Instead of rating hundreds of items, what if an AI could learn your taste from just ten smart questions?

This explores how a system can learn what you like by asking a few smart questions instead of making you rate hundreds of things. The collection has one direct answer and several related ones that show what that answer leaves out. The direct answer comes from PReF Can user preferences be learned from just ten questions?. It starts by learning a small set of shared 'base' preference dimensions from many users. Each new person is then described as a mix of those dimensions. Active learning picks each next question so that it removes as much uncertainty about your mix as possible. About ten adaptive questions turn out to be enough. Nothing in the model gets retrained; it is steered toward you at the moment it answers. The key move is factorization. Once preferences are described as a few shared dimensions, the system only has to estimate a handful of numbers about you, not your whole taste profile.

The collection also raises a quieter problem: some answers don't measure a real preference. Behavioral science work on annotation Do all annotation responses measure the same underlying thing? splits responses into three kinds. Some are genuine preferences. Some are non-attitudes, meaning the person has no real view and answers anyway. Some are constructed preferences, made up on the spot because the question was asked. This matters a lot for active learning. If ten questions are carrying all the weight, a few made-up or empty answers can skew your whole profile. A question that seems highly informative to the algorithm may be exactly the kind that pushes people to invent an answer.

One option is to skip questions altogether. M3-Agent Can agents learn preferences by watching rather than asking? learns preferences by watching continuous multimodal input. It stores what happened (events) separately from what it concludes about each person, and builds lasting knowledge about them over time. That moves the cost of stating preferences from the user's effort to the system's attention. Related work on personalization Does abstract preference knowledge outperform specific interaction recall? finds that compact summaries of what someone prefers beat pulling up raw past interactions. Like PReF's few mixing weights, this suggests a short summary of you is more useful than a long record.

Two findings complicate the idea that cheap elicitation solves the problem. First, getting the preference is not the hard part. Tests across 16 agent systems Why do LLM agents remember preferences but not act on them? show agents often remember a stated preference and still fail to act on it. Most failures came from misunderstanding how the preference applied, not from forgetting it. Second, in shopping, preferences may not exist before the conversation starts. Bilgihan and colleagues Do AI shopping agents reshape what consumers actually want? argue that AI agents that make purchases help shape what people want through repeated back-and-forth. In that case a few well-chosen questions don't just uncover your preferences. They may also change them.

The collection is thin here. PReF is its only direct study of active preference elicitation, and nothing in it tests active learning with real shoppers. The surprise from the surrounding material is this: asking fewer questions only saves effort if the answers reflect real preferences, the agent acts on them, and the asking itself doesn't change what the person wants.