Could an AI learn which 'persona' to put on for your question — physicist, lawyer, coder — to get a better answer?
Can machine learning predict which persona will help on a specific question?
This explores whether a model can learn to pick, for each individual question, the persona prompt ("you are a physicist," "you are a lawyer") that will make an LLM answer it better, and why that turns out to be harder than it sounds.
This explores whether you could train something to look at a question and choose the persona that will make an LLM answer it best. The most direct test in the corpus says no, at least for factual questions. One study tried 162 personas on 2,410 questions across nine models. Personas didn't raise accuracy on average, and none of the methods tried could reliably predict which persona would help on a given question Do personas in prompts actually improve LLM factual accuracy?. The authors concluded that the per-question effects look mostly random. A predictor can't learn a pattern that isn't there.
The interesting part is why the signal is missing, and other notes help explain it. When the same persona prompt is run several times, the outputs vary as much as, or more than, they do across different personas Why do LLM persona prompts produce inconsistent outputs across runs?. So much of what looks like a 'persona effect' is run-to-run noise. Persona prompts also seem to work only on the surface. They change how the model phrases things without changing what it knows or believes underneath Can persona prompts actually reduce bias in language models?. Research mapping the model's internal 'persona space' finds that one direction dominates it: how far the model has moved from its default Assistant character How stable is the trained Assistant personality in language models?. Put these together and asking a model to be a chemist doesn't open up hidden chemistry knowledge. It mostly changes tone, plus some noise.
The idea of predicting the right persona isn't dead, though. It works well in a nearby field. Recommender systems already do something very similar. AMP-CF splits each user into several latent personas and uses attention to choose which one matters for each candidate item, which improves accuracy and also explains the recommendation Can modeling multiple user personas improve recommendation accuracy?. PersonaAgent goes further and updates a persona at test time, adjusting it against feedback from recent interactions Can personas evolve in real time to match what users actually want?. In preference learning, about ten well-chosen questions are enough to work out which blend of reward functions fits a person Can user preferences be learned from just ten questions?.
The difference is where the persona comes from and whether anything grades it. Persona selection works when the personas are learned from real data and there's a feedback signal to optimize against. Even then, personas built from behavioral data reach only 75–90% accuracy at predicting the direction of A/B test results, and they're least reliable when the true effect is small Can behavior-based personas predict A/B test outcomes?. Persona selection fails when the persona is a label in a prompt and the task is factual recall, where the model's knowledge doesn't depend on its role. So the takeaway is not to build a persona router for factual Q&A. If you want personas to matter, ground them in data and give them something measurable to improve against.
Sources 8 notes
Testing 162 personas across 2,410 factual questions and nine models showed no average accuracy gain from personas, and no method could reliably predict which persona helps per question—suggesting effects are largely random rather than systematic.
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.
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.
AMP-CF separates user representation into latent personas weighted by attention to the candidate item. This candidate-conditional approach improves accuracy by adapting the user representation at prediction time and produces inherent explanations for why items were recommended.
Show all 8 sources
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.
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
LLM agents conditioned on anonymized behavioral data predicted A/B test directions with 0.75–0.90 accuracy across 40 experiments. Predictions were most reliable for large effects and least trustworthy for near-zero effects, making the approach viable for fast pre-screening but not full replacement of live testing.
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
- 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
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Persona Generators: Generating Diverse Synthetic Personas at Scale