PaperlayerView exact revision on arXiv

Source-linked AI summary

Personality Latents: Conditioning Frozen Language Models on Individuals from Their Own Writing

Gunhee Cho, Yun-Gyung Cheong

arXiv:2610.36805v1cs.CL

TL;DR

Personality control commonly relies on authored profiles or named trait categories, which do not target an individual beyond what is written or named. Personality Latents learn a free representation from a person’s writing and condition a frozen model, with an encoder for unseen people. On MBTI, the paper reports that refined latents outperform persona-embedding and prompt-based baselines, with the ranking holding across three scoring spaces.

  • Problem

    Profile- and trait-based personality control represents authored descriptions or taxonomy categories, while the paper targets the personality of a specific individual.

    Source: §1 · p. 1, §2.1 · p. 2

  • Method

    Personality Latents optimize a free latent from a person’s writing through a frozen model and train an encoder to predict latents for unseen people.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    Refined latents outperform persona-embedding and prompt-based baselines on MBTI axes, and the ranking holds across three scoring spaces.

    Source: §6 · p. 8

  • Takeaways & Limitations

    A latent learned with one model can transfer personality to a different model through a small new adapter.

    Source: §6 · p. 8

    Generated-text scorers detect trait-associated signals rather than establishing that the model holds a person’s personality, and may respond partly to topic or vocabulary.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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 propose Personality 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 on held-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 of , 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.

1 Introduction

Personality control matters for role-playing agents, personalized assistants, and simulated respondents, but existing profile- and trait-based approaches limit what can be expressed. Personality Latents learn an individual representation from their writing and condition a frozen model, with an encoder for unseen people and a latent that can transfer across models.

1 Introduction

  • Profile-based methods describe characters in text and prompt or fine-tune models, limiting conveyed personality to author-written descriptions; trait-based methods condition on categories, with resolution limited to the taxonomy.Source: §1 · p. 1
  • Persona embedding methods constrain representations to encoder outputs, tie them to one model, and have focused on behavior or task accuracy rather than personality.Prior work notes that embeddings capture behavior rather than how a person writes (Liu et al., 2025), and evaluation has been confined to a single setting.Source: §1 · p. 1
  • Personality Latents optimize a free latent from each person’s writing through a frozen model, while an encoder predicts latents for unseen people.The method uses no personality labels or authored profile, and allows optional test-time refinement and transfer through separate adapters.Source: §1 · pp. 1–2
  • Evaluation reads personality from conditioned-model writing with scorers trained on human posts and compares held-out people against documented labels and baselines.The benchmarks include MBTI and Big Five labels, with persona embedding and prompt-based methods compared under identical settings.Source: §1 · p. 2

2 Related Work

Prior work conditions models on authored profiles, taxonomy categories, or embeddings of a person’s behavior. Personality Latents instead target a specific individual and learn a free latent evaluated for personality.

2.1 Giving a persona to a language model

  • Profile-based methods convey only authored details and require separate training runs for each character.Character-LLM converts profiles into experiences (Shao et al., 2023), while RoleLLM builds profiles for a hundred characters.Source: §2.1 · p. 2
  • Trait-based methods use psychological categories, so control is limited to the taxonomy’s granularity rather than a specific individual.Examples include Big5-Chat, Machine Mindset (Cui et al., 2023), and steering along a named trait (Bhandari et al., 2026).Source: §2.1 · p. 2

2.2 Persona embeddings

  • Persona embedding methods learn person vectors from behavior or interaction histories, but Personality Latents use free parameters, separate adapters across models, and personality evaluation.Prior approaches include user encoders and query-dependent vectors (Ning et al., 2025; Liu et al., 2025), peer-relative embeddings, and user-specific or generated adapters (Tan et al., 2024, 2026).Source: §2.2 · p. 2

3 Method

Personality Latents learns person-specific vectors from writing and uses them to condition a frozen language model. Free latents address encoder constraints during training; an encoder supports unseen people, with optional refinement and adapters that transfer a shared latent across models.

