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

Abstract: Are LLMs creative in the same way humans are, and can the same interventions increase creativity in both? We evaluate a promising but largely untested intervention for creativity: forcing creators to draw an analogy from a random, remote source domain (''cross-domain mapping''). Human participants and LLMs generated novel features for ten daily products (e.g., backpack, TV) under two prompts: (i) cross-domain mapping, which required translating a property from a randomly assigned source (e.g., octopus, cactus, GPS), and (ii) user-need, which required proposing innovations targeting unmet user needs. We show that humans reliably benefit from randomly assigned cross-domain mappings, while LLMs, on average, generate more original ideas than humans and do not show a statistically significant effect of cross-domain mappings. However, in both systems, the impact of cross-domain mapping increases when the inspiration source becomes more semantically distant from the target. Our results highlight both the role of remote association in creative ideation and systematic differences in how humans and LLMs respond to the same intervention for creativity.

• The paper demonstrates that cross-domain mapping significantly elevates human originality while LLMs show only marginal creative gains.
• It utilizes a controlled generative task with rigorous evaluations using semantic distance metrics, transformer embeddings, and human ratings.
• Implications suggest that tailored prompt engineering based on semantic remoteness can optimize creative outputs and inform hybrid innovation strategies.

Evaluating Cross-domain Mapping as a Driver of Human and LLM Creativity

Introduction

This paper interrogates whether cross-domain analogical prompts—an intervention inspired by structure-mapping accounts of creativity—uniformly elevate creative output in humans and LLMs, or whether the computational substrates of these systems produce divergent responses. The authors systematically compare the efficacy of cross-domain mapping versus a user-need prompting baseline in both systems, leveraging a broad, carefully controlled generative task focused on product innovation. Their analysis emphasizes not only quantitative measures (originality, feasibility, usefulness, investment-worthiness), but also the systematic dependence of originality on semantic remoteness between conceptual domains.

Experimental Paradigm

The research employs a two-condition idea-generation protocol with both human participants and seven state-of-the-art LLMs (o3, gpt-4o, claude-sonnet-4, deepseek-v3, llama-3.3-70b, Qwen2.5-72B, gemini-2.5-pro). In the cross-domain mapping condition, actors are compelled to transfer a property from a randomly assigned, semantically distant inspiration source to a given consumer product. The user-need condition, by contrast, involves innovating to fulfill unmet user requirements. Generation prompts and normalization protocols are stringently harmonized between humans and LLMs, with all outputs rephrased via the o3 model to control for style and content presentation.

Subsequently, 1002 independent raters evaluate 2800 resulting ideas along four axes, with analytical focus on originality. The methodology integrates semantic distance computations via pretrained transformer embeddings and Jensen-Shannon divergence, providing a principled measure of conceptual remoteness and its effects on creative generation.

Key Quantitative Results

The results demonstrate robust divergences in how cross-domain mapping influences creativity across systems. While LLMs consistently yield more original ideas than humans under both prompting regimes, cross-domain mapping significantly boosts the originality of human-generated ideas but fails to produce a comparable effect in LLMs.

The joint originality-feasibility trade-off is also pronounced ($r = -0.74$), consistent across systems (Figure 1).

Figure 1: Cross-domain mapping increases originality for humans but produces weaker effects in LLMs; density contours highlight the trade-off between feasibility and originality.

The analysis of model-level effects reveals significant inter-LLM variability. The o3 and Claude-Sonnet-4 models outperform both humans and other LLMs in originality across tasks, but not all models—e.g., Llama-3.3-70b—reliably outperform human baselines. User-need prompts elicit less original generations, across both humans and LLMs, confirming the augmentation potential of guided cross-domain analogy for at least one system.

Figure 2: Distributions of originality scores segmented by model and prompting condition, indicating significant differences between top-performing LLMs and human responses.

Critically, the benefit of cross-domain mapping for humans manifests most strongly in the upper tail of originality, where the most original human ideas rival or exceed top LLM generations. However, in aggregate, LLMs’ baseline generative diversity—possibly a consequence of large-scale training across heterogeneous data—renders explicit analogical prompting dispensable for originality enhancement.

Cross-domain Pairings and Semantic Distance Effects

A salient contribution is the demonstration that the semantic distance between target and inspiration domains predicts resultant originality for both humans and LLMs. Target-inspiration pairs with higher semantic divergence elicit more highly rated novel ideas, implicating remote association as the primary driver of cross-domain creativity in both systems. Notably, the effect size is larger for LLMs ($\beta=0.18$), but remains significant for humans ($\beta=0.1$).

The correlation in originality ratings across identical cross-domain pairs between humans and LLMs (Pearson $r = 0.44$) further establishes that certain domain mappings are universally productive, irrespective of generative substrate. The analysis thus points to semantic remoteness, rather than explicit analogy-prompt framing, as the dominant predictor of originality in both system classes.

Figure 3: (a) Correlation in originality of identical target-inspiration pairs across systems; (b) systematic increase in originality as a function of Wikipedia-derived semantic distance; (c, d) originality heatmap across domain pairings for LLMs and humans.

Theoretical Implications and Model Computation

The results stress system-level architectural differences in creativity. Humans exhibit strong fixational biases—constrained retrieval within local semantic regions—such that cross-domain mapping interventions reliably “flatten” associative hierarchies and force more original generative outputs. LLMs, trained to maximize prediction over a vast multimodal corpus with cross-domain regularities, may operate over intrinsically “flatter” associative structures, minimizing the marginal impact of explicit analogical prompting.

Furthermore, LLMs and humans show qualitative differences in the mapping process: human analogical transfer often exploits salient perceptual or sensorimotor features, whereas LLMs gravitate towards mechanistic or functional alignments, likely reflecting distributional representations formed during pretraining.

Consistent with structure-mapping theory, explicit analogical interventions remain valuable for systems susceptible to domain-local fixation, but in architectures capable of wide, rapid, and perhaps “unbiased” recombination, the mechanisms for creative augmentation may require re-conceptualization.

Practical Implications and Future Directions in AI

Practically, these findings urge caution in directly translating human creativity interventions to LLM-based or hybrid human-AI systems without adaptation. Cross-domain mapping is highly effective for humans—particularly in surfacing highly novel concepts—but does not further enhance LLM originality, which is already augmented by construction. Nevertheless, manipulating semantic distance at the level of prompt engineering remains a viable strategy for maximizing creative yield in both settings.

From a methodological perspective, evaluating not only the quantity but the quality and domain of analogical transfer is indispensable; future research should operationalize frameworks for coding perceptual versus mechanistic mappings at scale. Incorporating iterative, adversarial, and multi-agent workflows—better reflecting real-world creative practice—may enable richer interaction between human and machine ideation, spurring new forms of collaborative innovation.

Conclusion

This study provides a rigorous comparative analysis of cross-domain mapping’s effects on creative output in humans and LLMs, revealing a clear asymmetry in sensitivity to analogical prompts. Humans benefit substantially from explicit cross-domain interventions, while LLMs exhibit high baseline originality with marginal further gain. However, semantic distance consistently amplifies originality in both systems, emphasizing remote association as a trans-systemic mechanism. These findings refine theoretical understanding of analogical creativity, highlight critical differences in associative dynamics between human and machine generation, and underscore the need for more nuanced, system-specific creative augmentation strategies.