Serendipity by Design: Evaluating the Impact of Cross-domain Mappings on Human and LLM Creativity

1. Problem Statement

The paper investigates a fundamental question in the intersection of cognitive science and artificial intelligence: Do Large Language Models (LLMs) exhibit creativity in the same manner as humans, and do the same interventions that boost human creativity apply to LLMs?

Specifically, the authors examine cross-domain mapping—the cognitive process of drawing analogies from a remote, unrelated source domain to solve a problem in a target domain. While this technique is a known driver of human creativity (e.g., Velcro inspired by burrs), it is unclear if it functions as a reliable intervention for LLMs, which possess fundamentally different knowledge architectures (massive, cross-disciplinary corpora vs. human semantic memory with potential fixation). The study seeks to determine:

• Does forcing a cross-domain mapping increase the originality of ideas for both humans and LLMs?
• How do the two systems compare in baseline originality?
• What characteristics of the source-target pairing (e.g., semantic distance) make a mapping generative?

2. Methodology

Experimental Design

The study employed a controlled experimental design involving 10 daily consumer products (e.g., smartphone, backpack, sofa) as targets. Participants (humans and LLMs) were assigned to one of two conditions to generate novel product features:

1. Cross-Domain Mapping Condition: Participants were forced to translate a property from a randomly assigned, remote source domain (e.g., octopus, cactus, GPS, tornado) to the target product.
2. User-Need Condition: Participants proposed innovations based on identifying unmet user needs without an external source domain.

Participants and Models

• Human Participants: $N=140$ English-speaking adults recruited via Prolific. Each completed 10 design challenges.
• LLMs: Seven state-of-the-art models were tested: o3, gpt-4o, claude-sonnet-4, deepseek-v3, llama-3.3-70b, Qwen2.5-72B, gemini-2.5-pro. Each model generated 10 iterations per target-product pair.
• Total Data: 700 human-generated ideas and 700 LLM-generated ideas per condition (Total $N=2800$ ideas).

Evaluation Protocol

• Normalization: To control for stylistic differences, all ideas were rephrased by a single LLM (o3) into short, 1–2 sentence product descriptions, stripping away the explicit analogy framing.
• Human Raters: $N=1002$ independent raters evaluated the ideas on a 5-point Likert scale across four dimensions:
  1. Originality (Primary metric: novelty/unexpectedness).
  2. Feasibility (Plausibility with current tech).
  3. Usefulness (Real-world benefit).
  4. Investment Worthiness (Overall viability).
• Statistical Analysis: Linear mixed-effects models were used to predict ratings based on source (Human vs. LLM), condition, and their interaction, with random intercepts for participants and ideas.

Semantic Distance Measurement

To quantify the "remoteness" of a mapping, the authors calculated the semantic distance between target and source concepts using Wikipedia embeddings:

• Retrieved up to 50 sentences per concept.
• Embedded sentences using all-mpnet-base-v2.
• Applied dimensionality reduction and soft quantization to create categorical distributions over a unit sphere.
• Computed Jensen–Shannon divergence between the distributions of the target and source concepts to derive a semantic distance metric.

3. Key Results

A. Baseline Originality: LLMs Outperform Humans

• Across both conditions, LLMs generated ideas rated as significantly more original than humans on average.
• Model Variance: o3 and Claude-Sonnet-4 achieved the highest originality scores ($M \approx 3.8$), significantly outperforming humans ($M \approx 2.9$) and other models. Llama-3.3-70b was the only model that did not reliably outperform humans.
• Human Ceiling: Even in the "User-Need" condition (no external prompt), LLMs were rated as more original than humans.

B. Impact of Cross-Domain Mapping Intervention

• Humans: Cross-domain mapping significantly increased the originality of human ideas compared to the User-Need condition ($b=0.48, p<.001$).
• LLMs: The intervention did not yield a statistically significant increase in originality for LLMs. LLMs maintained high originality regardless of whether they were prompted with a remote source or a user need.
• Interaction: There was a significant interaction effect, indicating that the "boost" from cross-domain mapping is specific to human cognitive constraints (fixation), whereas LLMs already access remote associations naturally.

C. The Role of Semantic Distance

• Universal Driver: For both humans and LLMs, ideas derived from semantically distant source-target pairs received higher originality ratings.
• Correlation: The correlation between human and LLM ratings for specific target-source pairs was positive ($r=0.44$), suggesting that certain combinations (e.g., Octopus + Car) are inherently more generative than others (e.g., Sneaker + Sofa).
• Quantitative Effect: As semantic distance increased (measured via Wikipedia embeddings), originality ratings increased ($\beta_{human}=0.1, \beta_{LLM}=0.18$).

D. Trade-offs

• Originality vs. Feasibility: A strong negative correlation ($r = -0.74$) exists between originality and feasibility. Highly original ideas are often perceived as less feasible.
• Originality vs. Usefulness: A weak negative correlation ($r = -0.08$) suggests that originality does not necessarily imply a lack of utility, though the most radical ideas often struggle with perceived feasibility.

4. Key Contributions

1. Empirical Comparison of Creativity: Provides the first large-scale, controlled comparison of how cross-domain mapping interventions affect human vs. LLM creativity, revealing a fundamental asymmetry in sensitivity to prompting.
2. Quantification of Semantic Distance: Introduces a rigorous, embedding-based metric to quantify the "remoteness" of analogies, demonstrating that distance is a universal predictor of originality across biological and artificial systems.
3. Identification of Cognitive Bottlenecks: Confirms that cross-domain mapping is a necessary intervention to overcome human "fixation" (sticking to known solutions), whereas LLMs, trained on vast cross-disciplinary data, do not suffer from the same fixation and thus do not require the same external "nudge" to access remote associations.
4. Qualitative Differences in Analogy: Preliminary analysis suggests humans tend to map perceptual/sensorimotor features (e.g., cactus spines = thorny texture), while LLMs tend to map functional/mechanistic structures (e.g., cactus = water harvesting mechanism).

5. Significance and Implications

• For AI Development: The results suggest that current LLMs possess a "latent" capacity for remote association that rivals or exceeds human creativity in terms of raw originality. However, they may lack the human ability to generate extremely high-creativity outliers when explicitly guided, or they may simply be "saturated" with creative potential even without prompts.
• For Human-AI Collaboration: Interventions designed to boost human creativity (like random word prompts) may be redundant or ineffective for LLMs. Future human-AI creative workflows should focus on iterative refinement and adversarial critique rather than just initial idea generation, as LLMs can already generate diverse, high-originality concepts.
• Theoretical Insight: The study challenges the notion that creativity is a monolithic trait. It highlights that while the outcome (original ideas) can be similar, the mechanisms differ: humans rely on external scaffolding to break cognitive fixation, while LLMs rely on the statistical breadth of their training data to naturally traverse semantic space.

Conclusion:
The paper concludes that while cross-domain mapping is a powerful tool for unlocking human creativity, it is not a necessary condition for LLM creativity. However, the semantic distance between concepts remains the critical factor for high originality in both systems. This suggests that the future of creative AI lies not in forcing analogies, but in leveraging the vast, pre-existing associative networks of LLMs while guiding them toward the most distant and generative conceptual pairs.