3 Method

  • Personality Latents represents each person with a latent learned from their own posts, without personality labels; the frozen LLM reconstructs that person’s text when conditioned on it.The method seeks a single low-dimensional vector that makes the model answer as that person would.Source: §3 · p. 2

3.1 Conditioning a Frozen LLM

  • The latent is converted into soft-prefix vectors placed before prompt embeddings, while the LLM remains frozen.Only the adapter, encoder, and latents are trainable.Source: Figure 1 · p. 3, §3.1 · p. 3

3.2 Latents as free parameters

  • Each training person receives a free latent optimized jointly with the adapter, allowing solutions beyond those an encoder can produce.This design addresses preliminary results in which a learned encoder did not outperform PCA at matched dimensionality, and changing its architecture did not remove the gap.Source: §3.2 · p. 3
  • Training uses the likelihood of a person’s own posts and a margin requiring their latent to explain those posts better than another person’s latent.A norm penalty controls free-latent drift; the margin addresses the degenerate solution of a generic prefix that ignores the latent.Source: §3.2 · pp. 3–4

3.3 Amortization: retaining an encoder

  • A permutation-invariant set encoder predicts each free latent from that person’s posts, trained alongside the latents using post embeddings and learned-query pooling.Stop-gradient makes amortization chase the latents without distorting them.Source: §3.3 · p. 4

3.4 Inference by optimization

  • For an unseen person, the encoder prediction can be refined by optimizing only the latent; this one-time cost is optional, and the stored result can be reused.The encoder prediction remains available when refinement cost is unacceptable.Source: §3.4 · p. 4

3.5 LLM-independent latents

  • A single person latent can serve multiple LLMs through separate adapters, with an LLM selected at each training step.Each adapter translates the shared latent into its model’s input space.Source: §3.5 · p. 4

4 Experimental Setup

The experiments use person-level writing corpora and documented personality labels held out from training. They evaluate personality read from generated answers with human-writing scorers, comparing conditioning methods under shared settings and paired uncertainty estimates.

4.1 Data

  • The evaluation uses 8,608 Kaggle MBTI, 2,081 Essays, and 1,178 PANDORA participants, split by person so no individual crosses splits; labels are not used in training.People contribute 8–64 posts; Essays authors’ long documents are segmented into units of at most 320 characters.Source: §4.1 · p. 4

4.2 Model and training

  • A frozen Qwen2.5-3B-Instruct model receives 24 soft-prefix vectors mapped from a 128-dimensional latent; post embeddings are pooled to predict the latent.The model is trained for 30,000 steps; free latents use a separate optimizer and larger learning rate.Source: §4.2 · p. 5
  • Optional test-time refinement starts from the encoder prediction and updates the latent using the person’s posts, taking about 30 seconds per person on one GPU.The refined latent is stored and reused.Source: §4.2 · p. 5

4.3 Evaluation protocols

  • Scorers trained only on real training-split human posts predict personality from model answers to eight neutral or twelve interview prompts, with checks repeated in two further embedding spaces: RoBERTa-large and all-MiniLM-L6-v2.Separately, latent-space probes predict trait labels directly from latents using five-fold cross-validation, without generating text.Source: §4.3 · p. 5

4.4 Baselines

  • Baselines include persona embeddings, text prompts, and unconditioned or shuffled-person controls, evaluated with the same model, people, prompts, and scoring.The persona embedding methods are reimplemented in this setting with the same latent size, prefix length, and training budget.Source: §4.4 · pp. 5–6

4.5 Metrics and statistics

  • Binary traits are scored with ROC AUC and balanced accuracy, continuous traits with Pearson r, and paired comparisons use bootstrap confidence intervals over people.The bootstrap uses 2,000 resampled sets and reports P(∆≤0) for the fraction in which the method does not outperform a baseline.Source: §4.5 · p. 6

5 Results

PL’s refined latents score highest on MBTI axes in generated answers, and the ranking holds across scoring encoders. Results also show transfer to a new LLM through a new adapter and recoverable, largely independent MBTI directions in the latent space.

