Source-linked AI summary
Luminate: Structured Generation and Exploration of Design Space with Large Language Models for Human-AI Co-Creation
Sangho Suh, Meng Chen, Bryan Min, Toby Jia-Jun Li, Haijun Xia
TL;DR
Creative LLM interactions can encourage rapid convergence instead of broad exploration, despite models’ ability to generate many outputs. The paper proposes a framework and Luminate system that structure responses through generated design dimensions, with a study of 14 professional writers indicating useful design-space exploration. The authors conclude that this approach can support human-AI creative exploration, while noting limits in response controls and comparative evaluation.
Problem
Current LLM interactions do not systematically structure output generation and can constrain users’ exploration of creative design spaces.
Method
The framework uses LLMs to generate key dimensions from a prompt, then generates responses using those dimensions for structured exploration in Luminate.
Results
A study with 14 professional writers found that Luminate had potential to help users effectively explore the design space during creative processes.
Takeaways & Limitations
The approach demonstrates how humans and LLMs can complement each other by combining rapid generation with human contextual comprehension and judgment.
Takeaways & Limitations
Luminate fixed responses at 150 words and 40 outputs, limiting control over response length and number and potentially affecting perceived value.
Abstract
from arXiv · showhide
Thanks to their generative capabilities, large language models (LLMs) have become an invaluable tool for creative processes. These models have the capacity to produce hundreds and thousands of visual and textual outputs, offering abundant inspiration for creative endeavors. But are we harnessing their full potential? We argue that current interaction paradigms fall short, guiding users towards rapid convergence on a limited set of ideas, rather than empowering them to explore the vast latent design space in generative models. To address this limitation, we propose a framework that facilitates the structured generation of design space in which users can seamlessly explore, evaluate, and synthesize a multitude of responses. We demonstrate the feasibility and usefulness of this framework through the design and development of an interactive system, Luminate, and a user study with 14 professional writers. Our work advances how we interact with LLMs for creative tasks, introducing a way to harness the creative potential of LLMs.
1 INTRODUCTION
Creative work benefits from generating and exploring many ideas, but current LLM interactions often steer users toward narrow, premature convergence. The paper proposes structuring LLM-generated responses around design dimensions so users can explore broader design spaces.
- Creative processes benefit from exploring multiple ideas before refining one, because premature convergence can cause fixation on a sub-optimal idea.
- LLMs can instantly produce tens to hundreds of outputs, creating an opportunity to help users understand design spaces comprehensively.
- Current LLM interaction paradigms often present similar or unstructured outputs that confine users to narrow spaces and encourage convergence.
- The proposed framework first uses LLMs to generate key dimensions, then structures responses around them for exploration.
- Luminate instantiates the framework, and a study with 14 professional writers found potential for effective design-space exploration during creative work.
2 BACKGROUND
The background positions Luminate at the intersection of creative-process support, human-AI collaboration, design-space reasoning, and visualization. Prior work motivates explicit dimensions and visual exploration to help people compare, generate, and navigate ideas.
- Supporting Creative Process: Creative thinking alternates between divergent generation of multiple ideas and convergent evaluation and synthesis into an outcome.
- Design Space: Design space denotes a conceptual space of possible ideas, designs, concepts, or solutions organized by dimensions and their possible values.
- Dimensional Reasoning: Making dimensions explicit enables designers to examine the problem space, desired characteristics, and alternative possibilities at a higher conceptual level.
- Dimensional Reasoning: Dimensional reasoning does not arise naturally, and even experts can struggle to identify relevant dimensions without explicit support.
- Co-creating with AI: Human-AI collaboration can combine AI’s rapid information generation with humans’ broader contextual comprehension and judgment of task suitability.
- Facilitating Exploration and Sensemaking: Visualization techniques help users inspect multivariate response distributions, while semantic zooming supports movement between overview and detailed text views.
- Facilitating Exploration and Sensemaking: Luminate extends visualization and interaction techniques to explore multivariate LLM responses and support iterative divergent and convergent creative processes.
