Quadrant Vacancy Prompting: Targeting Blank Regions in AI-Labeled 2×2 Semantic Maps to Steer Large Language Models Beyond Typical Responses

Large language models (LLMs) tend to concentrate their outputs around semantically similar "typical" answers even when repeatedly prompted on the same theme—a phenomenon documented as post-alignment mode collapse and output homogenization. Neither repeated chat sampling nor reasoning-enhancement techniques such as Chain-of-Thought reliably expand the breadth of ideation. This paper proposes Quadrant Vacancy Prompting (QVP), a practical method that (1) has an LLM generate a large set (tens to hundreds) of answers to a theme, (2) has the LLM itself label two evaluation axes and map all answers onto the resulting 2×2 plane, (3) repeatedly re-maps the answers under newly generated axis pairs, (4) identifies vacancies—blank regions ranging from entire quadrants to smaller local gaps—and (5) verbalizes the coordinates of a vacancy as explicit constraints in a new generation prompt, forcing the model to produce answers that fill the gap. We show that QVP is structurally analogous to filling unreached cells in Quality-Diversity search, to vacancy analysis in keyword-based patent maps, and to Boden's transformational creativity, and we argue that conditioning on vacancy coordinates operates as a forced shift of the generation distribution from its mode toward its tails. We illustrate the method with a case study conducted on ProAI Studio, a card-based ideation application (theme: new services to reduce food waste), in which all ten vacancy-targeted generations landed in the designated blank quadrant and both an opportunity-type ("blue ocean") and a worthless ("dead ocean") vacancy emerged within a single coordinate system. We further present an evaluation protocol for systematic empirical validation. This record contains two files: the English version (Furuta_2026_QVP_en.pdf) and the Japanese version with full appendices of all generated outputs (Furuta_2026_QVP_ja.pdf).