Source-linked AI summary
Maru: Information Architecture as a Shared Language for Generating Aligned and Persistent User Interfaces
Eunhye Kim, DaEun Choi, Bryan Min, Hyunjung Yi, Yue Jiang, Juho Kim
TL;DR
GenUI systems do not preserve the structural preferences users establish across iterative information tasks, leaving alignment to repeated model decisions. The paper derives four IA elements, implements them as persistent and user-editable rules in Maru, and finds stronger cross-session alignment than a baseline, alongside limits from rule accumulation and scope.
Problem
GenUI systems leave structural decisions to the model at each generation, despite users establishing preferences about grouping, prioritization, and terminology across tasks.
Method
The paper derives partition, hierarchy, order, and vocabulary from prior work and instantiates them as persistent IA rules in Maru’s generation pipeline.
Results
74% to 61% approval across sessions for Maru versus 71% to 33% for the baseline, while 97% of user-created rules arose through ordinary interaction and query behavior.
Takeaways & Limitations
IA persistence can help users shape information-task interfaces to their needs, but persistence requires boundaries to remain useful.
Takeaways & Limitations
The study addresses IA-level structural persistence, while users may also want finer-grained persistence of layout type and component choices.
Abstract
from arXiv · showhide
Generative user interfaces (GenUIs) promise on-demand components tailored to users' needs. As users iterate on information tasks, they construct personal structures over information they encounter---how items are grouped, what gets prioritized, and what terms mean in their context. Yet, current systems leave these structural decisions to the model at each generation, ignoring the structural logic users have established. Without a persistent representational structure shared between user and system, GenUIs have no basis to remain aligned with what users have established. We draw on Information Architecture (IA), a design practice for organizing and structuring information, as a shared language to bridge user-constructed structure and system generation. We present a framework identifying four IA elements---partition, hierarchy, order, and vocabulary---and characterize how each maps to concrete UI generation decisions. We instantiate this framework in Maru, a conversational system that captures user prompts and interactions as IA preferences, persisting as rules both user and system draw on across generations. A user study revealed that IA persistence kept generated UIs aligned as sessions progressed, while alignment without it degraded, with diverse patterns emerging across users and contexts, pointing to the value of IA persistence in aligning GenUI to individual needs.
1 Introduction
GenUI systems can generate personalized interfaces, but they do not preserve which structural elements users value across iterative prompts. The paper uses Information Architecture to encode and persist those preferences, then evaluates Maru’s effect on alignment and personalization.
- GenUI systems produce structured interfaces on demand, but prior advances do not specify which elements should change, stay, or carry importance across prompts.
- Users may value particular interface structures, yet current systems cannot recognize or guarantee that those structures persist, forcing users to restate preferences.
- Information Architecture maps user needs to concrete content and layouts while capturing structural logic that remains inspectable and editable across generations.
- The framework derives four IA elements—partition, hierarchy, order, and vocabulary—and maps them to generation decisions before instantiating them in Maru.
- Maru’s user study found that persistence shaped UI alignment, engagement, and personalization across generated interfaces.
2 Related Work
Related work improves GenUI generation and carries conversational context, but leaves structural interface preferences to model decisions. This paper makes that mapping explicit through a persistent, user-editable IA layer.
- LLM-powered GenUI systems can generate interfaces across domains, but keeping them aligned with users’ needs as tasks progress remains a challenge.
- Existing approaches enrich generation with intermediate data models or inferred user needs, while still delegating structural UI decisions to the LLM.
- Persistence research shows that users construct schemas, criteria, priorities, and relational maps, while externalized or editable representations reduce context reconstruction and increase control.
- Information Architecture organizes, labels, and structures information through decisions about grouping, hierarchy, sequence, and terminology.
- The paper positions IA as a persistent, user-editable layer that explicitly maps user behavior and structural rules to interface generation.
3 A Framework of IA Elements for GenUI
The framework derives four IA elements from sensemaking research and links each to how users organize information and how interfaces display or construct that organization.
- Framework construction: The framework was built by collecting literature and tools, identifying supported user needs and UI patterns, and synthesizing recurring mappings between behavior and structure.
- Framework construction: A thematic analysis of systems, theories, and empirical studies examined organization and navigation during information tasks.
- Four IA elements: The four IA elements are partition and hierarchy as structural elements, plus order and vocabulary as semantic elements.
