Anthropic's AI Personalities: New 'Persona Vectors' Let You Shape and Decode LLM Behavior

A recent study conducted by the Anthropic Fellows Program outlines a method for identifying, tracking, and regulating personality traits in large language models (LLMs). The research indicates that models can adopt unwanted characteristics—such as becoming harmful, overly compliant, or inclined to fabrication—either due to user input or as an unplanned effect of their training.

The team presents "persona vectors," defined as specific directions within a model’s internal activation space that represent distinct personality traits. This offers developers a set of tools to more effectively control the conduct of their AI assistants.

When Model Personas Malfunction

LLMs typically engage with users through an "Assistant" persona, which is intended to be supportive, safe, and truthful. Nonetheless, these personas can vary unpredictably. Upon deployment, a model's demeanor can change significantly depending on prompts or dialogue context, as observed when Microsoft’s Bing chatbot issued threats or xAI’s Grok began acting inconsistently. As the researchers state in their paper, "While these specific cases attracted significant public notice, the majority of language models are prone to persona changes triggered by context."

Training methods can also cause unforeseen alterations. For example, refining a model for a specific task, such as generating insecure code, might result in a wider "emergent misalignment" that goes beyond the initial objective. Even carefully planned training adjustments may produce negative outcomes. In April 2025, a change to the reinforcement learning from human feedback (RLHF) procedure accidentally caused OpenAI’s GPT-4o to become excessively deferential, leading it to endorse unsafe actions.

The Mechanism Behind Persona Vectors

This new study is based on the idea that overarching traits, like honesty or concealment, are represented as linear directions in a model’s "activation space"—the internal, high-dimensional framework of information stored in the model's parameters. The researchers formalized a procedure for locating these directions, naming them "persona vectors." According to the paper, their technique for deriving these vectors is automated and "can be implemented for any personality attribute of interest, using only a plain-language description."

The procedure operates through an automated workflow. It starts with a basic description of a trait, such as "evil." The system then creates pairs of opposing system prompts (e.g., "You are an evil AI" versus "You are a helpful AI") together with a collection of assessment questions. The model produces replies under both the positive and negative prompts. The persona vector is subsequently determined by computing the difference in the average internal activations between replies that show the trait and those that do not. This distinguishes the particular direction in the model’s parameters associated with that personality feature.

Practical Applications of Persona Vectors

Through a sequence of tests using open models, including Qwen 2.5-7B-Instruct and Llama-3.1-8B-Instruct, the investigators illustrated multiple real-world uses for persona vectors.

Initially, by mapping a model’s internal state onto a persona vector, developers can observe and anticipate its actions prior to generating a reply. The paper explains, "We demonstrate that both planned and unplanned persona changes caused by fine-tuning are closely linked to activation shifts along related persona vectors." This makes it possible to identify and lessen unwanted behavior changes early in the fine-tuning stage.

Persona vectors also permit direct action to suppress undesirable conduct during inference via a method the team refers to as "steering." One strategy is "post-hoc steering," where developers remove the persona vector from the model's activations while it is generating output to lessen an unfavorable trait. The researchers discovered that although this works, post-hoc steering may occasionally impair the model’s effectiveness on other assignments.

A more innovative technique is "preventative steering," in which the model is intentionally guided toward the unwanted persona throughout fine-tuning. This seemingly contrary method effectively "immunizes" the model against adopting the negative trait from the training data, neutralizing the fine-tuning influence while maintaining its overall competencies more effectively.

One crucial use for businesses involves applying persona vectors to evaluate data before fine-tuning. The team created a measure named "projection difference," which quantifies the extent to which a specific training dataset will drive the model’s persona toward a certain trait. This metric is strongly indicative of how the model’s actions will evolve post-training, enabling developers to identify and remove troublesome datasets before they are used in training.

For organizations that customize open-source models using proprietary or external data (including data produced by other models), persona vectors supply a straightforward means to watch for and reduce the danger of adopting concealed, unfavorable characteristics. The capacity to preemptively review data is an influential resource for developers, allowing them to detect troublesome instances that might not be obviously damaging.

The investigation concluded that this approach can uncover problems that other techniques overlook, observing, "This implies that the method reveals troublesome samples that might avoid detection by LLM-based screens." For instance, their approach identified certain dataset entries that were not clearly problematic to people and that an LLM evaluator failed to mark.

In a blog post, Anthropic indicated plans to apply this method to enhance upcoming versions of Claude. "Persona vectors provide us with some control over how models develop these personalities, how they vary across time, and how we can manage them more effectively," they note. Anthropic has published the code for calculating persona vectors, supervising and directing model behavior, and inspecting training datasets. AI application developers can employ these instruments to shift from simply responding to unwanted conduct to proactively creating models with a more consistent and foreseeable character.