Large Language Models(LLM) for Automating 20 Questions Game

Abstract

The 20 Questions game, a classical problem in artificial intelligence and natural language processing, involves one player thinking of an object while the other player asks yes/no questions to guess the object within 20 questions. Traditional approaches to this game have relied on decision trees, heuristic algorithms, and rule-based systems. However, the advent of Large Language Models (LLMs) such as GPT-3 and GPT-4 has opened new avenues for enhancing the performance and interactivity of AI players in this game. This paper explores the application of LLMs in playing the 20 Questions game, focusing on the model's ability to generate and interpret natural language questions and responses. We detail the architecture and training process of an LLM tailored for this task, highlighting modifications made to optimize the model's performance in understanding and narrowing down possible objects. Through extensive testing, we demonstrate that LLMs can achieve a higher success rate and engage in more coherent and contextually appropriate dialogues compared to traditional methods. Additionally, we analyze the model's performance across various object categories, noting the strengths and limitations of LLMs in dealing with different types of objects. Our results show that while LLMs excel in categories with well-defined and common attributes, they face challenges with abstract or less common objects. We propose several strategies for overcoming these limitations, including hybrid approaches that combine LLMs with specialized knowledge bases and reinforcement learning techniques. The findings of this study suggest that LLMs hold significant potential for not only improving the 20 Questions game but also advancing broader applications in conversational AI, interactive entertainment, and educational tools. Future work will focus on refining the model's questioning strategy, enhancing its knowledge representation, and exploring its application in other interactive scenarios.