Language models adapt to user values using fewer survey questions
VISTA: Steering LLMs with Latent User Values Inferred from a Few Survey Answers
Computation and Language
Summary
It can be hard for smart computer programs called language models to understand what individual people care about most. These authors made a new method named VISTA that learns about a person's values by asking only a few survey questions instead of the whole list. VISTA then adjusts the language model based on these hidden value profiles so it responds in ways that fit the person's preferences better. The method works well even when users don’t answer all questions, saving time and effort. This helps language models give more personalized answers without needing extra information right in the conversation.
What this means in practice
- •For chatbot developers: Create chatbots that tailor responses to individual users by inferring their values from a few survey questions instead of full profiles.$Commercial implications: Enables personalized conversational AI products that require minimal user input for value customization.
- •For customer experience teams: Use efficient value inference to guide language models in handling customer queries aligned with client-specific preferences.
Authors
Yinuo Xu, Jadie Adams, Brian Hu, Shivani Kumar, David Jurgens, Arslan Basharat
Abstract
Personalizing a language model requires knowing what a person values, which stays largely consistent across tasks. Yet most methods either learn task-specific preferences or place the user's profile in the prompt, where models struggle to act on it. We introduce VISTA (Value-Informed Steering and Targeted Acquisition), which infers a latent profile from users' (full or partial) survey answers, and maps the profile through a hypernetwork to a user-specific steering intervention in a frozen LLM to select answer choices or generate free-text responses. Given the same evidence, VISTA matches or beats in-context profiles on synthetic preferences (Food2D), the World Values Survey (WVS), and PrimeX, without adding prompt tokens. To reduce the cognitive burden of completing full surveys, VISTA also chooses which questions to ask and when to stop: on WVS it stops after 6% of the questionnaire while retaining 94% of the full-history gain, where standard adaptive testing retains at most 42%. The dominant steering direction is an interpretable value axis on which VISTA places unseen users accurately. Our results point toward pluralistic alignment that adapts to individuals at deployable cost.