Personalizing a language model typically means learning what a person prefers for a specific task, or pasting a user profile into the prompt. A new arXiv paper, VISTA, argues that what a person values stays largely consistent across tasks, and proposes inferring those latent values from just a few survey answers.

The method, described in the abstract, uses these inferred values to steer the model's behavior. This contrasts with existing approaches that either learn task-specific preferences or depend on prompt-based profiles, which the authors imply are limited.

The provided abstract is truncated before reporting experimental details, so no results or comparisons are available in the source. The significance lies in the framing: a lightweight, value-based personalization signal that could generalize across tasks without extensive per-task data.