LLMike: Exploring Large Language Models’ Abilities in Wheel of Fortune Riddles

A riddle from the game show “Wheel of Fortune” consists of a hidden sentence that can be discovered starting from a simple clue and by iteratively guessing its letters. Although the game is very popular and intuitive, solving one of these riddles is not trivial. In fact, for interpreting the clue, identifying the most probable letters, and leveraging the game’s mechanics effectively, a player requires linguistic abilities, world knowledge, and even some form of strategic thinking. The goal of this study is to verify whether Large Language Models (LLMs) possess the necessary abilities to solve Wheel of Fortune riddles. We propose a software framework called LLMike in which an algorithmic Game Master interacts with an LLM: prompting it, enforcing the game’s rules, updating the hidden sentence based on the model’s guesses, and evaluating their correctness. We study several models with different sizes, evaluating their performance, behavioural patterns, and common types of errors. Our dataset and code are available at https://github.com/ejdisgjinika/LLMike.