VISTA preprint reports AI personalization with fewer survey answers

Researchers report that VISTA, a method for personalizing language models, retained 94% of the gain from a full survey history while stopping after 6% of the World Values Survey questionnaire. Standard adaptive testing retained at most 42% in that comparison. The method infers a user's values from survey answers and converts that profile into an intervention that steers a frozen language model without adding prompt tokens. With the same evidence, it matched or beat profiles placed in prompts across Food2D, the World Values Survey and PrimeX. The findings concern survey and synthetic preference tasks.