3 PROMPTING FOR DESIGN SPACE: A FRAMEWORK FOR DESIGN SPACE THINKING IN HUMAN-AI CO-CREATION
The framework reframes early LLM-assisted creation as structured design-space exploration rather than direct production of individual artifacts. It addresses premature convergence and difficult dimensional reasoning by generating dimensions, organizing responses across them, and supporting flexible exploration.
- 3 PROMPTING FOR DESIGN SPACE: A FRAMEWORK FOR DESIGN SPACE THINKING IN HUMAN-AI CO-CREATION: The framework argues that generative AI should help construct design spaces early in creative work, supporting fixation avoidance and systematic idea generation and evaluation.
- 3.1 Current Interaction Paradigms (Figs. 1A & B): Existing single-output and multi-output LLM interactions rely on unstructured prompting that encourages users to refine or compare outputs without systematic dimensions.
- 3.1 Current Interaction Paradigms (Figs. 1A & B): When prompts omit elements such as setting, plot, personality, and tone, the LLM fills them in, potentially steering users toward a single narrative direction.
- 3.2 Structured Multi-Output Approach (Fig. 1C): The structured multi-output approach is motivated by evidence that dimensions aid comparison and idea generation while explicit multiple-idea generation counters premature convergence.
- 3.2.1 Motivation.: The system combines generated dimensions and values into requirements that guide response generation across multiple design-space subspaces.
- 3.2.1 Motivation.: Luminate uses LLMs to generate relevant categorical dimensions and values, helping address experts’ difficulty identifying dimensions for a problem space.
- 3.2.2 Design Goals.: Its design goals include reducing dimension-generation difficulty, supporting structured exploration, and enabling transitions between divergent and convergent thinking.
- 3.2.2 Design Goals.: The interface pairs a text editor with an exploration view containing response clusters, dimensions, adjustable axes, and semantic zoom controls.
4 LUMINATE
The section introduces Luminate’s interface, workflow demonstration, and implementation overview for a creative story-writing task.
- The section presents Luminate’s interface features and interactions.
- It demonstrates the system through a creative story-writing workflow.
- It concludes with a concise overview of the implementation.
4.1 Interface & Features
Luminate combines an editor with a visual exploration canvas to generate, organize, inspect, and iteratively steer responses through explicit dimensions.
- Interface: The interface pairs a text editor with an exploration view that initially presents generated responses as a single cluster.Selecting a node maps its response into the editor and highlights the corresponding node.
- Dimension-Guided Response Generation: Luminate generates categorical dimensions and values from a prompt, then uses randomly selected dimension values as requirements for response generation.Each requirement list is sent to the LLM to produce responses matching the selected attributes.
- Exploration: The response space supports exploration by adding generated examples and using dimensions to cluster and arrange them.Examples and responses can be inspected as part of a structured design-space workflow.
- Dimension Selection: Selecting one or two dimensions repositions responses into one- or two-dimensional layouts aligned with their dimension values.For example, Mood and Tone can organize responses spatially by values such as Romantic or Motivating.
- Generation Steering: Users can steer generation by creating similar responses, generating responses within filtered subspaces, or adding dimensions to revise existing responses.The system describes these as three techniques for generating new responses or attributes.
4.1.5 Semantic Zoom (Fig. 8, DG5).
Luminate supports navigation and targeted exploration by changing response detail, generating related content, filtering results, and using selected text as generation context.
- Semantic Zoom: Semantic zoom changes the amount of response information shown at different zoom scales.Users can navigate between overview and detailed views of responses.
- Response Generation: Users can generate similar responses from a node or add responses to a filtered subspace.New nodes join the canvas and align with their dimension values.
- User-Defined Dimensions: Adding a user-defined dimension generates values and revises existing responses so they can be searched, filtered, and clustered by the new dimension.The example uses Time Period values such as Victorian, Medieval, Colonial, and Apocalyptic.
- Search and Filtering: Keyword and dimension filters lower the opacity of nonmatching responses to help users locate responses of interest without discarding spatial information.Search can target words, while filters can target dimension values.
- Contextual Generation: Users can highlight selected editor text to request targeted generation focused on a specific phrase or line.This supports expansion of selected content rather than relying on the entire text as context.
4.2 Example Workflow: Writing Short Story
The example workflow shows a writer using Luminate to explore many story ideas, inspect their attributes, reorganize them by dimensions, bookmark promising options, and edit selected text into a draft.
