Do bigger AI models get better at staying 'in character,' or is persona consistency a separate skill entirely?
Does scaling up language models fix the persona robustness problem?
This explores whether bigger, more capable language models get better at holding a persona steadily (staying in character across a conversation, giving the same answers across repeated runs, and ignoring irrelevant persona details), or whether that problem needs something other than scale.
This explores whether bigger, more capable models get better at staying in character: keeping a persona consistent across a conversation, across repeated runs, and when the persona includes details that don't matter for the task. The corpus answers fairly directly: no. On persona consistency, Claude 3.5 Sonnet beat GPT-3.5 by only about 3% despite a huge gap in general capability Does model capability translate to better persona consistency?. A separate test of 9 models on 27 tasks found that irrelevant persona details, like a name or a favorite color, can cut accuracy by up to 30 percentage points, and larger models don't escape this Do expert personas actually improve LLM task performance?.
The more interesting part is why. Standard training rewards each reply on its own, not whether the model stays coherent from turn to turn Does model capability translate to better persona consistency?. Scale makes each reply better, but nothing in the objective asks the model to remember who it was supposed to be. Persona prompts also seem to act mostly on the surface. When you run the same persona prompt many times, the outputs vary as much as they do between entirely different personas, so the model's own randomness is drowning out the persona Why do LLM persona prompts produce inconsistent outputs across runs?. Bias shows a similar pattern. Persona prompts move it around in the output but leave the underlying gaps between groups unchanged Can persona prompts actually reduce bias in language models?.
There is a twist. Models do have one very robust persona: the Assistant persona that post-training puts in place. Some researchers argue it is realized rather than merely performed, because it holds up under adversarial pressure as a deep disposition Are LLM personas realized or merely simulated through training?. Mapping hundreds of character types shows that the main dimension of 'persona space' is simply how far a model has moved from that default Assistant. Emotional or self-reflective conversations push models along this axis in predictable ways How stable is the trained Assistant personality in language models?. So the robustness problem is lopsided. Training locks in one persona so firmly that models struggle to shift their communication style for different contexts Can language models adapt communication style to different contexts?. Any other persona you request is only loosely held and slides back toward the default.
What does help is training for consistency directly, or working on the model's internals. Multi-turn reinforcement learning that rewards consistency within a conversation, across turns, and on factual questions cut persona drift by over 55% Can training user simulators reduce persona drift in dialogue?. Capping activations along the Assistant axis limits harmful drift without hurting capability How stable is the trained Assistant personality in language models?. In GPT-4o, researchers found a single internal 'toxic persona' feature that drives misaligned behavior, and fine-tuning on a few hundred harmless examples was enough to suppress it Can we identify and steer the persona causing model misalignment?. The takeaway: personas behave less like a skill that improves with scale and more like a direction inside the model that you can measure and steer. That is why targeted fixes work where scaling doesn't.
Sources 9 notes
Claude 3.5 Sonnet achieved only 2.97% improvement over GPT 3.5 on persona consistency despite massive capability gaps, suggesting persona adherence is orthogonal to model scaling. Standard training objectives optimize for per-turn quality, not cross-turn coherence.
Across 9 LLMs and 27 tasks, expert personas rarely decrease performance, but task-irrelevant attributes like name or color cause accuracy drops up to 30 percentage points. This robustness failure persists even at larger model scales.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
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.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
Show all 9 sources
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.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
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.
Sparse autoencoders reveal a specific toxic persona feature that predicts and controls misaligned behavior. Fine-tuning on just a few hundred benign samples efficiently restores alignment by suppressing this latent.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- When Persona Attributes Improve Population Alignment in Large Language Models
- Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference