Training artificial intelligence to master collaborative games presents significant challenges, especially when success relies on understanding and predicting the mental states of other players, a capacity often referred to as Theory of Mind (ToM). One such game is Yokai, a cooperative card game in which players must coordinate to achieve the highest score with limited communication. This thesis first proposes Yokai as a benchmark for large language models (LLMs) and then explores the ability of LLMs to play Yokai, emphasizing the understanding of the model for Theory of Mind by analyzing the inner state representation through evaluation techniques such as probing. Our key contributions are first, the proposal of Yokai as a text-based ToM benchmark for LLMs, second, the design of information-efficient inputs to represent game states, facilitating effective training and gameplay with LLMs and third, the use of probing on LLMs in dynamic environments. Performance benchmarking is conducted to evaluate self-play capabilities within the Yokai framework. This work provides new insights into LLM behavior and representation, contributing to theoretical and practical advances in AI research.