Adaptive Querying with AI Persona Priors
| Authors | Kaizheng Wang et al. |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2605.00696 |
| Download | |
| Categories | stat.ML, cs.CL, cs.LG |
Abstract
We study adaptive querying for learning user-dependent quantities of interest, such as responses to held-out items and psychometric indicators, within tight question budgets. Classical Bayesian design and computerized adaptive testing typically rely on restrictive parametric assumptions or expensive posterior approximations, limiting their use in heterogeneous, high-dimensional, and cold-start settings. We introduce a persona-induced latent variable model that represents a user's state through membership in a finite dictionary of AI personas, each offering response distributions produced by a large language model. This yields expressive priors with closed-form posterior updates and efficient finite-mixture predictions, enabling scalable Bayesian design for sequential item selection. Experiments on synthetic data and WorldValuesBench demonstrate that persona-based posteriors deliver accurate probabilistic predictions and an interpretable adaptive elicitation pipeline.
Engineering Breakdown
Plain English
This paper tackles the problem of learning user-specific preferences and responses with a limited budget of questions—a critical constraint in real applications like personalized testing or recommendation systems. The authors introduce a method that uses a finite set of "AI personas" (response distributions from language models) to represent user diversity, enabling efficient sequential question selection through Bayesian experimental design. Instead of fitting expensive parametric models or running complex posterior approximations, their approach leverages closed-form updates and mixture-model predictions to scale across heterogeneous, high-dimensional, and cold-start settings where classical adaptive testing methods fail. The core innovation is treating the user's latent state as membership in a persona dictionary, which makes both inference and active query selection tractable and efficient.
Core Technical Contribution
The paper's key novelty is the persona-induced latent variable model: representing users as draws from a finite mixture of AI-generated response distributions rather than learning individual parameters or relying on restrictive parametric families. This sidesteps the classical tension between flexibility and tractability—by anchoring to LLM-produced personas, the model gains expressiveness without requiring expensive posterior approximations or custom variational inference. The technical breakthrough is enabling closed-form Bayesian posterior updates in this mixture setting, which permits scalable sequential item selection (adaptive querying) even with thousands of users and high-dimensional item spaces. This design patterns shift from "fit a model to each user" to "identify which of a few learned personas best explains this user's behavior," making cold-start and high-dimensional settings tractable.
How It Works
The system begins with a dictionary of K AI personas, each defined by a large language model that produces a response distribution over items (e.g., ratings, yes/no answers to psychometric questions). When a new user arrives, the algorithm maintains a posterior over which persona(s) the user belongs to, initialized uniformly (cold-start). At each step, the algorithm selects the next item to query by using Bayesian experimental design: it picks the item that maximizes expected information gain about the user's persona membership given the current posterior. After observing the user's response, the posterior is updated in closed form using Bayes' rule on the finite mixture—this avoids expensive MCMC or variational approximations. The system repeats this query-update cycle until the question budget is exhausted, at which point it makes predictions about held-out items by marginalizing over the posterior distribution of persona memberships. The mixture structure enables efficient computation: with K personas, each update is O(K) rather than O(D) where D is dimensionality.
Production Impact
In production, this approach directly reduces the cost of user profiling: instead of asking 50 questions to characterize a user (as in classical adaptive testing), you might achieve similar accuracy with 10–15 queries by using information-theoretic query selection. This cuts operational cost in recommendation systems, user research platforms, and diagnostic applications where each question has real expense (user fatigue, API calls, expert annotation). Integration is straightforward: you pre-compute a persona dictionary once (using an LLM or offline learning), then for each new user, maintain lightweight mixture-model updates on the application server. The trade-off is upfront work building high-quality personas—they must cover the actual diversity in your user population, and personas trained on one domain may not transfer. Latency is sub-millisecond for query selection (matrix operations on K personas) and inference, making it suitable for real-time systems; compute cost is dominated by initial LLM calls to generate persona dictionaries, not by per-user inference.
Limitations and When Not to Use This
The approach assumes users can be well-represented by a finite mixture of fixed personas—it will fail if user behavior is truly idiosyncratic or requires continuous parameter learning. The quality and diversity of personas are critical but not formally specified: the paper relies on LLM-generated distributions, which may have biases or gaps relative to true user behavior, and adapting or expanding the persona dictionary after deployment is not addressed. The method assumes the response model (how personas generate answers to items) is correctly specified; misspecification in the LLM's response distribution would corrupt posterior updates and subsequent predictions. Finally, the paper does not address multi-step feedback loops (e.g., how user behavior changes over time or adapts to previous interactions), treating the user's true persona as static, which is unrealistic in long-term personalization scenarios.
Research Context
This work extends the classical literature on computerized adaptive testing (CAT) and Bayesian experimental design by replacing restrictive parametric models (e.g., item response theory) with expressive LLM-based priors. It builds on recent progress in using language models as prior distributions and marries that with efficient sequential decision-making, addressing a known gap: most adaptive testing assumes Gaussian or logistic models, but real user heterogeneity is often high-dimensional and non-parametric. The paper contributes to the broader shift toward persona-based user modeling (seen in recommendation systems and NLP) by formalizing it as a probabilistic inference problem with information-theoretic optimality. Future work likely explores dynamic persona adaptation, transfer learning across domains, and extensions to settings where personas themselves are learned from data rather than pre-specified by LLMs.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::