The embedding of vector spaces, which numerically represent textual meaning, is a foundational technology in natural language processing (NLP). In the educational domain, difficulty scales are considered an essential aspect of meaning and are presumed to manifest as axes within these spaces. Although various difficulty scales, such as linguistic complexity and scientific content difficulty, exist, how these dimensions are embedded within vector spaces remains poorly understood. Recent advancements in inverse embedding techniques, originally developed for security applications, allow the reconstruction (generation) of textual examples from vector representations, even when the input vectors do not correspond to explicitly observed training data. In this study, we adapted the inverse embedding technology for educational applications and proposed a method for generating and evaluating textual examples corresponding to the candidate axes of cognitive difficulty in embedding spaces. By interpreting and assessing these difficulty-related axes using the generated examples, our approach provides insights into how different difficulty dimensions are encoded in vector spaces. This method has potential implications for personalized learning, automated readability assessments, and adaptive educational content generation.