Do LLMs Recognize Your Latent Preferences? A Benchmark for Latent Information Discovery in Personalized Interaction

TL;DR

This paper tackles latent information discovery in personalized LLM interactions by introducing a unified tri-agent benchmark spanning 20 Questions, Personalized Question Answering, and Personalized Text Summarization under a passive-user setting. It formalizes a turn-based elicitation framework with a Judge to assess alignment, and evaluates multiple models to reveal how effectively latent user preferences can be uncovered and applied. Key findings show that LLMs can surface latent information, but performance is highly task- and context-dependent (32%–98%), with common failure modes including preference reinforcement and dilution across turns. The work establishes a first systematic framework for studying latent information discovery in personalized interaction and highlights core challenges for building truly adaptive, user-centered AI systems.

Abstract

Large Language Models (LLMs) excel at producing broadly relevant text, but this generality becomes a limitation when user-specific preferences are required, such as recommending restaurants or planning travel. In these scenarios, users rarely articulate every preference explicitly; instead, much of what they care about remains latent, waiting to be inferred. This raises a fundamental question: Can LLMs uncover and reason about such latent information through conversation? We address this problem by introducing a unified benchmark for evaluating latent information discovery - the ability of LLMs to reveal and utilize hidden user attributes through multi-turn interaction. The benchmark spans three progressively realistic settings: the classic 20 Questions game, Personalized Question Answering, and Personalized Text Summarization. All tasks share a tri-agent framework (User, Assistant, Judge) enabling turn-level evaluation of elicitation and adaptation. Our results reveal that while LLMs can indeed surface latent information through dialogue, their success varies dramatically with context: from 32% to 98%, depending on task complexity, topic, and number of hidden attributes. This benchmark provides the first systematic framework for studying latent information discovery in personalized interaction, highlighting that effective preference inference remains an open frontier for building truly adaptive AI systems.

Figures (9)

• Figure 1: Illustration of a question answering task. Strategy A: The LLM directly provides a response. This response is generic and violates the user's preferences. Strategy B: The LLM tries to elicit the user preferences through a multi-turn conversation, requesting specific information at each turn. At the end of the interaction, the LLM produces a personalized response.
• Figure 2: The benchmark’s general setting, including the example of Fig. \ref{['fig:intro_ex']}. The boxes with a solid border depict the generic elements of the setting, while the boxes with a dashed border describe the example. Both the multi-turn conversation between the Assistant and the User, and the evaluation phase (where the Judge is utilized) after each turn are depicted.
• Figure 3: The two personalization tasks and the 20 Questions Game task of our benchmark. The LLM tries to elicit the user preferences (or the hidden object) through a multi-turn conversation, requesting specific information at each turn. After each turn, the Assistant produces a personalized response to the User's task (here we only illustrate the response after the final turn).
• Figure 4: Results of the "20 Questions Game" experiments across different models, topics, and conditions; Topic Known ↑ (the upper part of each figure) vs Topic Unknown ↓ (the lower part of each figure).
• Figure 5: Results of the Personalized Question Answering experiments across different models, topics, and preference settings. The figure shows Success Rate (y-axis) vs. Average Stop Turn (x-axis) for #Preferences $\in\{1,2,3\}$ (three panels). Color encodes topic/domain and marker shape encodes model; error bars reflect reported uncertainties on both axes. (i) Domain effects are prominent: clusters by topic are clearly separated; (ii) As the number of preferences increases, points generally drift toward the lower-right (lower success, higher turns), indicating that deeper personalization settings tend to be harder.