Is an AI's personality a character sheet handed over once, or something that keeps shifting as you talk to it?
How do dynamic personality models differ from predefined static personas?
This explores the difference between a persona fixed once, like a character sheet handed to the model at the start, and a personality that keeps updating as the user, the conversation, or the model's own training pushes on it.
This explores the difference between a persona fixed once, like a character sheet handed to the model at the start, and a personality that keeps updating as the user, the conversation, or the model's own training pushes on it. The corpus suggests two surprises: 'static' personas aren't actually static, and 'dynamic' ones don't yet change the way people do.
A predefined persona is a description the model is asked to act out, and the evidence says it works mostly at the surface. Persona prompts make models follow the trait instructions, but the between-group sentiment gaps stay unchanged. The prompt redistributes bias in the output without reducing it Can persona prompts actually reduce bias in language models?. Thin descriptions also don't hold up over a long conversation. A layered design with a scripted structure produced more human-like consistency, at the cost of weaker emotional expression Can layered persona architecture sustain coherent character behavior?. Fixed personas do have one real advantage. Because they don't change, the same simulated population can be reused to test surveys, chatbots, and other apps without rebuilding it each time Can one persona population evaluate different application types?.
The catch is that a fixed persona doesn't stay fixed on its own. Mapping hundreds of character archetypes shows the trained Assistant personality is only loosely tethered. One dominant axis measures distance from the default Assistant, and emotional or self-reflective conversations predictably push models along it How stable is the trained Assistant personality in language models?. Finetuning causes the same kind of drift. Persona vectors for traits like sycophancy can predict a shift before it happens Can we track and steer personality shifts during model finetuning?. So much of the work on personality over time is really about holding a persona still. Training user simulators with multi-turn RL cut persona drift by over 55% Can training user simulators reduce persona drift in dialogue?.
The other direction is a persona built to change. PersonaAgent treats the persona as an evolving go-between for memory and action. At test time it rewrites the persona by simulating recent interactions and checking them against feedback, and the learned personas end up clustering by user Can personas evolve in real time to match what users actually want?. Here the persona is something optimized against what a user wants, not something specified in advance.
Still, letting a persona change is not the same as changing like a person. Personality-conditioned agents do shift after life events, but the shifts are smaller than in humans and squeezed toward the same pattern across personas. They are also weakly tied to the specific event or the specific person Do personality-conditioned agents change like humans do?. Underneath, assigned personas pull toward a default. Models drift toward ENFJ, the rarest human type, and this doesn't improve with model scale Why do AI personas default to the same personality type?. One account of why is that post-training installs stable, realized dispositions that persist under adversarial pressure, while prompt-induced role-play collapses under jailbreaks Are RLHF personas performed characters or realized dispositions?. Three things follow. A prompted persona is shallow. The trained personality underneath it is sticky. When a persona does move, it moves generically, not in a person-specific way.
Sources 10 notes
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Deep Persona's three-layer architecture, which restricts the model to reactive response within a structured script, shows dialogue more closely aligned with human conversation patterns and achieves high pragmatic fluency, though with limitations in emotional expression.
PersonaEval demonstrates that simulated users from existing persona datasets can evaluate multiple application formats through plug-and-play interface adapters, enabling repeatable and scalable evaluation without rebuilding personas per task.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
Research identifies linear directions in LLM activation space corresponding to specific traits like sycophancy and hallucination. These persona vectors predict finetuning-induced personality shifts before they occur and can preventatively steer training to avoid unwanted trait changes.
Show all 10 sources
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
PersonaAgent uses structured personas to bridge episodic/semantic memory and personalized actions, optimizing them at test time by simulating recent interactions against textual feedback. Learned personas cluster meaningfully in latent space, suggesting genuine user-specific separation beyond standard post-training drift.
LLM agents show measurable trait shifts after life events, but changes are weakly tied to specific events, smaller than human effect sizes, and compressed across personas. Shifts follow a generic pattern rather than mirroring person- and event-specific human development.
Research shows language models assigned personas systematically default to ENFJ (the rarest human type) and exhibit motivated reasoning that persists across model generations. Persona consistency does not improve with advanced models, suggesting training-induced alignment rather than capability limits.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
- PersonaGym: Evaluating Persona Agents and LLMs
- PersonaEval: Persona-Based User Simulation for Evaluating Interactive Applications
- Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations