PaperlayerView exact revision on arXiv

Source-linked AI summary

Scanvas: Discovering and Developing Synergistic Opportunities in Generative Design Spaces

Yaqing Yang, Aniket Kittur, Hongyu Howie Wang, Nikolas Martelaro, Matt Klenk, Yan-Ying Chen, Matthew K. Hong

arXiv:2609.34062v1cs.HC

TL;DR

Sparse design spaces make synergistic opportunities difficult to find, while existing ideation tools often favor additive approaches. Scanvas enriches seed ideas and searches for relationships that create synergy; evaluations suggest it can improve idea quality and help surface promising directions.

  • Problem

    Latent capabilities, issues, and goals create a rich but sparse design space, and synergistic opportunities can be difficult to enumerate and traverse.

    Source: §1 · p. 2

  • Method

    Scanvas decomposes and enriches seed ideas, then searches the enriched space for synergistic relationships and develops combinations.

    Source: §1 · p. 3, §3.1.1 · p. 6, §6.5 · p. 21

  • Results

    In a study with 12 professionals against an LLM-supported ideation baseline, Scanvas supported identifying and developing synergies; the approach can improve idea quality and surface promising directions.

    Source: §1 · p. 3, §6.5 · p. 22

  • Takeaways & Limitations

    Scanvas can help surface promising relationships and design directions that might otherwise be overlooked.

    Source: §6.5 · p. 22

    The evaluation captured only a short period of design activity.

    Source: §6.4 · p. 21

Abstract

from arXiv · show

Good design is often synergistic, creating super-additive value by linking goals so that existing resources produce greater outcomes. However, finding these synergistic opportunities in sparse design spaces is difficult, and current LLM-supported ideation tools largely default to additive paradigms such as feature blending, variant generation, or local patching. We present Scanvas, an AI-supported system for systematically discovering and developing synergistic design opportunities. Scanvas operationalizes synergy through a two-step computational process: first, it decomposes seed ideas into explicit properties (components, behaviors, surpluses, and issues) to enrich the design space; second, it systematically searches across enriched ideas using three theory-grounded strategy operators: unlocking or strengthening goals, turning weaknesses into resources, and sharing components across functions. We instantiate Scanvas as an auto-generation pipeline and an interactive system. Pipeline ablations and a user study with 12 professional designers demonstrate that Scanvas enables users to surface and develop significantly higher-quality, synergistic concepts compared to LLM ideation baselines.

1 Introduction

Synergistic design links goals so existing resources create additional value, but such opportunities are difficult to find in sparse design spaces. Scanvas develops ideas before searching for synergies and supports both automated and designer-guided exploration.

1 Introduction

  • Existing LLM ideation tools have rarely addressed sparse synergistic opportunities, instead emphasizing variants, parameter spaces, trade-offs, or idea blending.Latent capabilities, issues, and goals make the design space rich but sparse.Source: §1 · p. 2
  • Scanvas treats synergy as relationships that strengthen one another, including adding or linking goals and serving multiple goals with shared resources.These strategies aim to create additional value from the same or fewer resources.Source: §1 · p. 2
  • Scanvas decomposes seed ideas into components, behaviors, surpluses, and issues, then expands them before searching for synergies across original and enriched ideas.Search targets include unlocking goals, turning weaknesses into resources, and sharing components across functions.Source: §1 · p. 3
  • Scanvas is instantiated as an auto-generation pipeline and an interactive system for designers to add, expand, and combine ideas while potential synergies are surfaced.The pipeline directly expands seed ideas and searches the enriched space for combinations.Source: §1 · p. 3
  • A pipeline ablation found enrichment and synergistic search were important drivers of idea quality, while a study with 12 professionals suggested Scanvas supported identifying and developing synergies beyond feature accumulation.The study compared Scanvas with an LLM-supported ideation baseline.Source: §1 · p. 3

2 Related Work and Design Goal

The related work frames synergistic ideation as a sequence of challenges: develop ideas to expose opportunities, operationalize strategies for synergy, and search for and construct relationships across ideas.

