Effectiveness of Human-Designed Creativity Interventions for LLMs

Determine whether creativity interventions developed for humans to help innovators overcome fixation and unknown unknowns are effective when applied to large language models that have fundamentally different capacities for storing, retrieving, and recombining knowledge.

Background

The paper investigates cross-domain mapping as a creativity intervention for both humans and LLMs. Humans often struggle with fixation and the "unknown unknowns" problem, whereas LLMs are trained on broad, cross-disciplinary corpora that might enable wider associative leaps. This raises the broader issue of whether creativity interventions designed to address human cognitive limitations translate effectively to LLMs, whose knowledge access and recombination mechanisms differ substantially.

The authors explicitly note that, despite progress in human creativity research on scalable methods to overcome fixation, there is an unresolved question about whether such human-oriented interventions will also be effective for systems with different knowledge architectures such as LLMs.