Title: IDEAFix: A Framework for Assessing Creative Defixation in Large Language Models

Abstract:

As large language models (LLMs) become more prevalent in creative problem-solving and idea generation, their actual creative potential remains a subject of debate. While some research highlights their superior performance relative to humans, other studies point to structural constraints, such as fixation and the tendency toward output homogenization. Current evaluation methods often fall short; they either utilize narrow, decontextualized tasks that fail to capture goal-oriented generation or employ broader settings where multiple variables of the creative process are confounded. This complexity makes it difficult to isolate the specific impacts of task formulation, prompting techniques, and evaluation design. Furthermore, the influence of structured prompting strategies on idea generation has received limited attention.

To address these gaps, we present IDEAFix, an evaluation framework designed to analyze divergent thinking within open-ended idea generation contexts. Our approach involves prompting models to produce multiple original solutions to controlled variations of short design scenarios, adjusting for task attributes and defixation prompting strategies. This methodology allows for a systematic examination of how structured guidance shapes LLM output.

Our findings indicate that both the formulation of the task and the selection of attributes play significant roles in model performance, with simple prompting strategies capable of enhancing the originality of the generated solutions. Despite these improvements, we observed persistent homogenization of outputs across different models, underscoring inherent limitations in their capacity to produce truly diverse ideas. Ultimately, IDEAFix offers a controlled and extensible platform for investigating the mechanisms that drive creativity in large language models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC