Adaptive Querying with AI Persona Priors

2026-05-04 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

This paper introduces "Adaptive Querying with AI Persona Priors," a novel framework for efficiently learning user-dependent quantities of interest, such as responses to unasked items or psychometric indicators, within strict question budgets. The method addresses limitations of classical Bayesian design and computerized adaptive testing (CAT), which often rely on restrictive parametric assumptions or computationally expensive posterior approximations. The core innovation is a persona-induced latent variable model where a user's state is represented by membership in a finite dictionary of AI personas. Each persona provides response distributions generated by a large language model (LLM) offline. This approach yields expressive priors with closed-form posterior updates and efficient finite-mixture predictions, enabling scalable Bayesian experimental design for sequential item selection. Experiments on synthetic data and the WorldValuesBench dataset demonstrate that persona-based posteriors provide accurate probabilistic predictions and an interpretable adaptive elicitation pipeline, outperforming CAT baselines, especially in cold-start scenarios.

Key takeaway

For Machine Learning Engineers developing interactive systems with tight query budgets, consider integrating AI persona priors. This approach offers a scalable and expressive alternative to traditional CAT, particularly beneficial in cold-start scenarios where historical data is scarce. Your team should leverage LLMs to pre-generate persona-specific response distributions, enabling efficient, closed-form Bayesian updates and robust adaptive querying. Be mindful that while greedy adaptive querying excels at small budgets, non-adaptive designs might offer more robustness against model misspecification at larger query budgets.

Key insights

AI personas from LLMs enable efficient, expressive Bayesian adaptive querying with closed-form posterior updates.

Principles

• LLM-generated personas can form expressive Bayesian priors.
• Closed-form posterior updates enhance scalability.
• Greedy adaptive querying balances efficiency and accuracy.

Method

The method involves generating persona-question response distributions offline via LLMs, then online, updating a prior over persona membership sequentially as user answers arrive, enabling efficient Bayesian experimental design.

In practice

• Use GPT-5-mini for persona-conditioned response elicitation.
• Employ the Twin-2K-500 persona bank for diverse profiles.
• Prioritize direct distribution elicitation over point predictions.

Topics

• Adaptive Querying
• AI Personas
• Large Language Models
• Bayesian Experimental Design
• Computerized Adaptive Testing

Code references

• yw3453/adaptive-query-ai-persona-priors

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Adaptive Interviewing for Persona Simulation in LLMs: Evidence-Grounded Reasoning Improves Decision Alignment — Adaptive interviewing improves LLM decision simulation by selectively grounding predictions in user-specific, follow-up derived evidence.
• Understanding the Behaviors of Environment-aware Information Retrieval — LLMs can learn to adapt query strategies for diverse retrievers using RL, revealing distinct optimal query styles.
• A Persona Dialogue Dataset of Lesser-Known Characters for Fairer Evaluation of Role-Playing LLMs — Evaluating role-playing LLMs with lesser-known characters can reveal biases and improve fairness.
• The persona selection model — LLMs simulate diverse personas during pre-training, with post-training refining a specific "Assistant" character.
• I Fine-Tuned an LLM to Reply Like My Wife — 10 Years of Chats, One Adapter, and Results That Amazed… — Fine-tuning LLMs with personal, messy data on consumer hardware can yield surprising persona learning.
• What’s the Best Way to Brainwash an LLM? — First-person statements are most effective for deeply embedding a persona in LLMs via SFT, enabling better generalization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.