Maru, a conversational system that records how information is divided into groups, placed in a hierarchy, ordered and named, had higher odds of being accepted than an otherwise similar baseline in the study. Overall, the odds of acceptance were 2.08 times higher with Maru (95% CI [1.25, 3.47], p = .005). After the analysis adjusted for prompt tokens used in each generation, the odds ratio was 2.40 (95% CI [1.28, 4.47], p = .006).
That result should be read as an association, not proof that storing rules caused the higher approval. Participant-level randomization was not reported. The main outcome was also subjective: participants accepted or rejected each generated interface, rather than being assessed on objective task performance.
A steadier approval pattern
The difference emerged most clearly over time. In the baseline condition, approval fell from 71% in the first half of sessions to 33% in the second half. Maru approval moved from 74% to 61%. The baseline trajectory produced an odds ratio of 0.19 (95% CI [0.08, 0.45], p < .001), while Maru's trajectory was not statistically clear (OR = 0.63, 95% CI [0.15, 2.65], p = .53).
Maru also reached an output judged satisfactory in fewer generations: an average of 5.58, compared with 7.75 for baseline (standard deviations 1.62 and 2.09, respectively; Wilcoxon signed-rank p = .008). In this study, that was a measure of how quickly a session reached an accepted result, not evidence of better performance on the underlying task.
How Maru stored preferences
The organizing scheme behind Maru was built from a three-phase review of 42 papers. Two coders analyzed 12 seed papers, followed by eight snowball-sampled papers, and one coder applied the codebook to 22 remaining papers. The analysis produced four information-architecture elements: partition, hierarchy, order and vocabulary. The first two were structural elements, while the last two were semantic elements.
Maru's information-architecture layer was implemented independently of its interface components, allowing other generative user-interface systems to consume the rule schema. The system detects, stores, retrieves and reapplies rules across generations through queries, UI interactions and panel editing. The baseline was an IA-layer ablation. It used the same base context and core data acquisition, rendering pipeline, LLM model, layouts, interactions and queries, but represented earlier interactions as an unstructured log rather than typed IA rules.
Across 36 sessions, participants generated 220 interfaces, made 585 interactions and constructed 1,292 IA rules. Of those rules, 838 were attributed to users and 454 were scaffolded by the system. Nine layout types appeared across the sessions, and Maru averaged 2.7 distinct types per session. Median prompt lengths were close: 3.4k tokens per model call in baseline and 3.3k in Maru.
Users generally expressed preferences through ordinary use rather than a dedicated schema editor. Among the 838 user-attributable rules, 76% came from interface interactions, 20% from natural-language query extraction and 3% from manual panel editing. The paper reports that 97% were built without deliberate schema engagement. Of 94 rules that received explicit feedback, 88 were accepted, a 94% acceptance rate (95% CI [87%, 97%]).
Persistence brought trade-offs
Persistence could also carry the wrong preference into a new context. During subtask transitions, cross-topic rule retrieval received a thumbs-down 23% of the time, compared with 18% for on-topic retrieval. The comparison was descriptive and did not isolate a causal reason for the difference.
Meanwhile, the active rule set grew quickly. Mean active-rule count rose from 15.2 in the first quarter of a session to 37.5, 59.1 and 101.6 in later quarters. System scaffolding added about 24 hierarchy rules per generation, and user refinement rules were never pruned, merged or expired. The authors describe this accumulation as a partial contributor to declining alignment, without reporting an independent causal estimate.
That creates a practical tension for Maru: persistence can preserve a choice, but it can also leave an irrelevant rule in force when the work shifts. The authors point to task-aware scoping, expiration or decay, and layered persistence across structural and component levels as open design needs.
Different rules, different pages
On the page, Maru preserved more individual variation. In the career-exploration task, all six baseline participants ended with tables, while the six Maru participants ended with four distinct layout types. In picnic planning, Maru produced six final layout types, compared with four in baseline. The paper presents layout diversity as a design outcome, not a direct measure of task success.
Several rule states were associated with particular layouts. Grouped layouts occurred in 33% of cases with partition rules and 0% of cases without them. Tables occurred in 10% of cases with five or more hierarchy rules and 0% with fewer. Table or list layouts occurred in 17% of cases with order rules, compared with 5% without them. Of 45 Maru layout transitions, 29, or 64%, were linked in time to a new rule, while average section count stayed between 1.2 and 1.6.
Users did not report a clear increase in control. Perceived-control ratings averaged 3.25 out of 5 with Maru and 3.00 with baseline, a difference of 0.25 points that was not statistically significant (p = .555).
A narrow test
The study involved 12 participants, all daily or near-daily LLM users, each completing three sessions. The first two used Maru or the baseline on assigned graduate-program/career-exploration and picnic-planning tasks, with system order counterbalanced. The third used Maru for a self-selected real task, so it did not provide a comparative condition.
The sample was small and recruited through an institutional online community. The study tested structural IA persistence, not visual or component-level persistence, and it did not establish transfer across separate task instances or longer-term use. These boundaries, along with the subjective approval measure and unpruned rule store, limit the findings to the tested prototype, participants and information-intensive tasks.
The document is a preprint identified as arXiv:2608.25565v1 in cs.HC and dated 26 Aug 2026. The paper says the Maru implementation and IA layer are available in an online repository. It reports support from an IITP grant, No. 2021-0-01347, and an NRF grant, No. RS-2024-00406715, both funded by the Korea government's MSIT.
Paper data and sources
Original title: Maru: Information Architecture as a Shared Language for Generating Aligned and Persistent User Interfaces
Authors: Eunhye Kim, DaEun Choi, Bryan Min et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: 10.1145/3830398.3830621
Original paper · Full text