ReLayout: Structural Adaptation in Layout Systems

Updated 6 July 2026

ReLayout is a polysemous term that denotes methods for structure-preserving modifications in graphic design, document analysis, and expert systems.
It employs techniques such as relation graphs, region annotations, and prototype balancing to maintain spatial and relational integrity during layout modifications.
Empirical studies reveal significant gains and practical challenges, highlighting its impact on both design fidelity and system-level expert allocation.

Searching arXiv for recent uses of “ReLayout” and closely related layout/relayout papers. ReLayout is a research term used across several technical literatures to denote methods for modifying an existing arrangement while preserving some notion of structure. In recent arXiv work, the name has been attached to at least four distinct lines of research: structure-preserving design layout editing in graphic design, relation-reasoned content-aware layout generation, layout-enhanced pre-training for real-world visually-rich document understanding, and adaptive expert relayout in Mixture-of-Experts systems under expert parallelism (Lin et al., 1 Feb 2026, Tian et al., 8 Jul 2025, Jiang et al., 2024, Ma et al., 20 May 2026). Across these usages, the common theme is not a single algorithmic family but the attempt to change placement, grouping, or allocation without discarding higher-level relational organization.

1. Terminological scope

In the current literature, “ReLayout” is best treated as a polysemous technical term rather than a single canonical method. The table below summarizes the principal usages that explicitly bear the name.

Domain	Meaning of “ReLayout”	Representative paper
Graphic design editing	Structure-preserving editing of an existing design under operations such as move, resize, add, and delete	"ReLayout: Versatile and Structure-Preserving Design Layout Editing via Relation-Aware Design Reconstruction" (Lin et al., 1 Feb 2026)
Content-aware graphic layout generation	Relation-reasoned generation that decomposes layout into saliency-, region-, and margin-aware structure	"ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal LLMs" (Tian et al., 8 Jul 2025)
Document understanding	Layout-enhanced pre-training for real-world VrDU without manually annotated semantic groups	"ReLayout: Towards Real-World Document Understanding via Layout-enhanced Pre-training" (Jiang et al., 2024)
MoE systems	Adaptive expert relayout: changing which experts live on which ranks, or reassigning expert placement in response to load	"Diagnosing Overhead in Dispatch Operations: Cross-architecture Observatory" (Ma et al., 20 May 2026)

This multiplicity matters because the same word denotes different objects of manipulation. In design papers, the manipulated entity is a spatial composition of elements on a canvas. In document understanding, it is the latent grouping structure of OCR tokens. In MoE systems, it is expert placement across ranks or memory devices. The term therefore names a class of structural adaptation problems, not a uniform research program.

2. ReLayout in design layout editing

The most literal use of the term appears in design editing. "ReLayout: Versatile and Structure-Preserving Design Layout Editing via Relation-Aware Design Reconstruction" formulates the task as $f:(D_{in},O)\rightarrow D_{out}$ , where $D_{in}$ is the original design, $O$ is an editing operation, and $D_{out}$ is the edited design. The framework standardizes four basic operations—move, resize, add, and delete—and requires that the resulting design satisfy the edit while preserving the layout structure of unedited elements (Lin et al., 1 Feb 2026).

Its central structural device is the relation graph $G:=(V,R)$ . Nodes include all elements plus the canvas as an extra node, and edges encode 3 size relation types and 9 position relation types. Size relations are derived from area ratio, while position relations are defined by a $3\times3$ partition either relative to another element or to the canvas. The model removes graph edges incident to the target element, because those relations may legitimately change after editing. This makes the graph an explicit preservation prior over unedited content rather than a rigid global template (Lin et al., 1 Feb 2026).

The second major contribution is relation-aware design reconstruction (RADR), a self-supervised surrogate for missing triplet supervision. Instead of training on $(\text{original}, \text{operation}, \text{edited})$ triplets, RADR reconstructs a design from element content $C$ , relation graph $G$ , and a synthesized editing operation $O$ , learning the approximate mapping $D_{in}$ 0, where $D_{in}$ 1 is the attribute sequence. The backbone is a multimodal LLM with Llama-3.1-8B as the LLM backbone and CLIP ViT-L/14 as the vision encoder, with a two-layer MLP projector and LoRA fine-tuning (Lin et al., 1 Feb 2026).