- Divergent Thinking Phase (Exploration): Chris begins with a rabbit-related prompt and explores generated stories through a canvas containing clustered response nodes.Selecting nodes updates the text shown in the editor.
- Divergent Thinking Phase (Exploration): Semantic zoom lets Chris move from titles to keywords, summaries, and full text while navigating story responses.This progressively reveals information such as plot, setting, genre, and tone.
- Divergent Thinking Phase (Exploration): Chris uses response attributes to recognize alternative story directions, including a squirrel protagonist and rabbit antagonist in a mysterious forest setting.The displayed attributes suggest other genres and settings he might explore.
- Divergent Thinking Phase (Exploration): He generates additional stories, bookmarks promising ideas, and rearranges the response space by Genre and Setting dimensions.The nodes move from a cluster to positions aligned with labels such as Mystery, forest, urban area, and underwater.
- Convergent Thinking (Synthesis): Chris retrieves bookmarked stories, embeds selected text into the editor, and edits the accumulated material into a story draft.Selecting another node appends its text beneath previously embedded text.
- Convergent Thinking (Synthesis): The scenario demonstrates an iterative interplay between divergent and convergent thinking but does not cover all Luminate features.It is presented as one of numerous possible workflows.
4.3 Implementation & Prompt Engineering
Luminate uses a web interface implemented with React, Editor.js, and d3.js, and relies on sequential LLM prompts to generate dimensions, requirements, and responses.
- Implementation: The system uses React, Editor.js, and d3.js for the interface, text editor, exploration view, clustering, and point positioning.The exploration view uses d3.js force layout for clustering.
- Prompt Engineering: Luminate first prompts an LLM for categorical dimensions and values, then forms requirement lists with one random value per dimension before generating responses.The implementation used OpenAI’s text-davinci-003 API for these generation steps.
5 USER STUDY
The study evaluated Luminate with 14 professional writers who used the system for creative writing tasks after an interface tutorial. Participants completed surveys, interviews, and think-aloud tasks during an 80-minute online study.
- Study Procedure: Researchers recorded participants’ screens and audio to support transcription and analysis of their interactions and feedback.Participants also thought aloud throughout the creative-writing task.
- Participants: 14 professional writers with prior AI-writing experience participated in the online user study.Participants had several years of creative-writing experience and had used tools including ChatGPT, JasperAI, and DALL-E.
- Tasks: The study assigned tasks according to participants’ prior experiences, including copywriting, stories, emails, poems, a video-game scene, and song lyrics.The task distribution covered six creative-writing categories.
- Study Procedure: Participants completed a pre-study survey, a 25-minute interface tutorial, and a 25-minute creative-writing task with AI support.The study concluded with a post-study survey and semi-structured exit interview.
6 RESULTS
Participants generally found Luminate usable and useful for creative exploration, reporting that its responses and dimensions supported divergent thinking and understanding of design space. Their workflows combined single batches, prompt chains, and dimension-driven generation, while response volume created some cognitive overload.
- Usability: 14 participants generally found Luminate usable, with most agreeing it was easy to use and easy to learn.Five participants were neutral or disagreed because the study provided insufficient time to learn all features.
- Creativity Support: 82.16 was Luminate’s Creativity Support Index score, and participants gave the strongest support to exploration of ideas.All but one would use Luminate for creative writing, while 13 would use it for exploring ideas generally.
- Creative Workflows: 8/14 participants used chains of prompts, while 5/14 used dimension-driven generation and 1/14 used one or multiple responses from a single prompt.These workflows demonstrate flexible ways of incorporating Luminate into creative work.
- Divergent Thinking: 9/10 was Luminate’s average score for helping participants track ideas and explore many possibilities.All participants reported that multiple options from one prompt benefited brainstorming and ideation.
- Divergent Thinking: Participants said Luminate’s varied responses exposed them to ideas, scenarios, and perspectives they had not previously considered.They associated the breadth of outputs with thinking outside the box and broader creative horizons.
- Design-Space Understanding: Participants generally agreed that Luminate improved their understanding of possible responses and AI capabilities, while helping them foresee response spaces.They also used generated dimensions to explore 1-dimensional and 2-dimensional design spaces.
