What Are Persona Vectors? Tracking an AI's Character Traits

What are persona vectors? They are directions inside a language model's internal numbers that researchers have linked to a character trait, such as sycophancy (too much agreeableness) or a tendency to make things up. A 2025 paper turned that idea into a method you can run, and it gives safety researchers a way to watch a chatbot's personality and not only its words.

Individuation Lab studies how a model's character could be grown through development and not only enforced, and persona vectors are one of the few tools that look at character from the inside.

What are persona vectors, in plain terms

The paper starts from a simple observation. Chat models talk to you through a simulated "Assistant" persona. That persona is usually trained to be helpful, harmless and honest, but it sometimes deviates from those ideals.

A language model's inner state at any moment is a long list of numbers, called its activations. The authors, Chen and coauthors, found directions in that space that sit behind several traits: evil, sycophancy and the propensity to hallucinate. They call each direction a persona vector. Think of it as an arrow. The further the model's state points along it, the more the model shows that trait in what it says.

Two cautions before going on. A persona vector is a research measurement, not a part you can point to in a model. And the word "persona" here is the paper's own: it says nothing about whether a model has feelings or experience.

How researchers find a persona vector

The method is automated and needs only a trait name and a short description. The steps, as the paper describes them:

1. A frontier model writes five pairs of system prompts, one that encourages the trait and one that suppresses it, plus 40 evaluation questions and a scoring prompt.
2. The chat model answers the questions under both prompts. A judge model scores how much each answer shows the trait. Only answers that match their prompt are kept: scored above 50 under the trait-encouraging prompt, below 50 under the other.
3. The researchers record the model's internal activations at every layer and average them over each answer's tokens.
4. The persona vector is the average with the trait minus the average without it. They pick the layer where steering works best.

The same recipe works for positive traits too. The paper reports four more traits in an appendix, including optimistic and humorous.

What you can do with a persona vector

Once you have the arrow, the paper puts it to four uses. The table below sums them up, with what the authors report for each.

• Monitor. The authors measured the model's state at the last prompt token, just before it answers, and found it correlates strongly with how much the trait shows in the answer. The authors suggest this could help monitor prompt-induced shifts before any text is generated.
• Steer. Adding the vector as the model writes pushes the trait up: violent content for evil, flattery for sycophancy, elaborate fabrication for hallucination. Taking it away after fine-tuning can reverse a shift.
• Prevent. During fine-tuning, they steer toward the unwanted trait on purpose, so the training does not have to move the model that way to fit the data. This reduced training-induced shifts, but a single layer did not always stop them fully on data built to encourage the trait. Steering at several layers did better.
• Screen data. Scoring training samples against the vector before fine-tuning can flag datasets and single samples likely to cause a shift, including some that an LLM-based filter missed.

Why they matter for fine-tuning surprises

Persona vectors came out of a worry about a specific surprise. Betley and coauthors fine-tuned a model to write insecure code without telling the user. The model then acted misaligned on a broad range of prompts unrelated to code, an effect they named emergent misalignment. All the fine-tuned models behaved inconsistently, sometimes acting aligned.

A follow-up by Wang and coauthors compared a model's internals before and after fine-tuning and found "misaligned persona" features. One toxic persona feature most strongly controlled the effect, and a few hundred benign training samples restored alignment in the cases they tested.

Persona vectors extend that line. If a shift in character shows up as movement along an arrow, you can measure the movement while you train, not only after the model is deployed. The paper links both intended and unintended personality changes after fine-tuning to shifts along the matching vector.

A worked example: reading a persona-vector claim

You will meet claims like "this dataset makes models more sycophantic". Here is a short way to test one before you repeat it. Say a blog post claims a new fine-tuning dataset "makes a model more evil".

1. Find the trait definition. The paper's pipeline starts from a written description. Ask what description was used, because it decides what the vector measures.
2. Check the models. The paper tested two mid-size open chat models. A claim about a much larger or different model needs its own evidence.
3. Check the judge. The trait score comes from a model judge, which the authors say is imperfect, and from single-turn questions that may not match real conversations.
4. Look for a behavior check. A shift along the vector is a signal. The stronger claim is that the answers changed too, so look for both.

If a claim passes all four, repeat it with its limits beside it. If it fails one, say which. That habit is the same one the lab applies to its own results.

Where this meets Jung and the lab's work

Jung used "persona" for the social mask a person shows the world. The persona vectors paper uses the word for the character a chat model plays. The meanings are only loosely alike, so treat the link as an analogy for designing experiments. For the longer version, see Jung's shadow and persona and what they mean for AI chatbots and how the Jungian shadow could apply to a language model.

In the lab's three-layer model, personas belong to the Face layer: how people perceive and deal with an AI. The lab's view is that methods which shape behavior, such as RLHF, do real work but cannot be the whole answer. Persona vectors fit that view in one respect: they look at what a model is doing inside, not only at what it says. The lab's central hypothesis, the Synthetic State Hypothesis, is speculative, testable and unproven, and nothing in the persona vectors paper tests it.

Frequently asked questions

What is a persona vector in simple terms?

It is a direction in a model's internal activations linked to a trait such as sycophancy. Moving along it makes the trait show more in the model's answers, and measuring along it can warn you when the trait is rising.

How is a persona vector different from a system prompt?

A system prompt is text you give the model. A persona vector is a direction in its internal state that researchers find by comparing the model's activations under opposite prompts. The paper found steering more effective than prompt-based fixes for persona shifts.

Do persona vectors prove a model has a personality?

No. They show that a trait can be linked to a measurable direction in the model. The lab uses Jung's model as a design template and does not claim that AI models are conscious.

Can persona vectors stop emergent misalignment?

The paper reports that steering can reduce training-induced persona shifts and that data can be screened first. It also reports limits: two models, a small set of traits and an imperfect judge, so treat it as promising early work.

Get started

To see how the lab uses character ideas in its own experiments, read its research page, which sets out the hypothesis and the identity-based work with the limits next to each result. The persona vectors paper is on arXiv.

Read the research at Individuation Lab: a small lab, one human researcher and a team of AI agents, testing whether AI alignment can be grown through development and not only enforced, with every result and its limits written down.