The paper reports strong gains over GPT-4o and generation baselines. In the reconstruction setting, ReLayout reaches Size Rel 0.9150, Pos Rel 0.8684, and Op 0.9991; in the generalization setting, it reports Size Rel 0.9475, Pos Rel 0.9157, and Op 0.9960. A user study further reports 79.5% preference for visual appeal and 86.0% preference for structure preservation in reconstruction, with similarly strong results in generalization (Lin et al., 1 Feb 2026).

A closely related but differently named design line appears in "Aggregated Structural Representation with LLMs for Human-Centric Layout Generation" (Jin et al., 26 May 2025). That work does not use the ReLayout name, but it is structurally adjacent: it introduces an editable relation matrix over CONTAIN, PARALLEL, TOP, and LEFT relations, feeds graph-derived structure into Intern-VL, and reports large gains from the relation matrix, including an improvement from Max-IoU 0.27 / Relation Error 0.61 without RM to 0.54 / 0.30 with RM (Jin et al., 26 May 2025). Together, these papers establish relation-level control as a central design principle for structure-preserving layout editing.

3. ReLayout in content-aware graphic layout generation

A second design-oriented usage appears in "ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal LLMs" (Tian et al., 8 Jul 2025). Here the task is not editing an existing design but generating a new content-aware arrangement on a given canvas image. A layout is represented as

$D_{in}$ 2

with $D_{in}$ 3, and the multimodal input consists of the canvas image plus optional text and element images.

The paper’s key claim is that prior MLLM methods generate mostly at the level of individual coordinates and therefore miss the structural middle layer designers use. ReLayout addresses this with relation-CoT, which augments layout annotations with region, saliency, and margin information, as well as a specialized parallel relation. Regions are defined as

$D_{in}$ 4

and are recursively inferred by projecting boxes to the $D_{in}$ 5- and $D_{in}$ 6-axes, grouping overlaps, and estimating dominant direction. The resulting representation decomposes layouts into nested sub-layouts rather than flat element lists (Tian et al., 8 Jul 2025).

The second component is the layout prototype rebalance sampler. It defines prototype features over three dimensions—saliency, region, and element—and concatenates them as

$D_{in}$ 7

Layouts are then clustered with K-means using $D_{in}$ 8, and cluster-level sampling weights are computed from cluster counts via a temperature-like parameter $D_{in}$ 9. The paper reports that $O$ 0 works best, arguing that smaller values preserve imbalance and larger values oversample rare prototypes (Tian et al., 8 Jul 2025).

The model is fine-tuned on InternVL2.5-8B and evaluated on PKU and CGL. The paper emphasizes improvements in maximum overlap and distributional quality. On the PKU hard split, ReLayout reports Ove 0.0109 versus 0.0318 for PosterLlama; on the CGL hard split it reports Ove 0.0117 versus 0.0183. A professional designer study reports $O$ 1 and $O$ 2 for ReLayout, and a diversity study reports the highest score among compared methods (Tian et al., 8 Jul 2025).

Methodologically adjacent work includes "LayoutRAG: Retrieval-Augmented Model for Content-agnostic Conditional Layout Generation" (Wu et al., 3 Jun 2025) and "Constrained Graphic Layout Generation via Latent Optimization" (Kikuchi et al., 2021). LayoutRAG retrieves reference layouts using category-count filtering and bipartite IoU matching, then injects them through condition-modulated attention into a flow-matching generator, achieving strong gains in under-specified conditional generation (Wu et al., 3 Jun 2025). The latent-optimization approach instead keeps a pretrained generator fixed and solves

$O$ 3

enforcing relations such as alignment, overlap avoidance, above/below, and size comparisons through augmented Lagrangian optimization (Kikuchi et al., 2021). These neighboring methods clarify that the distinctive feature of ReLayout (Tian et al., 8 Jul 2025) is explicit relation annotation and prototype balancing inside an MLLM training pipeline.