5.1 Personality expressed in behavior

  • PL’s refined latents score highest on MBTI axes under both question types, while differences on Essays and PANDORA are small and no method leads consistently.The reported MBTI AUCs are 0.754 on neutral prompts and 0.782 on interviews; Essays AUCs are below 0.61 and PANDORA correlations below 0.21.Source: §5.1 · pp. 6–7
  • PL stays first across LUAR, RoBERTa-large, and MiniLM scoring; its advantage over PPlug is significant for both question sets, and over DEP for neutral prompts.Table 2 reports MBTI AUC for the same generations scored by three encoders, with scorers refitted in each space.Source: §5.1 · p. 7, Table 2 · p. 7
  • With 60 refinement steps, MBTI AUC rises from 0.649 to 0.754 on neutral prompts and from 0.729 to 0.782 on interviews.Under RoBERTa-large scoring, the improvement is significant for both question sets (+0.075, P = 0.013); the refined latent has cosine similarity 0.635 to the encoder prediction.Source: §5.1 · p. 7

5.2 One latent, Several LLMs

  • On 120 held-out MBTI people, Llama-3.1-8B-Instruct with Qwen-trained latents and a new adapter reaches AUC 0.592, versus 0.492 with another person’s latent and 0.451 unconditioned.The gains over SHUFFLE and NONE are significant; the latents and encoder remain fixed, with no gradient reaching them.Source: §5.2 · p. 7, Table 3 · p. 7

5.3 What the latent space looks like

  • Linear probes recover all four MBTI axes above chance, and their trait directions are nearly perpendicular, with mean absolute cosine similarity 0.105 and no pair above 0.21.Figure 3 shows probe AUCs on the left and cosine similarities between trait directions on the right.Source: §5.3 · p. 8, Figure 3 · p. 8
  • None of the ten leading latent-space principal components significantly correlates with an MBTI axis after Benjamini–Hochberg correction at false discovery rate 0.05.The authors conclude that trait directions are probe-findable but are not among the largest differences between people encoded by the latents.Source: §5.3 · p. 8

6 Conclusion

Personality Latents condition a frozen language model on an individual’s writing using a directly optimized latent, with an encoder enabling use for unseen people. The conclusion reports transfer across scoring spaces, an advantage from test-time refinement, and transfer to another model through a new adapter.

6 Conclusion

  • A latent optimized through the frozen model represents an individual, while an accompanying encoder predicts latents for people unseen during training.The method uses each person’s own writing and does not require personality labels or an authored profile.Source: §6 · p. 8
  • PL transfers MBTI axes better than persona-embedding and prompt-based baselines, with this ranking holding across three scoring spaces.Test-time refinement accounts for most of the advantage, indicating that an encoder alone does not find the best representation of a person.Source: §6 · p. 8
  • A fixed latent can transfer personality to a different model through that model’s small adapter.The conclusion identifies joint training of one latent with several models as a natural next step.Source: §6 · p. 8
  • Linear probes find directions corresponding to MBTI axes, but those traits are not the main source of variation between people in the latent space.The conclusion suggests making these traits more prominent in the latent as a future direction.Source: §6 · p. 8

Limitations

The authors qualify their findings by noting uncertainty about what scorers detect, weak Big Five signals, limited model coverage, and practical constraints on refinement. They also flag potential impersonation or profiling and call for consent.

Limitations

  • Scorers detect trait-associated signals in generated text, which may reflect topic or vocabulary rather than the model holding a person’s personality.Scoring in three embedding spaces reduces but does not eliminate this concern.Source: Limitations · p. 9
  • The clearest gains are on MBTI labels from self-reported forum flair, while every tested method shows weak signal on the Big Five corpora.The authors note that MBTI has weaker psychometric support than the Big Five.Source: Limitations · p. 9
  • Main results use one frozen three-billion-parameter model, and cross-model testing covers only one additional model with latents kept fixed..Source: Limitations · p. 9
  • Test-time refinement requires a person’s own posts and a short optimization run for each person.Source: Limitations · p. 9
  • Because the method retains some individual style, it could enable impersonation or profiling without consent; the authors call for consent in deployment.They do not release latents of real individuals.Source: Limitations · p. 9

A Principal Components of the Latent Space

The analysis finds no significant correlation between the ten leading latent-space principal components and any MBTI axis after correction. The leading components also account for a limited share of variance.

A Principal Components of the Latent Space

Related papers

Loading 2610.36805v1…