IDEA: Augmenting Design Intelligence through Design Space Exploration

Chuer Chen, Xiaoke Yan, Xiaoyu Qi, and Nan Cao · 2025

Read IDEA: Augmenting Design Intelligence through Design Space Exploration by Chuer Chen, Xiaoke Yan, Xiaoyu Qi, and Nan Cao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-10587.txt.

Introduction

Design choices explode into vast option spaces, and Chuer Chen and co-authors build a framework to search them in a 2025 paper. Designers now lean on gut feel, since option spaces lack a math form that computers can search. Their model lays out a design space as clear axes, each with a set of pickable elements.

A language model writes the constraints for a task, and a tree search explores picks under those rules. Chosen abstract picks are then built into real artifacts. Two test scenes are shown: laying out data-driven articles and making pictorial charts. The frame aims to pair human taste with tireless search, so small teams can test far more options than time once allowed.

Catalog record

CREATORChuer Chen, Xiaoke Yan, Xiaoyu Qi, and Nan Cao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-10587

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

IDEA: Augmenting Design Intelligence through Design Space Exploration Chuer Chen, Xiaoke Yan, Xiaoyu Qi, and Nan Cao Abstract Design spaces serve as a conceptual framework that enables designers to explore feasible solutions through the selection and combination of design elements. However, effective decision-making remains heavily dependent on the designer’s experience, and the absence of mathematical formalization prevents computational support for automated design processes. To bridge this gap, we introduce a structured representation that models design spaces with orthogonal dimensions and discrete selectable elements. Building on this model, we present IDEA, a decision-making framework for augmenting design intelligence through design space exploration to generate effective outcomes. Specifically, IDEA leverages large language models (LLMs) for constraint generation, incorporates a Monte Carlo Tree Search (MCTS) algorithm guided by these constraints to explore the design space efficiently, and instantiates abstract decisions into domain-specific implementations. We validate IDEA in two design scenarios: data-driven article composition and pictorial visualization generation, supported by example results, expert interviews, and a user study. The evaluation demonstrates the IDEA’s adaptability across domains and its capability to produce superior design outcomes. I Introduction Adesign space serves as a structured repository of design knowledge that codifies design possibilities through interdependent dimensions and their valid element combinations. It has been widely adopted across design communities, such as visualization design [ 1 ] and architectural design [ 2 ] , to assist designers in making rational decisions. Recently, with the rise of generative AI, large-scale models have been increasingly used in design automation. However, their outputs often violate domain constraints or lack precise control over stylistic details [ 3 ] . The construction of effective design spaces has thus become pivotal — not merely as academic exercises, but as foundational infrastructure enabling reliable design automation. In the visualization community, researchers have progressively formalized design spaces by analyzing design exemplars and abstracting recurring patterns into structured references. Prior research can be broadly categorized into two directions: guiding designers through spaces for aesthetic or expressive choices [ 4 , 5 ] , and supporting tool development by structuring dimensions around system requirements [ 6 , 7 ] . However, despite its widespread adoption as a conceptual tool, the construction and application of design spaces face three persistent challenges. First, traditional space construction predominantly relies on subjective expertise rather than systematic methodologies, resulting in ad-hoc frameworks that struggle to generalize across domains or evolving requirements. Second, existing spaces exhibit limited expressiveness in modeling intricate interdependencies among multidimensional design parameters, constraining their utility in complex scenarios requiring dynamic trade-off analysis. Third, the decision-making process remains manual, exacerbating cognitive load while stifling efficiency and innovative potential. To address the above challenges, we propose a structured design space model that mathematically formalizes design spaces as executable decision structures with orthogonal dimensions and discrete elements, enabling systematic construction, constraint modeling, and automated decision-making. Building on this foundation, we introduce IDEA (i.e., i ntelligent d esign space e xploration and a ugmentation), a decision-making framework that leverages the design space to guide generative models, producing effective design outcomes from given requirements