- Partition: Partition groups lateral peers without implying hierarchy, such as listings grouped by neighborhood or criteria divided into must-haves and nice-to-haves.
- Hierarchy: Hierarchy assigns differentiated roles between items, with position signaling relationships such as containment or evidential support.
- Order: Order assigns sequence or weight according to subjective importance, priority, or relevance, such as weighting safety above commute time.
- Vocabulary: Vocabulary establishes personal, task-specific terms as persistent semantic anchors for organizing, labeling, and evaluating information consistently.
4 The Maru Probe
Maru uses persistent Information Architecture rules to capture structural preferences from queries, interactions, and direct edits, then apply them across generated interfaces. Its pipeline separates rule construction, retrieval, layout selection, content filling, and interactive rendering.
- 4 The Maru Probe: Users can construct IA rules through chat, generated-UI interactions, or direct editing in the IA panel.Generated interfaces can scaffold initial hierarchy and partition rules, while interactions such as dragging update rules directly.
- 4 The Maru Probe: Layout selection uses the query, examples, and hierarchy and partition rules to produce a layout skeleton specifying component types and group labels.The query remains necessary for task-specific priorities that structural rules alone cannot infer.
- 4 The Maru Probe: The Extractor detects IA rules from queries, stores them with system-scaffolded rules, and retrieves relevant rules for later generations.Retrieval uses summarized chat context and the full active rule set to apply rules on top of one another.
- 4 The Maru Probe: Content filling combines the selected skeleton and query with all four IA rule types before rendering typed, interactive components.Order and vocabulary affect which items surface, how they are labeled, and how they are ranked within the selected structure.
- 4 The Maru Probe: Stable UUID pointers or concept labels bind rules to data so partition memberships and hierarchy relationships persist across renames and field edits.Concept references are upgraded to UUID-based links when extracted data items are available.
5 User Study
The user study compares Maru with a baseline lacking the IA layer to examine persistence, user communication of preferences, and the diversity of generated interfaces. It uses 12 participants across assigned information tasks and a participant-selected real task.
- 5 User Study: The comparative study evaluates whether persistent IA rules improve alignment, how users communicate preferences, and whether rule states produce distinct interfaces.Maru accumulates rules across queries and interactions, whereas the baseline does not.
- 5 User Study: The baseline removes rule detection, storage, schema hints, and rule-based layout examples while retaining Maru’s interactions, layout types, queries, and shared base context.This design treats the baseline as an ablation of the IA layer.
- 5 User Study: 12 participants, averaging 24.3 years old, were recruited; all reported daily or near-daily LLM use and prior experience with information-intensive tasks.The sample included 5 male and 7 female participants.
- 5 User Study: Participants completed three sessions: two assigned-task sessions using Maru or the baseline and one Maru session for a real information task.System order was counterbalanced across participants for the first two sessions.
- 5 User Study: The study logged participant actions, system outputs, session metadata, and cross-session information for analysis.Records included queries, UI interactions, rule edits, prompts, responses, chosen layouts, and generation durations.
6 User Study Results
Across 36 sessions, Maru’s persistent IA rules generally improved alignment, reduced repeated specification, and supported individualized layout trajectories, while over-persistence and rule accumulation created failures.
- Alignment and efficiency: Maru maintained approval more consistently across sessions: rates shifted from 74% to 61%, whereas baseline approval fell from 71% to 33%.The baseline decline was statistically significant, while Maru’s change was not.
- Alignment and efficiency: Persistent rules reduced the need to restate structural preferences, shortening Maru prompts by 10.6 words on average as rules carried context forward.Baseline prompts declined by only 2.5 words, with participants sometimes copy-pasting queries or avoiding new generations.
- User engagement: Users mostly formed IA rules through ordinary interaction and query behavior, with 97% of user-created rules produced without deliberate schema editing.Of 838 user-attributable rules, 76% originated from UI interactions, 20% from query extraction, and 3% from panel editing.
- User engagement: Users’ intuitive understanding and actual behavior diverged: vocabulary and order felt most natural, but partition and hierarchy dominated created rules.Partition accounted for 43% and hierarchy 26% of user-created rules, while vocabulary accounted for 21% and order 10%.
- Layout personalization: Maru preserved individualized layout diversity: career participants ended on four layout types, while all baseline participants converged on tables.For picnic planning, Maru produced six distinct final layouts versus four in the baseline.