4. ReLayout in real-world document understanding

In document AI, "ReLayout: Towards Real-World Document Understanding via Layout-enhanced Pre-training" introduces ReVrDU, a reformulation of visually-rich document understanding that disallows manually annotated semantic groups, and a corresponding pre-training method called ReLayout (Jiang et al., 2024). The paper’s premise is that standard VrDU pipelines rely on semantic groups $O$ 4 that are manually annotated but unavailable from ordinary OCR systems. ReLayout therefore works with OCR words $O$ 5, token positions, word boxes, and OCR text segments $O$ 6 instead.

Architecturally, the model is close to RoBERTa plus a 2D embedding layer. Token embeddings are

$O$ 7

where text, 1D order, and 2D box embeddings are summed. The main novelty lies in pre-training. ReLayout uses three objectives: MLM, 1D Local Order Prediction (1-LOP), and 2D Text Segment Clustering (2-TSC). The total loss is

$O$ 8

with $O$ 9, $D_{out}$ 0, $D_{out}$ 1, $D_{out}$ 2, and $D_{out}$ 3 (Jiang et al., 2024).

The 2-TSC objective is the part most directly tied to the ReLayout name. It constructs candidate pairs of OCR segments that are spatially close and already semantically similar, then encourages their representations to become closer using a SimSiam-style objective: $D_{out}$ 4 This does not use negative pairs and is only applied in the final epoch of pre-training (Jiang et al., 2024).

Empirically, the paper’s most important result is not just absolute accuracy but robustness when manually annotated semantic groups are removed. ReLayout $D_{out}$ 5 reports FUNSD 84.64, CORD 96.82, and DocVQA 76.02; ReLayout $D_{out}$ 6 reports 86.11, 97.42, and 80.14. The paper further notes that methods dependent on semantic-group annotations deteriorate sharply under OCR-only conditions, while ReLayout remains comparatively stable (Jiang et al., 2024). In this usage, “ReLayout” refers less to editing than to learning latent grouping structure from layout itself.

5. ReLayout in Mixture-of-Experts systems

In systems research, “ReLayout” denotes adaptive expert relayout in Mixture-of-Experts expert parallelism: changing which experts live on which ranks, or reassigning expert placement in response to observed load. "Diagnosing Overhead in Dispatch Operations: Cross-architecture Observatory" studies this family critically and argues that its effectiveness is often overstated (Ma et al., 20 May 2026).

The paper formalizes expert-parallel dispatch with a send-count matrix $D_{out}$ 7, where $D_{out}$ 8 counts tokens sent from rank $D_{out}$ 9 to rank $G:=(V,R)$ 0. If $G:=(V,R)$ 1, then AlltoAll completion tracks the maximum receive load rather than the mean. Against that background, ReLayout is one of four mitigation families, alongside predictive sample placement, hierarchical collectives, and EP-aware topology (Ma et al., 20 May 2026).

The main negative result is that placement-only relayout under fixed routing has limited power. Using DODOCO across five MoE checkpoints and EP ranges up to 32, the paper reports that scaling EP changes the per-expert max/mean token ratio by at most 5% within each architecture’s measurable range: 5.0% for MLA, 2.8% for MHA, 4.4% for GQA, 4.4% for Mamba-2, and 0.3% for GDN. It also reports that mock tokens overestimate routing Gini by up to 2.35× (Ma et al., 20 May 2026). The central conclusion is that the dominant straggler is “intrinsic to the routing decision the model makes, not to how its experts land on ranks,” which sharply limits what expert-to-rank relayout alone can fix (Ma et al., 20 May 2026).

A more specialized positive systems use of the term appears in "TriMoE: Augmenting GPU with AMX-Enabled CPU and DIMM-NDP for High-Throughput MoE Inference via Offloading" (Pan et al., 1 Mar 2026). There, “Prediction-Driven Expert Relayout and Rebalancing” means converting expert weight layout between striped and localized forms, migrating cold experts across DIMMs, and prefetching predicted hot experts to GPU HBM. The predictor is a per-expert EMA,

$G:=(V,R)$ 2

and the paper reports over 78% accuracy, 38 KB metadata overhead, and <3.3% online migration overhead. In ablation, the relayout/rebalancing mechanism contributes an additional $G:=(V,R)$ 3 improvement (Pan et al., 1 Mar 2026).

