Can a few smart questions help an AI learn what you prefer?

The challenge of starting from scratch

Personalizing an AI assistant is difficult when it has no history with the person asking. There may be many possible preferences—such as tone, detail, rigor or examples—but any one person may care about only a few. Asking about every option could be tedious; asking without a plan could miss what matters.

The authors of Cold-Start Personalization via Bayesian Adaptive Questioning propose that preferences are not completely independent. Someone who likes detailed explanations, for instance, may also value worked examples. A system could learn such population-level associations in advance, then use a person’s answers to make informed guesses about preferences it has not asked about.

Learning from profiles, then asking questions

The proposed method, CAPE, separates the process into two stages. Offline, it learns a model of how preference criteria relate across complete profiles. Online, when meeting a new user, it updates its beliefs as answers arrive and selects follow-up questions intended to reveal useful information. It then predicts a fuller preference profile, including dimensions the user did not directly address, and passes that profile to a separate language model to shape its response.

The team tested CAPE on four reasoning benchmarks: medical questions, mathematics, commonsense reasoning and social reasoning. Each domain included 100 problems, with 50 simulated users per problem. The system could ask about five criteria, selected from 15 used by its belief model. The benchmark focused on task-specific preferences, and simulated users answered minimally rather than volunteering extra information.

CAPE’s main belief model was Bayesian linear regression, a relatively simple statistical approach. The authors compared it with several alternatives, including a reinforcement-learning system trained to improve its questioning policy using feedback on completed responses. All methods used the same downstream response generator, helping isolate the effects of preference elicitation and inference.

Better alignment in the benchmark

Across the four domains, CAPE reached 77–87% of the benchmark’s “oracle” alignment score—the score obtained using a complete preference profile. The paper reports an overall figure of 80.8%, compared with 68.5% for the reinforcement-learning baseline. CAPE also needed between 2.5 and more than 15 times fewer questions to reach equivalent alignment in the reported comparisons. In two domains, the RL baseline did not match CAPE’s single-question alignment even after 15 questions.

CAPE also changed its next question more often in response to different answers. Its reported adaptivity was 39–62%, compared with 0–28% for the RL baseline across the domains. In the authors’ analysis, the main gains came from the learned preference model: one answer could inform predictions about several unasked criteria. Adaptive question selection added a smaller benefit; in an ablation test, it saved about one question compared with random selection.

The results suggest that, in this experimental setup, explicitly learning patterns among preferences can be more useful than asking a system to learn both those patterns and a questioning strategy from sparse final feedback. They do not show that every Bayesian model or questioning rule will outperform reinforcement learning in other settings.

What remains uncertain

The evaluation used simulated users, not people interacting with a deployed assistant. The simulations followed an established benchmark protocol, but held user cooperation and verbosity fixed. Real people may answer inconsistently or change their minds, possibilities the study does not test. The work also does not address learning across sessions or helping users articulate preferences they have not yet considered.

CAPE relies on predefined criteria and preference profiles collected for offline learning. The authors identify natural-language questions and automatic discovery of preference dimensions as future work. They also warn that population data can encode social biases, and that inferred preferences can be wrong. They recommend giving users a way to correct personalization and making clear when preferences have been inferred.

The paper text identifies the work as appearing in the 2026 International Conference on Machine Learning proceedings, while the linked arXiv record is version 2 of a preprint. This account draws on the supplied paper text, not an independent assessment of publication history or a new replication. The reported gains should be understood as results within the paper’s benchmarks, simulated-user protocol and evaluation method—not proof that assistants will personalize reliably for everyone.