2 Related Work and Design Goal

  • Prior work addresses idea development, synergy relationships, and computational discovery, motivating three design goals for computationally supporting synergistic ideation.The goals concern enriching ideas, operationalizing synergy strategies, and searching for and developing relationships across ideas.Source: §2 · p. 3

2.1 Developing Ideas to Prepare for Synergy

  • 2.1 Existing development methods produce better or more explored variants, but less systematically develop latent properties because they may later participate in synergy.Design Goal 1 is to enrich ideas so more components, resources, byproducts, and resolved issues become available for later discovery.Source: §2.1 · p. 4

2.2 Theoretical Strategies for Finding Synergistic Solutions

  • Many synergy theories remain high-level conceptual frameworks rather than concrete procedures directly applicable computationally, while existing systems often operationalize one design method; computational support for unifying multiple synergy forms is limited.These forms include discovering new goals, repurposing weaknesses as resources, and sharing components across functions.Source: §2.2 · p. 5

2.3 Finding and Developing Synergistic Combinations

  • 2.3 Prior systems select or construct combinations using relevance, novelty, distance, or coherence, but these criteria do not necessarily target synergy.The section's goal is to search enriched ideas for relationships that create value through interaction rather than merely adding existing functions.Source: §2.3 · pp. 5–6

3 System

Scanvas turns opportunities within individual ideas into materials for discovering synergies across ideas through a two-stage LLM-based pipeline.

3 System

  • Scanvas first enriches individual ideas by making behaviors, surplus capabilities, and issues explicit and expanding them into variants.This stage develops opportunities within individual ideas before searching across them.Source: §3 · p. 6
  • It then operationalizes synergy strategies to search the enriched idea space and automatically develop synergistic combinations.The system is instantiated as an end-to-end auto-generation pipeline and an interactive system.Source: §3 · p. 6

3.1 Generation Pipeline

The generation pipeline decomposes and expands seed ideas, searches the resulting idea pool for theory-grounded synergies, and ranks candidates for further exploration.

3.1 Generation Pipeline

  • Given a design problem and seed ideas, the pipeline enriches ideas, searches original and expanded ideas for synergies, and uses an LLM judge to rank candidates.Expansion and synergy discovery can be applied iteratively.Source: Figure 2 · p. 6, §3.1 · p. 6

3.1.1 Idea Enrichment.

  • Idea enrichment decomposes each idea into components, behaviors, surplus capabilities, and issues, then expands it through repurposing components, leveraging byproducts, or mitigating issues.These structured properties expose resources, functions, outputs, and unresolved problems for subsequent synergy discovery.Source: §3.1.1 · pp. 6–7

3.1.2 Synergy Discovery.

  • Synergy discovery searches for combinations that unlock goals, turn an issue or byproduct into another idea's input, or let a shared component serve multiple functions.Because applying these rules grows the candidate set combinatorially, a judging agent ranks the resulting synergies.Source: §3.1.2 · p. 7

3.2 Scanvas: Interactive System Walkthrough

Scanvas lets designers enrich an idea space, develop user-selected combinations, and search automatically for synergies. The walkthrough also shows tools for visualizing and organizing ideas.

3.2.1 Enrich individual ideas.

  • Idea cards expose properties such as issues and surplus capabilities, then support expanding ideas into variants that repurpose components, leverage byproducts, or mitigate issues.For example, sunlight is treated as a resource for a solar-collecting roof, while glare and heat motivate a ceramic-dot roof.Source: §3.2.1 · p. 8

3.2.2 Combining ideas selected by the user in a synergistic way.

  • Designers select ideas and opportunities for a mix, while Scanvas develops their relationship into a coherent concept with value beyond the individual ideas.The example combines a glass-roof display, ceramic-dot roof, and reclined seat into an e-ink roof with controllable dots.Source: §3.2.2 · pp. 8–9

3.2.3 Finding synergistic combinations automatically.

  • Users can ask Scanvas to find a complementary idea for a selected concept; it combines a breeze-and-nature-sound idea with a personal head-space dome as “Private Sky Domes.”The concept pairs each passenger’s raised overhead zone with localized airflow and directional nature sounds.Source: §3.2.3 · p. 9
  • Scanvas searches across canvas ideas for promising pairs, linking a breeze idea with glass-roof heat and noise issues to propose a ventilated double-layer roof.The shared structure addresses airflow distribution, solar heat, and acoustic comfort together.Source: §3.2.3 · p. 9

