Evaluating Large Language Models in a Complex Hidden Role Game