Gradual Generation of User Interfaces as a Design Method for Malleable Software CCS: Human-centered computing HCI design and evaluation methods CCS: Human-centered computing Interaction paradigms

Bryan Min, Peiling Jiang, Zhicheng Huang, Haijun Xia · 2026

Read Gradual Generation of User Interfaces as a Design Method for Malleable Software CCS: Human-centered computing HCI design and evaluation methods CCS: Human-centered computing Interaction paradigms by Bryan Min, Peiling Jiang, Zhicheng Huang, Haijun Xia (2026) 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-2601-17975.txt.

Introduction

Bryan Min, Peiling Jiang, Zhicheng Huang, and Haijun Xia propose a gradual way to design AI-made user interfaces. Their 2026 preprint deals with generative UI, where a model builds screens from a user's prompt. Building good tools of this kind is hard. Users need to find the many ways a screen can be tuned, while the model keeps its full expressive range. Prompt boxes alone give few hints about what can be changed.

Classic design tools, on the other hand, show options but fix the form. The authors suggest growing the interface step by step as a design method for malleable software. Gradual growth can reveal choices as users need them. The work speaks to anyone building AI tools that users must shape to fit their tasks.

Catalog record

CREATORBryan Min, Peiling Jiang, Zhicheng Huang, Haijun Xia
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-17975