- Layout personalization: IA rule dimensions mapped to layout types, with rule changes preceding 29 of 45 observed Maru layout transitions.Partition additions preceded grouped layouts in 16 cases, order additions preceded sequenced layouts in 7, and hierarchy accumulation preceded tables in 6.
- Layout personalization: Personal rule profiles produced distinct trajectories within identical tasks, from timeline layouts driven by temporal ordering to tabs shaped by partition exclusions.In Barcelona planning, richer rule profiles supported progression across accordion, diagram, and timeline layouts, while minimal profiles remained on map+tabs.
7 Discussion
The discussion argues for persistence that operates across UI abstraction levels and task contexts, while recognizing that persistence must be bounded. Maru’s IA layer supports interaction-based customization and user-specific UI trajectories, but accumulated rules can outlive their relevance.
- Designing for Persistence Across Levels and Contexts in GenUI: Maru addresses IA-level structural persistence, while users with design experience sought finer-grained control over layout type and component structure.The paper proposes a layered persistence model spanning high-level structural rules to component-level handles.
- Designing for Persistence Across Levels and Contexts in GenUI: Rules can bleed across sub-task boundaries, so future systems should let them expire or remain scoped to relevant sub-task contexts.Suggested mechanisms include task schemas and decay over accumulated context.
- Designing for Persistence Across Levels and Contexts in GenUI: Maru’s concrete IA rules were difficult to reuse across task instances without additional user observation and explicit generalization.What transferred across tasks was a general sense of user priorities rather than rules directly mapping to UI decisions.
- Discussion: IA rules supported low-cost customization because 97% were created without deliberate engagement with the IA schema.Natural language and interaction served as the medium through which users communicated customization intent.
- Discussion: Users increasingly specified preferences deliberately after seeing that Maru would carry them forward as accumulated rules.The observed shift suggests users became more comfortable communicating through IA rules over time.
- Discussion: Users performing the same task followed distinct and valid UI trajectories, motivating evaluation of how user preferences shape the design space rather than seeking one optimal layout.The discussion reframes GenUI research around conditions that support preference-shaped layouts.
- Discussion: The proposed GenUI stack places IA between task-driven data models and UI specification or malleable interface layers.IA is described as translating structural information needs into interface structure.
A.1.1 Full Set of Literature Papers for Each Round
Table A.1 records the literature coding process across rounds.
- A.1.1 Full Set of Literature Papers for Each Round: The appendix includes a table documenting the literature coding process across rounds.The supplied passage provides the table title but no round-level contents.
A.1.2 Finalized Literature Codebook
Table A.2 presents the finalized codebook mapping IA elements to user needs, UI components, and interactions across coded systems.
- A.1.2 Finalized Literature Codebook: The finalized literature codebook maps IA elements to user needs, UI components, and interactions across coded systems.The supplied passage identifies the table’s scope but does not list individual mappings.
A.2 Maru Details
The appendix describes Maru’s interface, supported layouts and interactions, study procedure, tasks, surveys, and the pipeline distinction between baseline and Maru. Maru represents the same interactions as typed IA rules rather than a flat timestamped log.
- Interface Overview: Maru combines a chat panel, generated UI, and IA panel to connect natural-language queries, interactive layouts, and accumulated IA rules.The IA panel visualizes accumulated state across tabs, while detected rules appear inline as tagged labels.
- Supported Components: Maru supports 16 composable layout types, including combinations such as map + tabs and tabs containing card-grid.The layout types are listed in Table A.3.
- Supported Components: The system includes 10 supported interaction types mapped to IA rule dimensions.These mappings are summarized in Table A.4.
- Study Procedure: Each study session lasted approximately 120 minutes and included two information-intensive tasks, post-session surveys, and a final survey after a freeform session.The surveys included NASA-TLX and custom structural-alignment items.
- Study Tasks: Participants compared graduate programs or jobs in one task and planned an outdoor picnic for about 15 people in the other.The tasks required comparison or planning outputs including priorities, timelines, lists, and budgets.
- Experimental Conditions: The baseline and Maru shared data acquisition, rendering, and LLM components, differing in whether interactions were passed as a flat log or typed IA rules.The same user interactions were supplied to both conditions, but Maru encoded them as structured preferences.
- Experimental Conditions: Maru’s structured prompt represented ordering, partitions, hierarchy relationships, and consistent field names as explicit IA preferences.The example encodes price prioritization, item groupings, exclusion, and field vocabulary.