What Do Llamas Really Think? Revealing Preference Biases in Language Model Representations