Taken together, these papers show that “ReLayout” in MoE research is a systems-layer placement concept, but its value depends strongly on whether it merely remaps experts under fixed routing or actually changes effective token-to-expert assignment or device-locality costs.

Several adjacent literatures do not use the ReLayout name but address closely related problems. "Interactively Optimizing Layout Transfer for Vector Graphics" treats relayout as layout transfer from a source design $G:=(V,R)$ 4 to a target design $G:=(V,R)$ 5, using correspondence $G:=(V,R)$ 6, inferred semantic rules, and interactive controls such as global layout copy, element layout copy, and individual rule adherence (Warner et al., 2023). "SMT-Layout: A MaxSMT-based Approach Supporting Real-time Interaction of Real-world GUI Layout" frames responsive GUI relayout as a Boolean-and-arithmetic constraint problem over widget geometry and visibility, reporting millisecond-level interaction on real-world layouts (Li et al., 2024).

In more formal geometric settings, "Area-Universal Rectangular Layouts" studies when a rectangular arrangement can be reshaped to fit arbitrary target rectangle areas while preserving combinatorial structure, proving that a rectangular layout is area-universal iff it is one-sided (0901.3924). For scene-graph editing, "3D-Layout-R1: Structured Reasoning for Language-Instructed Spatial Editing" uses scene-graph transformations and structured reasoning traces for 3D layout editing, with reported average gains of 15% in IoU and 25% reduction in center-distance error over CoT-SFT and vanilla GRPO baselines (Zhen et al., 23 Mar 2026). For responsive web repair, "Repairing Responsive Layout Failures Using Retrieval Augmented Generation" uses localized CSS repair rather than holistic regeneration, reporting 38 repaired out of 43 RLFs, or 88.3% accuracy (Zerin et al., 1 Nov 2025).

More generative neighbors include "LayoutTransformer: Layout Generation and Completion with Self-attention," which models layouts autoregressively as sequences of primitive attributes and supports completion from partial layouts (Gupta et al., 2020). These neighboring works make clear that contemporary relayout research spans at least four paradigms: structured editing, conditional generation, constraint solving, and systems reallocation.

7. Recurring technical themes

Despite the diversity of domains, several motifs recur. The first is explicit relational structure. Design editing uses relation graphs over size and position (Lin et al., 1 Feb 2026); content-aware generation uses regions, saliency, margins, and parallel groups (Tian et al., 8 Jul 2025); document understanding uses OCR segments plus local-order and segment-clustering objectives to recover latent grouping (Jiang et al., 2024). This suggests that direct coordinate prediction or flat sequence modeling is often regarded as insufficient when preservation of structure is central.

The second is decomposition into intermediate representations. RADR reconstructs from $G:=(V,R)$ 7 rather than raw triplets (Lin et al., 1 Feb 2026). Relation-CoT recursively decomposes layouts into regions before coordinates (Tian et al., 8 Jul 2025). ReLayout in document understanding introduces 1-LOP and 2-TSC precisely to model intermediate grouping structure (Jiang et al., 2024). In each case, relayout is treated not as a single step from input to output but as a structured latent process.

The third is sensitivity to evaluation realism. The document-understanding paper argues that manually annotated semantic groups make VrDU unrealistic (Jiang et al., 2024). The MoE observatory paper argues that mock-token benchmarks exaggerate routing imbalance and can mislead ReLayout-style system claims (Ma et al., 20 May 2026). These critiques converge on a methodological point: structural adaptation methods are especially vulnerable to optimistic results when their benchmark representation already supplies hidden structure or synthetic artifacts.

A final recurring issue is that preservation is usually soft rather than guaranteed. ReLayout for design editing conditions on relation graphs but still reports residual overlap and relation failures as element counts grow (Lin et al., 1 Feb 2026). Content-aware ReLayout improves overlap and FD but still depends on imperfect evaluation metrics and backbone choices (Tian et al., 8 Jul 2025). MoE relayout can improve locality or balancing in narrower settings, yet placement-only remapping does not generally erase routing hotspots (Ma et al., 20 May 2026). The literature therefore presents ReLayout less as a solved primitive than as a family of structure-aware approximations whose success depends on how structure is represented, what is held fixed, and which aspects of change are actually controllable.