Serendipity by Design: Cross-Domain Mappings on Human and LLM Creativity

> Paper: Serendipity by Design: Evaluating Cross-domain Mappings on Human and LLM Creativity > Authors: Qiawen Ella Liu, Marina Dubova, et al. > arXiv: https://arxiv.org/abs/2603.19087

Background

Creativity research has long drawn on Sarnoff Mednick's (1962) Associative Hierarchy theory: highly creative people have a *flat* associative hierarchy, allowing them to jump to remote concept nodes rather than only the most obvious ones. Many famous inventions came from cross-domain analogies—the Wright brothers studying bird flight, George de Mestral inventing Velcro from burdock burrs.

Humans, however, suffer from design fixation: when improving a product, we anchor on its existing features. Cross-domain mapping—deliberately linking unrelated domains—acts as a forced nudge out of this comfort zone. The Princeton team asked: does this technique work equally well for humans and LLMs?

Experimental Design

• Participants: human crowdworkers and multiple mainstream LLMs (including GPT-4)
• Task: design new features for 10 everyday products (backpack, TV, bicycle, alarm clock, coffee cup, umbrella, desk, car, watch, refrigerator)
• Two prompt conditions:
1. *User-need*: open-ended design task 2. *Cross-domain mapping*: draw inspiration from a randomly assigned source (octopus, cactus, GPS, bees, spider webs, sunflowers, bats, coral reefs, dandelions, camels, shells, etc.)
• Evaluation: human raters scored originality, feasibility, and usefulness; semantic distance between source and target domains was quantified with Wikipedia-trained word embeddings

Key Findings

1. LLMs are natively more divergent

Without cross-domain prompts, LLM-generated ideas were on average more original than human ideas—suggesting LLMs already possess a flat associative hierarchy from their broad training data, while humans fall into design fixation.

2. Cross-domain mapping helps humans, not LLMs

• Humans: cross-domain prompts significantly increased originality.
• LLMs: no significant effect—the operation is already natural to them.

Possible explanations: LLMs have internalized cross-domain associations; humans need the external nudge; prompt effects may only appear for LLMs at extremely large semantic distances.

3. Semantic distance drives originality

For both humans and LLMs, the larger the semantic distance between source and target domains, the more original the ideas—supporting remote association as a core mechanism of creativity.

4. Different creativity distributions

• Humans: bimodal—many mundane ideas plus occasional brilliant leaps.
• LLMs: uniform—consistently decent, but rarely exceptional.

Implications

• For creativity training: cross-domain mapping is an empirically validated strategy for humans—random-word techniques, cross-disciplinary learning, analogy databases.
• For AI-assisted design: don't expect explicit prompts to "unlock" LLMs—they're already there. Semantic distance can serve as a predictive metric for inspiration quality.
• For human-AI collaboration: a promising model is humans doing the cross-domain leaps while AI elaborates and expands on those seed ideas, potentially yielding a 1+1>2 effect.

Conclusion

The title "Serendipity by Design" captures the paradox: serendipity (chance discovery) can be *designed*—systematically introduced cross-domain mappings create the happy accidents that breakthroughs often depend on. But the effect is asymmetric: humans need this nudge, while LLMs already inhabit a permanently flat associative space. The best creative system may be neither pure human nor pure AI, but humans opening the conceptual doors and AI filling in the details.

Further reading: Mednick (1962), *The Associative Basis of the Creative Process*; related concepts: remote association theory, structure mapping theory, semantic distance, design fixation.