benchmarking, eliciting, and enhancing role\-playing abilities of large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 14743–14777\.External Links:[Link](https://aclanthology.org/2024.findings-acl.878/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.878)Cited by:[§1](https://arxiv.org/html/2610.36805#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36805#S2.SS1.p1.1)\.

## Appendix APrincipal Components of the Latent Space

![Refer to caption](https://arxiv.org/html/2610.36805v1/fig/fig_pca.png)Figure 4:Correlation between the ten leading principal components of the latents and each MBTI axis\. No pair survives correction at a false discovery rate of0\.050\.05, and the three leading components carry only22\.6%22\.6\\%of the variance\.We computed the ten directions along which the latents of held out people vary most and correlated each with the four MBTI axes\. Figure[4](https://arxiv.org/html/2610.36805#A1.F4)shows these forty correlations\. None is significant after Benjamini Hochberg correction at a false discovery rate of0\.050\.05, and the strongest, between the second component and the thinking versus feeling axis, is0\.3930\.393\. The variance is also spread out, with the top three components explaining only22\.6%22\.6\\%of it\.