How Context Injection Shapes Creativity in LLM-Generated Alternative Uses

Context injection influences large language model (LLM) outputs across many tasks, but its effect on divergent-ideation measures is underexplored. We systematically inject four types of context (object-focused, semantically related, unrelated, and random) into Alternative Uses Test (AUT) prompts and evaluate responses from four open-weight LLMs using both LLM-judged proxy metrics (creativity, novelty, value) and automated text metrics, including clustering-based flexibility. The main finding is metric divergence: no single context type optimises all evaluated dimensions at once. Instead, different context types steer outputs toward distinct evaluated profiles; for instance, random context maximises judged novelty while unrelated context uniquely increases categorical flexibility. These results indicate that context type determines the direction of measured output change rather than providing a uniform creativity boost, offering practical guidance for matching prompting strategy to target outcome.