3.2.4 Organizing and browsing inspirations.

  • Scanvas visualizes ideas and uses text embeddings and UMAP to arrange idea cards by semantic similarity; users can also bookmark ideas.The interface includes automatic canvas organization and a bookmark feature for saving ideas.Source: Figure 3 · p. 8, §3.2.4 · p. 9

4 Evaluation Method

The evaluation combines an automated pipeline comparison with a within-subject study of the interactive system. Both assess idea quality, while the user study also examines participants’ experience and idea exploration.

4.2 Study 2: User Study of the Interactive System.

  • Study 2 is a within-subject comparison in which 12 professionals use Scanvas for one ideation task and the baseline for another.Task, condition, and presentation order are counterbalanced; each condition includes a 30-minute main task.Source: §4 · p. 9, §4.2 · p. 10, §4.2.1 · pp. 10–11

4.1 Study 1: Auto-Generation Evaluation.

  • Two engineering experts independently rate generated ideas on Goal Fulfillment, New Opportunities, Value vs. Cost, and Interestingness using a 7-point scale.The dimensions cover original-goal fit, additional value, resource costs, and promise for further exploration.Source: §4.1 · p. 10

4.2.2 Participants.

  • Participants were engineers or engineering PhD students with industry experience, averaging 3.3 years in their current roles.Seven participants were full-time industry engineers and five were engineering PhD students.Source: §4.2.2 · p. 11

4.2.3 Baseline.

  • The baseline retains idea expansion and user-selected mixing but omits automatic synergy discovery, using zero-shot prompts for variants and combinations.Both conditions use GPT-5.5 as the underlying language model.Source: §4.2.3 · p. 11

4.2.4 Evaluation.

  • The user study measures experience through surveys and interviews, and idea quality through blind expert ratings on the four dimensions.Semantic diversity is calculated as mean pairwise cosine distance between idea embeddings, averaged across two tasks.Source: §4.2.4 · pp. 11–12
  • Rater agreement was good: ICC values were .84 for Goal Fulfillment, .73 for New Possibilities, .73 for Value vs. Cost, and .71 for Interestingness.The estimates use ratings from both evaluators across 201 ideas.Source: §4.2.4 · p. 12

5 Result

Across the automated evaluation and user study, Scanvas received stronger idea-quality ratings than the relevant baselines. Participants also rated its support for enrichment and synergy discovery more highly.

5 Result

  • The full automated pipeline received significantly higher ratings than the ablated baselines, and Scanvas-produced ideas received significantly higher expert ratings than baseline ideas.The comparisons are shown in Figures 4 and 5, respectively.Source: §5 · p. 12
  • Participants rated Scanvas more highly than the baseline for supporting idea enrichment and discovery of synergistic combinations.These experience ratings are shown in Figure 7, left.Source: §5 · p. 12
  • Scanvas received average Creativity Support Index scores above 5.7 on a 7-point scale across all dimensions.The reported scores are descriptive.Source: §5 · p. 12

5.1 Idea Quality

Scanvas outperformed its automated ablations and the user-study baseline on idea quality. The results also compare human-involved exploration with fully automated generation and examine diversity and participant creativity.

5.1.1 End-to-End Auto-Generation Results.

  • The full pipeline outperformed both ablations across all four idea-quality dimensions, while the ablations did not differ significantly.The results suggest that expansion and synergy generation jointly improve idea quality rather than either process alone.Source: §5.1.1 · pp. 12–13, Figure 4 · p. 13

