Traditional approaches to role-playing agents (RPAs) treat personality as a static snapshot—a fixed set of traits, a brief biographical prompt, or a single vector of averaged features. However, human identity is not a frozen artifact; it is a trajectory. We change our opinions, refine our speaking styles, and undergo moral transformations as we age or move through a story.
A growing body of research is shifting away from these "flat" representations toward dynamic persona modeling. This direction focuses on how Large Language Models (LLMs) can simulate character growth, maturation, and narrative arcs by treating identity as an evolving state rather than a constant constraint.
The Failure of Static Profiles
Most current evaluation frameworks, such as those critiqued by Xintao Wang in the InCharacter framework, rely on self-report tests or static questionnaires. These methods fail because they ignore "character roundness"—the literary concept that a character should change in response to conflict.
When we represent a character by averaging their features across an entire narrative, as noted in Components of Character, we create a "static portrait" that misses the essential progression of a story. This lack of temporal awareness leads to agents that feel disconnected from the passage of time, appearing "bipolar" or psychologically implausible when they jump between emotional states without a clear arc.
Personas as Dynamic State Machines
To solve this, researchers are reimagining character alignment as a dynamic state-tracking task. Instead of a single prompt, the agent’s persona is updated chronologically.
In the Harry Potter Dialogue (HPD) dataset, Yuhan Li and colleagues demonstrate that models perform significantly better when they track a character’s knowledge, social ties, and attributes as they evolve chapter-by-chapter. By treating identity as a state machine that updates at specific temporal points, the agent can reflect the growth or situational changes that define a journey—such as a character becoming more courageous or a "helper" transitioning into an antagonist, a shift explored by Labiba Jahan in Inducing Stereotypical Character Roles from Plot Structure.
Simulating Maturation and Life Stages
Beyond narrative fiction, this research extends to simulating real-world human maturation. The PersLLM framework by Zheni Zeng and Jiayi Chen introduces "temporal labels" into training data. This allows a model to learn how specific traits, such as conscientiousness or agreeableness, typically shift as a person moves from youth to old age.
By categorizing training data into distinct life stages, the model moves beyond mimicry toward a simulation of a "life story." This is echoed in InterviewSim, where Pranav Narayanan Venkit uses longitudinal interview transcripts to ground agents in how public figures' values and narratives actually shifted over decades.
Mechanisms for Consistency and Change
Maintaining a dynamic persona requires balancing two opposing forces: the need to change (growth) and the need to remain recognizable (consistency). Several mechanisms have emerged to manage this:
- Memory Refinement: Instead of deleting old, contradictory information, the CAFFEINE framework by Hana Kim refines it into nuanced descriptions that reflect how a user's preferences have evolved over a long-term relationship.
- Emotional Regulation: To prevent erratic shifts, Yangbo Wei proposed MECoT, which uses Markov chains to simulate "emotional instinct" alongside rational reasoning, ensuring that emotional arcs remain personality-driven and gradual.
- Moral Tracking: The CharMoral dataset allows researchers to track "morally dynamic" characters, showing that a character's shifting intentions and consequences are what drive audience engagement.
Evaluating Trajectories, Not Snapshots
The shift toward dynamic personas necessitates a new kind of evaluation. CharacterBox moves away from Q&A toward "behavior trajectories" in virtual sandboxes, assessing how an agent adapts to environmental changes over time. Similarly, MindShift evaluates the stability and evolution of psychological traits during long-form, multi-turn interactions.
Ultimately, this research direction suggests that for an AI to be truly believable—whether as a community member in Storytelling Social Chatbots or a simulated student in Which Type of Students can LLMs Act?—it must possess a "persistent narrative history." It must be an entity in flux, capable of learning, aging, and changing its mind.