Rethinking Design Fixation in LLMs: A Process-Based Perspective Toward Linguistic Fixation

Large Language Models (LLMs) are increasingly employed in design tasks, raising the need for robust evaluation methods. Existing studies focus on assessing LLM-generated outputs using human-centered design constructs, such as design fixation, overlooking the underlying process through which LLMs generate text. This outcome-based approach neglect that design fixation may stem from the LLMs generation process, and that the "creativity" of LLMs is inherently (upper) bounded by training data. This study investigates design fixation at the process level by examining its origins within LLM architecture. We perform a clustering analysis of hidden states — high-dimensional vectors that LLMs use to predict the next token. Results show that hidden states are primarily organized by linguistic features, such as part-of-speech categories, rather than by design-oriented categories. This suggests that design fixation may emerge as a text-level property attributable to human evaluation of LLM outputs, rather than as a characteristic of the LLM's generative process.