A personalized model that sees only five to twenty examples from a new user cannot learn that user from scratch. In practice it borrows a prior: a belief, learned offline from many other people, about what users tend to want. This works for typical users. It fails quietly for people whose tastes or opinions the training population underrepresents, and those are often the people personalization is meant to serve. We propose to give each few-shot personalization step a cheap, per-user check of whether the population prior is fighting that user's own data. When the check fires, the system should do two things: weaken the prior for that user, and change which questions it asks next, choosing questions that test the prior rather than ones that assume it.
The two-sentence version is this. Bayesian personalization systems treat the population prior as fixed. We measure, for each user after a handful of answers, how much the personalized prediction depends on that prior, and we use this signal to decide per user how much to trust the crowd and what to ask next. We test whether the signal predicts which users the prior will hurt, and whether acting on it improves accuracy for the worst-served groups without costing typical users.
Core Claim
Cold-Start Personalization via Bayesian Adaptive Questioning (CAPE) shows that a small Bayesian linear model with a population-learned prior beats an 8-billion-parameter reinforcement-learning questioner on PrefDisco: 80.8% preference alignment against 68.5%, with 3–5 times fewer questions. That result is the methodological base we reuse. CAPE treats its offline prior as ground truth, though. Whose Opinions Do Language Models Reflect? documents that opinion data systematically under-serve groups such as people aged 65 and over, widowed respondents, and Mormons, so a prior fit to that data is wrong in a predictable direction for predictable users. This motivates the proposal. On the theory side, Bayesian Active Meta-Learning under Prior Misspecification (sOED-ML) proves that under a misspecified prior, information-greedy questioning can produce negative actual information gain, meaning the system queries points that reinforce its false belief. It proposes switching to task-specific questioning when a credible lower bound turns negative. That paper was evaluated only on synthetic participants, and its switching margin follows a hand-set decay schedule.
Our claim is that the sensitivity diagnostic of Detecting and diagnosing prior and likelihood sensitivity with power-scaling can be computed per user, in closed form or near it, inside a few-shot personalization loop on real human data. That paper perturbs the prior by raising it to a power and measures how far the posterior moves, using a cumulative Jensen–Shannon distance with an actionable threshold of 0.05. We further claim the diagnostic carries information that global shrinkage tuning and robust heavy-tailed priors do not. Specifically, it identifies users whose posterior is prior-dominated and in conflict with the prior, and redirecting questions for those users improves worst-group accuracy. If heavy-tailed priors alone close the gap, the claim is false, and we say so below.
Research Program
The user model is deliberately small. Each user is a weight vector $w$ over fixed features from a frozen backbone. For aesthetics, these are penultimate-layer image features, following the frozen-backbone logic of Personalised Aesthetics with Residual Adapters. For opinions, they are question and answer embedding averages of the kind used by Group Preference Optimization. The population prior is a Gaussian $\mathcal{N}(\mu_{pop}, \Sigma_{pop})$ fit to training users.
For the Gaussian-likelihood case, power-scaling is exact and cheap: raising the prior to power $\alpha$ gives $\mathcal{N}(\mu_{pop}, \Sigma_{pop}/\alpha)$. The perturbed posterior is therefore available in closed form, and the importance-sampling step of Kallioinen et al. is only needed for non-conjugate likelihoods. Those are the Bradley–Terry ranking model on aesthetics and the multinomial choice model on opinions, where we use a Laplace approximation plus Pareto-smoothed importance sampling.
That paper notes a limitation: power-scaling probes only how spread out the prior is, not where it is centered. A minority user's problem is usually location. Their mean preference sits away from $\mu_{pop}$. We therefore add a location perturbation that tilts the prior mean toward the data-only estimate. The per-user conflict score $s$ is the posterior predictive distance under both perturbations, computed after each answer.
Two actions depend on $s$. The first is tempering. We pick a per-user prior strength $\alpha^\ast$ as the largest value at which the predictive distance from the data-only posterior stays below threshold; this replaces sOED-ML's heuristic margin. The second is conflict-aware questioning. When $s$ exceeds threshold, the acquisition rule stops maximizing information about $w$ along directions where the prior is confident. It instead maximizes information about the deviation $w - \mu_{pop}$ along the directions driving the conflict. This is the operational analogue of sOED-ML's switch to task-specific gain, but defined through the measured predictive sensitivity rather than nested Monte Carlo over transferable parameters. That makes it cheap enough to run per turn.
The program has three stages. First, we validate the diagnostic as a predictor using passive few-shot data. Second, we test tempering. Third, we test the active questioning policy on PrefDisco, where CAPE provides the matching non-adaptive-prior system.
Baselines
The naive reference point is the population model with no personalization: $\mu_{pop}$ applied to everyone. Its counterpart is the same Bayesian head with a broad, nearly flat prior. These two bracket the problem. Any method that loses to the flat prior on minority users and to the population model on typical users is useless.
The strongest direct competitor is the fixed-population-prior Bayesian head, which is CAPE's approach. It is the system whose blind spot we target, and it is already strong on PrefDisco, where its reported alignment is 80.8% (CAPE). On PrefDisco we reimplement it with identical features, the same question budget, and the same offline training users. Beating it means higher alignment on users drawn from held-out profile clusters, at equal question counts, with no drop on in-distribution users.
The strongest tuned competitor is a global prior strength, a single $\alpha$ shared by all users and selected by cross-validation over training users. It gets the same hyperparameter search budget as our threshold.
The robustness competitors are a multivariate Student-t prior and a mixture-of-groups prior. The mixture is fit with one component per training demographic or rater cluster, with mixture weights inferred per user. Both are fit on the same training users with the same tuning budget.