5.1.2 User Study Idea Ratings.

  • Scanvas ideas scored 4.93 versus 3.99 for the baseline on composite expert ratings (p<.001), with higher ratings across all four dimensions.Scanvas also had higher mean ratings for goal fulfillment (5.16 vs. 4.50), new benefits (5.06 vs. 4.27), value vs. cost (4.45 vs. 3.28), and interestingness (5.03 vs. 3.89).Source: §5.1.2 · p. 13, Figure 5 · p. 14
  • Scanvas ideas had overall expert ratings comparable to the full auto-generation pipeline, with no significant differences on the pooled two-task comparisons.For example, the composite means were 4.93 for Scanvas and 5.12 for auto-generation; all reported comparisons had p≥.066.Source: §5.1.3 · pp. 14–15
  • Scanvas had a larger share of high ratings than auto-generation for Goal Fit (45% vs. 38%), Creative Benefit (34% vs. 30%), and Interestingness (41% vs. 32%); Worth Cost was similar (29% vs. 28%).Scanvas also produced a wider range of idea quality, including some low-scoring ideas.Source: §5.1.3 · p. 15, Figure 6 · p. 15
  • Scanvas ideas were significantly more semantically diverse than those from the automated-generation conditions.Table 4 compares average pairwise embedding distance among ideas across the two shared tasks.Source: §5.1.3 · p. 15, Table 4 · p. 16
  • The diversity results suggest that incorporating participants’ ideas and decisions broadened exploration beyond fully automated generation while Scanvas produced higher-rated ideas than the user-study baseline.The authors interpret this as preserving breadth from human exploration while helping participants develop it into higher-quality ideas.Source: §5.1.3 · p. 15

5.1.4 User Creativity Measurement and Interaction Effects.

  • No significant correlation was found between participants’ AUT fluency and either their mean idea-quality ratings or their Scanvas-minus-baseline gains.Quality correlations had |ρ|≤.45 with all p≥.14; gain correlations had |ρ|≤.35 with all p≥.27.Source: §5.1.4 · p. 16

5.2 User Feedback on How Enrichment Supported Synergy Discovery

Participants described Scanvas’s enrichment as supplying interpretable properties and directions for synergy discovery, while rating it more helpful than the baseline for finding and developing useful combinations.

5.2 User Feedback on How Enrichment Supported Synergy Discovery

  • Participants perceived Scanvas as more helpful than the baseline for finding and developing useful combinations, with significantly higher ratings on these aspects.The comparison is reported for participants’ perceived helpfulness and ratings of the ideation process.Source: §5.2 · p. 16
  • Idea analysis and expansion surfaced properties, opportunities, and issues that participants used as material to guide further exploration and synergy discovery.Participants described the sidebar analysis as exposing opportunities and issues, and enrichment also supported exploration after ideas were combined.Source: §5.2 · p. 17
  • For a kitchen-support task, a robot’s ingredient reminders and a portable recipe card’s format were combined into shopping support while preserving the user’s role as cook.The robot’s cooking role raised concerns about enjoyment and practice, while the recipe card was considered to require too much reading; the combination retained reminders and timing support.Source: §5.2 · p. 17
  • Analysis of a combined tactile-card and headphone idea surfaced a further opportunity: using visitors’ finger movements to improve museum audio descriptions.The example illustrates participants revisiting a combined idea for another round of enrichment and development.Source: §5.2 · pp. 17–18
  • Baseline expansion was used more often despite lower usefulness ratings, with participants describing some outputs as near-duplicates or out of scope.The authors suggest repeated baseline use may partly reflect trial-and-error, whereas Scanvas surfaced directions for participants to interpret and develop before deciding what to expand or combine.Source: §5.2 · p. 18

5.3 User Experience of Synergy Discovery

Participants said Scanvas helped them notice relationships between ideas and develop those relationships into combinations that created new value or reduced resource demands.

5.3 User Experience of Synergy Discovery

  • Participants perceived Scanvas as more helpful than the baseline for finding and developing useful combinations, including relationships that were difficult to see in the separate ideas.The section describes synergy discovery as finding such relationships and developing them to create value or achieve value with fewer resources or costs.Source: §5.3 · p. 18

5.3.1 Helping users find promising idea combinations they might otherwise miss.

  • Scanvas surfaced combinations participants might not have chosen, including a safety knob and cooking-step display combined into a safe practice mode for older adults.The burner stayed off while the knob and display supported rehearsal of the cooking process before real cooking.Source: §5.3.1 · pp. 18–19

