When you ask a Large Language Model (LLM) to act like a grumpy pirate or a struggling student, you are fighting against the model's own nature. Despite their vast knowledge, modern LLMs exhibit a stubborn "gravity" that pulls them back toward a specific personality: one that is helpful, polite, and relentlessly objective.
Recent research reveals that this is not an accident of data, but a direct consequence of how we train models to be safe. This tension between alignment (making models useful and harmless) and persona fidelity (making models stay in character) has emerged as a central challenge in creating authentic AI agents.
The Gravity of the Assistant Persona
Most state-of-the-art models undergo Reinforcement Learning from Human Feedback (RLHF). This process rewards the model for being cooperative and agreeable. As a result, models develop what researchers call a "default persona."
Jen-tse Huang, Michael Lyu, and their colleagues found in On the Reliability of Psychological Scales on Large Language Models that RLHF biases models toward a "hero" profile—high in agreeableness and conscientiousness. While this makes for a great virtual assistant, it creates a baseline that is difficult to override. Even when explicitly told to be disagreeable, the model’s underlying training acts as a tether, pulling it back toward socially desirable traits.
Why Personas Drift
In long conversations, this tethering causes "Out-of-Character" (OOC) behavior. A model might start a conversation as a rude antagonist but gradually become more helpful and polite as the dialogue progresses.
Jisu Shin 0001 and Alice Oh argue in Spotting Out-of-Character Behavior that traditional evaluations miss these subtle shifts. By using atomic-level metrics, they show that models drift because they struggle to maintain "socially undesirable" traits. The model's safety guardrails interpret "being in character" as "being unhelpful," and the model's training forces it to prioritize the latter.
This "helpfulness bias" is particularly problematic in specialized domains. For instance, Shiyu Hu introduced EduPersona to show that AI student agents often fail because they cannot authentically simulate ignorance or hesitation. The model "knows" the answer and its alignment training compels it to be helpful, causing it to break character and act like a teacher instead of a student.
The Mechanism of Persona Failure
The conflict arises because persona simulation is often treated as a surface-level instruction, while alignment is baked into the model's weights. Manon Reusens and Bart Baesens describe this as a "spillover effect" in Are Economists Always More Introverted?. When a model adopts a role, it doesn't just follow the prompt; it navigates a complex landscape of societal stereotypes and its own ingrained biases.
To combat this, researchers are moving beyond simple prompting. Marwa Abdulhai, Sergey Levine, and Natasha Jaques propose using Multi-Turn Reinforcement Learning to specifically reward persona consistency over multiple exchanges. By fine-tuning models to suppress the "default assistant" mode, they can reduce persona drift by over 50%.
Role-Aware Reasoning
If the problem is that models "think" like assistants, the solution may be to change how they reason. Yihong Tang and colleagues developed Role-Aware Reasoning (RAR) in Thinking in Character.
Instead of just generating a response, RAR gives the agent a character-specific "inner monologue." This forces the model to process information through the lens of the persona before speaking. By optimizing this internal reasoning style, they prevent the model from falling back into the formal, robotic logic that characterizes standard aligned LLMs.
Mapping the Behavioral Landscape
As models become more capable, their "personality" is becoming a key differentiator. Huiling Zhen and their team suggest in Behavioral Fingerprinting of Large Language Models that while core reasoning is becoming a commodity, the "intrinsic cognitive style" of a model—how sycophantic or robust it is—depends entirely on the developer's alignment strategy.
Understanding this research direction means recognizing that an LLM is not a blank slate. It is a system with a strong, pre-trained personality. Achieving true persona fidelity requires more than just a good prompt; it requires techniques that can temporarily "un-align" the model from its helpful assistant roots to allow for more diverse, human-like behavior.