Uncovering the Computational Ingredients of Human-Like Representations in LLMs