Provably Efficient Personalized Multi-Objective Bandits with Proactive Conversational Queries

A new algorithm called MO-PQUCB aims to improve personalized recommendations by combining user queries with traditional feedback, according to a paper submitted to arXiv in 2026. The approach addresses a long-standing challenge in multi-objective decision-making systems where user preferences must be learned alongside unknown rewards. [1] The algorithm, formally introduced in a preprint titled "Provably Efficient Personalized Multi-Objective Bandits with Proactive Conversational Queries," tackles the problem of learning user-specific trade-offs among competing objectives. In standard multi-objective multi-armed bandit (MO-MAB) problems, existing methods infer preferences only from utility feedback, which the authors argue entangles preference learning with reward exploration. [1] [2] MO-PQUCB integrates what the researchers call proactive conversational queries—structured signals where users reveal priorities through statements like "cheap and clean hotel"—directly into the learning process. These queries are modeled using a Plackett-Luce subset choice model. The paper demonstrates that relying solely on queries is insufficient due to a fundamental shift-invariance barrier, necessitating a hybrid design. [2] The algorithm's architecture combines query-based preference anchoring with bandit feedback through shift-invariant regularization and what the authors term dual-exploration UCB (Upper Confidence Bound). The researchers provide theoretical proofs that proactive queries accelerate preference estimation and yield improved regret scaling over prior preference-aware MO-MAB methods. [2] For scenarios where user-provided queries may be unreliable, the paper characterizes the statistical limits of learning under corrupted queries and introduces a robust estimator. This estimator achieves near-optimal performance when the corruption is sparse, according to the preprint. The authors report that experiments validate both the theoretical and practical gains of MO-PQUCB. [2] The paper was submitted on June 7, 2026, to arXiv, an open-access repository for electronic preprints that has been operating since 1991 and now receives approximately 24,000 submissions per month. [6] The work appears under the Computer Science > Machine Learning category and is accessible through arXiv's standard abstract page, which also features community-developed tools through the arXivLabs framework. [1] [5] arXivLabs, launched in 2020, provides a formalized structure for third-party collaborators to develop experimental features such as bibliographic explorers and recommender systems that appear alongside papers on the site. [5]