through three synergistic modules: (1) a LLM-powered constraint generation module that translates requirements into logical rules, (2) a constraint-guided Monte Carlo Tree Search that navigates the space for optimal solutions, and (3) an adaptive action execution module that synthesizes domain-specific implementations from abstract decisions. We demonstrate IDEA’s cross-domain effectiveness in cases of data-driven article composition and pictorial visualization generation. The major contributions of the paper are as follows: • We propose a structured design space model that formalizes design knowledge into orthogonal dimensions and discrete elements as machine-interpretable decision structures, thereby enhancing computational understanding and intelligent design. • We introduce IDEA, an intelligent design decision-making framework that transforms user requirements into high-quality outcomes through constraint-guided exploration of design spaces, with adaptable structures that support diverse domains. • We validate IDEA ’s effectiveness and generalizability through two scenarios: data-driven articles and pictorial visualizations, with two sets of example results, a user study confirming outcome quality, and expert interviews affirming human-aligned decision-making and practical applicability. II Related Work In this section, we review the most relevant prior work across three areas, including design space, computational design, and algorithmic decision-making in design. II-A Design Space The conceptualization of design spaces as structured frameworks has evolved across design disciplines. Mackinlay [ 8 ] first characterized it in the 1980s as ”a space of possible designs”. Recently, Halskov and Lundqvist [ 2 ] refined this perspective to designer agency, defining it as ”the space of design possibilities and alternatives that a designer considers to meet a design brief”. Another influential perspective came from mathematical formalization. Shaw [ 9 ] defined design spaces as ”discrete Cartesian spaces where design decisions form orthogonal dimensions, alternatives constitute dimension values, and complete designs are space points”. Similarly, Schulz et al. [ 10 ] decomposed the concept into actionable parameters, proposing that ”a design space comprises finite dimensions, each capturing a specific design decision required to fully specify a solution”. Building on these foundational definitions, visualization researchers have explored design spaces across various applications. In traditional visualization, structured design spaces have been developed for core techniques including scatterplots [ 11 ] , maps [ 12 ] , graph layouts [ 13 , 14 ] , and matrix visualizations [ 15 ] . The emergence of narrative visualization has expanded the design space to encompass more dimensions and choices, accommodating storytelling logic and structural coherence. For example, Bach et al. [ 16 ] proposed an 8-dimensional dashboard design space, distinguishing content-oriented dimensions (e.g., data, visual representation) from composition-focused ones (e.g., page layout, color, structure). Yang et al. [ 17 ] analyzed data stories using Freytag’s Pyramid [ 18 ] narrative structure, and identified a design space consisting of three dimensions: narrative pattern, data flow, and visual communication. While existing design spaces provide structured references for human designers, their formulation lacks computational interpretability, preventing intelligent design. Building on Shaw’s Cartesian model [ 9 ] and Schulz’s parametric decomposition [ 10 ] , we propose a structured design space model that formalizes dimensions as orthogonal axes, elements as discrete nodes, and actions as executable functions, bridging the gap between conceptual exploration and automated implementation. II-B Design Automation Design emerges from a series of deliberate decisions, such as selecting color schemes, arranging visual hierarchies, and configuring typography. The first step toward design automation lies in understanding how these decisions collectively shape effective designs. For instance, Zhao et al. [ 19 ] proposed a deep ranking framework to learn a convolutional neural network for exploring the effects of various design factors (e.g., color, font) on graphic designs. Baechler et al. [ 20 ] designed a vision-language model for UI and infographics understanding, including tasks such as question answering, element annotation, and summarization. To automate the complicated design creation process, researchers take a further step toward intelligent design generation. Early studies have investigated template-based methods. Cui et al. [ 6 ] developed an authoring system that automatically converts statements to a set of infographics with pre-designed styles. Yang et al. [ 21 ] proposed topic-dependent layout templates grounded in aesthetic principles and a computational framework to integrate layout elements under template constraints. Recent advance in multimodal foundation

Read the complete text