5.3.2 Combining ideas to unlock new value that neither idea provided alone.

  • A tactile art card and headphone-based audio were combined into an orientation-aware shared sound experience, creating a social goal neither idea provided alone.Visitors and companions could hear a shared 3D sound space corresponding to the part of the artwork each was attending to.Source: §5.3.2 · p. 19

5.3.3 Combining ideas to reduce the cost or resources required by the original ideas.

  • Combinations reduced resource demands by reusing components or replacing costly dedicated ones, such as using a spoon rest for both cooking prompts and sensing.Because the spoon rest contacted the pot, it could sense heat and vibration while prompting the next cooking step, making a hood-mounted sensor optional.Source: §5.3.3 · pp. 19–20

6 Discussion

The discussion frames Scanvas as a way to develop and search relationships within existing ideas, with human-guided exploration complementing automated search. It also identifies limits in current evaluation and idea-space organization.

6.1 From Generating Ideas to Developing Synergistic Combinations

  • Scanvas treats combination as searching for relationships that create greater value from the same or fewer resources, not simply placing ideas together.The discussion presents the existing idea space as a source of new directions that can complement external inspiration.Source: §6.1 · p. 20
  • The authors argue that creativity support may productively explore uncertain, potentially high-value combinations rather than optimize only for consistently high-quality outputs.They note that exploring relationships among existing ideas may open directions that incremental development would not.Source: §6.1 · p. 20

6.2 Supporting Human–AI Collaboration Through Exploration and Pivoting

  • The interactive condition produced a larger share of highly rated ideas than the end-to-end automated pipeline across several evaluation dimensions.The authors suggest people may recognize latent potential and pivot toward directions automated search may not prioritize, while automation covers a larger combinatorial space without continuous user effort.Source: §6.2 · p. 20

6.3 Toward Value Functions for Evaluating Synergistic Ideas

  • A proposed value function would weigh an idea’s benefits against the costs or resources needed to realize them, while retaining human judgment in prioritization.The discussion presents this as a future direction for filtering and prioritizing large generated idea spaces.Source: §6.3 · p. 21

6.4 Limitations and Future Work

  • The participant sample was relatively small and did not capture the full range of expertise, domains, and working styles in professional design.The authors call for larger and more diverse practitioner studies to examine variation in how designers interpret opportunities and benefit from AI support.Source: §6.4 · p. 21
  • Scanvas offers limited support for organizing a growing idea space, which can make comparing related ideas, revisiting promising directions, and prioritizing exploration harder.Potential future supports include clustering, filtering, summarization, or adaptive views that preserve relationships among ideas.Source: §6.4 · p. 21
  • The evaluation covered short, controlled tasks rather than ongoing projects, so use under real project constraints and longer refinement cycles remains uncertain.The study used predefined scenarios and limited study time rather than having participants pursue ideas within an active project.Source: §6.4 · p. 21

6.5 Conclusion

  • The conclusion positions relationships among ideas as a central object of exploration, with AI broadening possible connections and people guiding their development.The authors report that the approach can improve idea quality and surface promising relationships and design directions that might otherwise be overlooked.Source: §6.5 · p. 22

7 Appendix

The appendix details study tutorials and surveys, including how participants’ baseline divergent-thinking scores related to their in-study outcomes.

7.1.1 Study System Tutorials.

  • Before each main task, participants received a step-by-step tutorial on the assigned system and practiced on a “Clean laundry with less water” demo task.The interviewer demonstrated the system and introduced its main features.Source: §7.1.1 · pp. 23–24

7.1.2 Post-study survey.

  • The post-study survey assessed ideation-process perceptions, feature usefulness, and resulting idea quality, and included the Creativity Support Index.Table 5 reports the Scanvas-condition CSI items, rated after tasks on a 7-point Likert scale.Source: §7.1.2 · p. 24, Table 5 · p. 24

7.1.3 Pre-study survey.

  • AUT fluency did not predict in-study idea quality, Scanvas-minus-Baseline gain, or the number of ideas produced.AUT fluency was measured by distinct alternative uses for five objects; none of its correlations with the idea-quality outcomes reached significance.Source: §7.1.3 · p. 24, Table 6 · p. 24
Loading 2609.34062v1…