Browse the full library · About the Tower of Babel

Full text

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

Gradual Generation of User Interfaces as a Design Method for Malleable Software CCS: Human-centered computing HCI design and evaluation methods CCS: Human-centered computing Interaction paradigms Bryan Min, Peiling Jiang, Zhicheng Huang, Haijun Xia Abstract . AI is growing increasingly capable of automatically generating user interfaces (GenUI) from user prompts. However, designing GenUI applications that enable users to discover diverse customizations while preserving GenUI’s expressiveness remains challenging. Current design methods—presenting prompt boxes and leveraging context—lack affordances for customization discovery, while traditional menu-based approaches become overly complex given GenUI’s vast customization space. We propose Gradually Generating User Interfaces—a design method that structures customizations into intermediate UI layers that AI gradually loads during interface generation. These intermediate stages expose different customization features along specific dimensions, making them discoverable to users. Users can wind back the generation process to access customizations. We demonstrate this approach through three prototype websites, showing how designers can support GenUI’s expanded customization capabilities while maintaining visual simplicity and discoverability. Our work offers a practical method for integrating customization features into GenUI applications, contributing an approach to designing malleable software. Figure 1 . We present Gradual Generation as a method for designing customizations in generative UI applications. Gradual Generation proposes UI/UX designers to organize UI customizations into multiple intermediate layers that AI progressively loads during interface generation, enabling users to discover and access customizations by rewinding to earlier stages. Four side-by-side panels illustrate ``Gradual Generation'' for customizing a video feed homepage through intermediate layers, with arrows showing a left-to-right progression. Layer 1 (Categories) shows a list of feed categories (e.g., ``Most Popular,'' ``Gaming,'' ``Sports,'' ``Shorts'') with controls to reorder and add new categories. Layer 2 (Layout) shows those categories rearranged and resized into sections, emphasizing ``Most Popular.'' Layer 3 (Content) shows each section populated with video cards and controls to refine what is shown. Layer 4 (Style) shows the same feed with visual styling controls (e.g., card radius, spacing, hover scale). Beneath each layer is a corresponding specification format: natural language (Categories), JSON schema (Layout), code (Content), and CSS variables (Style). 1. Introduction Advances in generative AI are driving a resurgence for Generative UIs (GenUIs)—user interfaces that AI can automatically generate based on user prompts and context. In fact, we are already seeing new capabilities emerge from GenUI, such as generating custom browser tabs and websites ( 6 ; 4 ; 2 ) , generating UI widgets within chatbots ( 10 ; 1 ) , and customizing UIs from user prompts ( 8 ; 9 ) . This demonstrates a promising future for making software more malleable ( 8 ; 5 ) and easily customizable according to users’ needs. However, it remains unclear how to best design GenUI so that users can discover diverse customizations while preserving GenUI’s expressiveness. Current design presents prompt boxes and leverages background context to automatically generate UIs. This paradigm lacks the affordances necessary for users to discover customization options ( 11 ; 12 ) . On the other hand, if designers revert to traditional UI customization methods by organizing sets of options into menu categories, the interface becomes overly complex given the vast space of possible GenUI customizations ( 7 ) . This challenge calls for the HCI community to explore new paradigms for integrating customization features into GenUI applications. We argue that the transition from prompt to generated UI itself presents an opportunity to surface diverse customization options. We posit that providing intermediate stages between the user’s initial prompt and the final generated interface creates a valuable design space for organizing different categories of UI customizations. In this paper, we aim to explore Gradually Generating User Interfaces —a design method that structures sets of customizations into intermediate UI “layers” that AI gradually loads during the process of generating the final interface. These intermediate UIs expose different sets of customization features along specific dimensions, making them discoverable to users. Users can then “wind back” the gradual generation process to access the customizations they need. This approach enables designers to support the broader space of customizations that GenUI promises while maintaining both visual simplicity and discoverability (Fig. 1 ). We first introduce the Gradual Generation design method, then demonstrate its potential through three prototype websites. Our prototypes show how UI/UX designers can effectively integrate GenUI’s expanded customization capabilities without sacrificing interface usability. Finally, we discuss design implications and future work for this approach to achieving malleable software with AI. 2. Gradual Generation of User Interfaces Gradual Generation is a GenUI design approach that enables users to discover possible customizations in an interface by presenting them while the UI generates step-by-step. Upon entering a GenUI page, the interface presents intermediate UIs that show key classes of customizations. For example, a video homepage illustrated in Fig. 1 may present four intermediate UIs before presenting the final interface: the user’s preferred video categories , the layout of these categories, the content to show in each video, and finally the style . The categories layer can allow users to modify the ranking of their preferences, the layout layer can allow users to resize and rearrange the categories on the page, the content layer can allow users to determine which attributes are shown or hidden in each video, and the style layer can allow users to personalize the aesthetic of their page. This design method differs from the traditional one in that designers are responsible for identifying the key intermediate UIs, as opposed to categorizing the set of customizations. Figure 2 . Intermediate UI layers provide specific sets of customizations. For example, designers can dedicate a “Layout” layer to customizations that let users resize, rearrange, and combine sections in the UI. Figure 3 . Example intermediate layers of (1) a team scheduling calender interface, and (2) course management system interface. Example intermediate layers of (1) a team scheduling calender interface, and (2) course management system interface. 2.1. The Gradual Generation Design Method The Gradual Generation design method involves four main tasks for designers. To illustrate these tasks, we continue the example of designing a video homepage UI presented in Fig. 1 . (1) Identify key intermediate stages of the UI Designers first identify key intermediate stages of the UI by reflecting on their design process for producing UI mockups from raw data. For example, when designing a video homepage UI, intermediate stages might include: a categories stage that lists category rankings from the software’s recommendation algorithm, followed by stages for visual layout of individual categories, then video content and data attributes within each category, and finally aesthetic styling of the homepage. (2) Design customizations for each intermediate stage For each stage, designers explore and design customizations specific to that level of the interface (Fig. 2 ). For instance, the category rankings stage might support customizing ranking order, introducing additional categories, or refining categories into subcategories (e.g., specifying "gaming" as "e-sports"). (3) Develop specifications for customizations in each stage Designers then create specifications documenting the available customizations for each intermediate stage. These specifications can take various forms: documentation pages listing all categories the application’s recommender system can output, JSON type definitions, or code blocks. GenUI uses these specifications to generate interfaces based on each stage’s parameter set, where UI changes map directly

Read the complete text