This work investigates the performance of large language models (LLMs) in a deductive reasoning task modeled after the “20 Questions” game. In this task, two LLMs, designated as the “guesser” and the “answerer,” collaborate to identify a secret word through a series of binary yes-or-no interactions. The guesser is responsible for formulating strategic queries aimed at narrowing down possible answers, while the answerer provides accurate responses informed by the secret word. This study evaluates the LLMs’ ability to perform deductive reasoning, gather relevant information, and make decisions in a constrained environment. Our results reveal the models’ proficiency in generating targeted, insightful questions that effectively reduce uncertainty, underscoring their potential in complex problem-solving scenarios.