This research introduces a novel approach for training LLMs to strategically elicit information about uncertain entities through natural language conversations.

• Develops adaptive elicitation algorithms that help LLMs refine their understanding through targeted questioning
• Creates models that can dynamically adjust questioning strategies based on previous responses
• Demonstrates effectiveness across multiple domains including education, healthcare, and user preference learning
• Achieves significant improvements in information gain compared to baseline methods

For education, this approach enables more personalized assessment of student knowledge, allowing for adaptive testing that efficiently identifies knowledge gaps and learning needs through conversational interaction rather than standardized questions.

Adaptive Elicitation of Latent Information Using Natural Language