The external-system baselines are the published leaders on each benchmark. On FLICKR-AES this is residual adapters, reported at Spearman $\rho = 0.639$ with all-layer adapters, and rising from 0.49 to 0.57 with ten images (Personalised Aesthetics with Residual Adapters). The ten-image figure is the relevant comparison point for our regime. On OpinionQA and GlobalOpinionQA the leader is GPO, reported as outperforming in-context fine-tuning by 7.1% and 8.4% on average (GPO). No absolute per-group score for GPO is reported on Lacuna, so we rerun it on our splits. On REAL-CUR, task-vector customization reports SROCC 0.621 (Task Vector Customization). We include it as a secondary aesthetics check.
Four ablations isolate the key ingredient:
- flagging the same fraction of users at random and tempering them, which tests whether the diagnostic's choice of users matters;
- tempering without changing questions;
- changing questions without tempering;
- dispersion-only power-scaling without the location perturbation, which tests our extension of Kallioinen et al.
Why not just…?
Why not just tune one global prior strength by cross-validation? This is cheap, and it will capture most of the average gain. Our claim is about the tail. A single strength set to suit the majority stays too strong for users the population misrepresents. A strength set to suit the minority throws away useful shrinkage for everyone else. The test is whether per-user tempering improves worst-group error over the best global value, at matched average error. If the gap is under the run-to-run spread across seeds, the per-user machinery is not justified for passive personalization. The remaining contribution would then rest on the questioning experiments.
Why not just use a heavy-tailed or mixture prior and let the posterior handle conflict automatically? This is the most serious threat. A Student-t prior downweights itself when data disagree, and a mixture can route a minority user to a better component. Both are run as full competitors. Our expected advantage is narrow but checkable. A robust prior repairs the estimate eventually, but it does not change which questions are asked, so it may spend scarce turns confirming the wrong component. If a mixture prior paired with standard information-gain questioning matches our system on held-out groups at every question budget, we will report the diagnostic as redundant. We will then reframe it as a monitoring tool that flags users for human review, not as an improvement to personalization.
Why not just pick each user's prior scale by maximizing their own marginal likelihood? With five to ten observations, per-user empirical Bayes is high-variance. The promises and pitfalls of deep kernel learning shows that marginal-likelihood optimization with few constraints drifts toward degenerate solutions that fit the training points. We nonetheless include per-user marginal-likelihood scale selection as a baseline. The comparison is whether our threshold rule gives lower variance of $\alpha^\ast$ across bootstrap resamples of a user's few answers, and lower held-out error.
Why not just prompt an LLM with the user's profile? The file's evidence is consistent here. PrefDisco reports negative normalized alignment in 29.0% of model–task pairs under interactive discovery (PrefDisco), and prompting baselines reach 22–31% alignment against CAPE's 80.8%. We include one prompting baseline on OpinionQA for completeness but do not expect it to compete.
Proposed Experiments
OpinionQA and GlobalOpinionQA. Each demographic group is treated as a "user" to be personalized from $n \in \{5, 10, 20\}$ answered questions. We fit the population prior on all groups except a held-out set of groups that Santurkar et al. identify as poorly represented (65+, widowed, Mormon, and their GlobalOpinionQA analogues by country). We also run a random held-out split as a control. The metric is 1-Wasserstein distance between predicted and survey answer distributions on unseen questions, reported separately for held-out-underrepresented groups, random held-out groups, and the worst group.
FLICKR-AES. We use the benchmark's 210 individual raters. Training raters supply the prior. Test raters are personalized with 5 to 20 images and scored by Spearman correlation. We stratify raters by distance of their full-data weight vector from $\mu_{pop}$, so atypical raters can be examined directly.
PrefDisco. The active-questioning test follows CAPE's four domains (MedQA, AIME, CSQA, SocialIQA). We fit the offline prior on profile clusters that exclude some trait combinations, then measure alignment and objective task accuracy on excluded profiles at question budgets of 1 to 5.
The first decisive check is whether the conflict score is a real predictor. At $n = 5$, does $s$ predict, with area under the ROC curve clearly above 0.5, which users end up with worse held-out error under the population prior than under the flat prior at full data? If not, the rest of the program stops.
The second decisive check is worst-group gain. On held-out underrepresented groups, per-user tempering must beat the cross-validated global strength and the mixture prior. On typical users it must not lose more than seed-level noise. The random-flagging ablation must do measurably worse than diagnostic flagging at the same flag rate.
The third check is the questioning effect on PrefDisco. Conflict-aware questions should reach CAPE's in-distribution alignment on excluded profiles with fewer questions than CAPE and fewer than mixture-prior information-gain questioning. We also report whether objective accuracy on AIME degrades less than the 12.1% loss PrefDisco documents for frontier models under personalization.
Finally, we report the diagnostic's latency per turn on CPU, to confirm it fits the interactive budget that makes small Bayesian heads attractive in the first place.
References
- Cold-Start Personalization via Bayesian Adaptive Questioning
- Whose Opinions Do Language Models Reflect?
- Bayesian Active Meta-Learning under Prior Misspecification
- Detecting and diagnosing prior and likelihood sensitivity with power-scaling
- Personalised Aesthetics with Residual Adapters
- Group Preference Optimization: Few-Shot Alignment of Large Language Models
- Scaling Up Personalized Image Aesthetic Assessment via Task Vector Customization
- The promises and pitfalls of deep kernel learning
- PrefDisco: Benchmarking Proactive Personalized Reasoning