- Design-Space Understanding: All participants agreed that the exploration view helped them comprehensively understand the design space, and 13 agreed that selecting dimensions helped build that understanding.Participants also reported that dimensions helped them reflect on important task factors and were accurate, useful, and relevant.
7 DISCUSSION
The discussion frames Luminate as a structured way for humans and LLMs to explore, evaluate, and guide creative outputs, while identifying scope and usability limitations. It also outlines future extensions beyond writing.
- Human-AI complementarity: Luminate combines LLM generation with human judgment by helping users navigate and evaluate responses through dimensions.LLMs generate information quickly, while humans assess broader context and task requirements.
- Human-AI complementarity: User-defined dimensions and values let users express preferences and guide the generation or updating of responses.The framework supports adding dimensions and values to generate additional responses and update existing ones.
- Limitations: The study scope is constrained by fixed outputs: each response contained 150 words and each task produced 40 responses.Participants sometimes preferred longer or fewer responses, which may have reduced Luminate’s perceived value relative to conventional tools.
- Limitations: The findings primarily reflect creative writers experienced with AI, leaving usefulness and scalability for non-professional writers untested.The authors call for testing with more diverse participant groups.
- Limitations: Luminate was not rigorously compared with existing AI tools because the comparison with ChatGPT relied on participant self-reports.The authors state that their study does not answer where Luminate stands among AI writing tools.
- Future work: Future work should support adjustable interfaces, diverse user groups, and comparisons with alternative AI tools where feasible.The proposed redesign also considers transitions between simpler and more complex interfaces.
- Future extensions: The framework may extend from writing to image and video tasks, with interaction techniques such as semantic zoom supporting exploration across domains.The authors describe this as a possible broader application rather than an evaluated result.
- Limitations: The framework currently uses categorical ordinal and nominal dimensions, but their usefulness can vary by task and user.Examples include unwanted creativity levels and irrelevant dimensions such as gender for a Peruvian-food prompt.
8 CONCLUSION
The paper presents Prompting for Design Space as a framework for systematically generating and exploring diverse LLM outputs. Luminate and a study with 14 professional writers demonstrate its practicality and effectiveness for writing support.
- Conclusion: Prompting for Design Space structures the creation of a design space where users can explore, assess, and integrate numerous LLM responses.The paper positions this as a way to better harness LLMs’ creative potential during creative processes.
- Conclusion: A Luminate implementation and study involving 14 professional writers demonstrated the framework’s practicality and effectiveness.The paper describes this as the first systematic text-to-text exploration of multiple responses from a single prompt, to the authors’ knowledge.
A.1 Examples of Study Participant’s Creative Writing: Copywriting, Story, Email, and Lyrics
The appendix presents participant prompts and generated writing examples spanning stories, emails, lyrics, and Peruvian-cuisine copy. It also documents a generation-time measurement table.
- Peruvian-cuisine copy: The Peruvian-cuisine examples discuss cuisine as a social space, international flavor fusion, accessibility, and preservation of original flavor profiles.They mention ingredients, income, foreign culinary influence, and modernist techniques alongside accessibility.
- Story writing: The examples include a story-writing task in which a participant prompted Luminate to write and then extend a futuristic adventure story.The generated story follows a time traveller exploring a futuristic city, gadgets, and an amusement-park ride.
- Email writing: The email-writing task asked for a kind first-person notice ending a translator’s current project work while preserving consideration for future projects.The generated email communicates the decision, expresses appreciation, and invites questions.
- Lyrics writing: The lyrics task developed a dark narrative about death, money, politics, corporations, ambition, and Sir William’s conflict with the world.Multiple verses and choruses elaborate the setting, corporate control, division, and Sir William’s truth-telling role.
- Lyrics writing: Additional lyric examples portray Sir William as a truthful figure whose confrontation with darkness changes both him and the world.The later verses emphasize his honesty, reputation, struggle, and transformation.
- Generation measurements: Table 2 records response-generation time for different numbers of dimensions across three trials and reports failed-call ratios.Failure includes API errors or responses that are not well-formed valid JSON; time is measured in seconds.