Reproduce “Fluid Transformers and Creative Analogies: Exploring Large Language Models' Capacity for Augmenting Cross-Domain Analogical Creativity” on OpenResearch

About this paper

Zijian Ding, Arvind Srinivasan, Stephen MacNeil, Joel Chan

This study demonstrates that Large Language Models can augment human cross-domain analogical creativity, showing that carefully prompted LLMs generate analogies useful for problem reformulation and ideation. While effective in assisting creative tasks, the research also identified that a notable portion of generated analogies contained potentially harmful or upsetting content.

Key insights

• Prompt engineering, particularly one-shot learning, significantly improved the semantic distance and potential usefulness of LLM-generated analogies, yielding unique cross-domain suggestions.
• The actual helpfulness of an analogy in a creative task is not reliably predictable through a priori expert judgments, emerging instead from user interaction and application.
• LLM-generated analogies can effectively broaden human thinking, leading to the addition of new concepts and reformulations of creative problems, with a median helpfulness rating of 4 out of 5 from users.
• A considerable percentage (up to 26%) of generated analogies contained potentially harmful content, primarily upsetting descriptions, highlighting a critical ethical concern for LLM deployment in creative support tools.

Summary by alphaXiv.