FlexDoc: Flexible Document Technology

• FlexDoc is a multi-faceted concept encompassing fluid document formats, adaptive layouts, and synthetic data generation for controllable variability in document technology.
• One variant employs a fluid-document metamodel using the RSL hypermedia framework to manage distribution, access rights, and context-driven rewriting.
• Another variant optimizes both content and layout via continuous and discrete optimization, while a third enhances document AI training with scalable synthetic data generation.

FlexDoc is a recurring name in document-technology research rather than a single canonical system. In the arXiv record, it denotes at least three technically distinct frameworks: a fluid-document format and on-line editor prototype built on the Resource–Selector–Link (RSL) hypermedia metamodel; a framework for adaptive documents that optimizes both content and layout during consumption; and a scalable synthetic data generation framework for multilingual semi-structured documents used to train document understanding models (Tayeh et al., 2021, Jiang et al., 2024, Dua et al., 2 Oct 2025).

1. Terminological scope

The published literature uses the name “FlexDoc” for multiple lines of work with different immediate objectives.

Variant	Problem domain	Core description
FlexDoc as fluid documents	Hypermedia and ubiquitous computing	A fluid-document format and on-line editor prototype built on the RSL hypermedia metamodel
FlexDoc as adaptive rendering	Cross-device document consumption	A framework for creating and consuming documents that seamlessly adapt to different devices, author, and viewer preferences and interactions
FlexDoc as synthetic data generation	Document AI training data	A scalable synthetic data generation framework that combines Stochastic Schemas and Parameterized Sampling to produce realistic, multilingual semi-structured documents with rich annotations

This polysemy is consequential. Two variants are primarily concerned with authoring and consumption, while another targets data generation for enterprise-scale document understanding. A common thread is controlled variability: in one case through hypermedia resources and selectors, in another through optimization over layout templates and content alternatives, and in another through probabilistic sampling over document structure and content (Tayeh et al., 2021, Jiang et al., 2024, Dua et al., 2 Oct 2025).

2. FlexDoc as a fluid-document metamodel

In "A Metamodel and Prototype for Fluid Document Formats" (Tayeh et al., 2021), FlexDoc is a fluid-document format and on-line editor prototype built on top of the Resource–Selector–Link hypermedia metamodel. The RSL metamodel, associated in the description with Signer & Norrie 2007, is organized around three core entities: Resource, an addressable unit of content; Selector, a way to point into a resource; and Link, a bi- or multidirectional association between resources and/or selectors. The FlexDoc formalization extends this with adaptation rules:

$\mathit{RSL} = (R,\,S,\,L,\,A).$

Each selector is defined as $s=(r,\sigma)$ with $r\in R$, each link as

$\ell = \bigl(E,\,\mathit{type},\,\alpha\bigr),$

and each adaptation rule as a function

$$a : (\mathit{Context}\times (R\cup S\cup L)) \to 2^{R\cup S\cup L}.$$

Graph-theoretically, the model is viewed as a directed hypergraph

$G = (V,\,E), \quad V = R\cup S,$

subject to the consistency constraints of no dangling selectors, existing link endpoints, rights coherence, and adaptation termination.

Within this formulation, distribution, user rights management, and adaptation are not peripheral features. Each Resource and Link can be stored, cached or served independently via RESTful URIs; each entity carries an access-control policy; and “context resolvers” and “property” components allow rules to rewrite or filter parts of a resource before rendering. The description states explicitly that no additions to the core RSL are required for fluid-document requirements such as scattering chapters over multiple servers, enforcing per-paragraph rights, or re-flowing layout to smartphones.

The associated FlexDoc prototype, linked in the description to Tayeh et al. 2018, comprises a Repository & REST API, Link Manager, Rights Manager, Adaptation Service, Rendering Engine, and Editor UI. The workflow is defined procedurally: a user requests a top-level Resource $r_0$; the repository returns an RSL subgraph; the Rights Manager prunes unauthorized fragments; the Adaptation Service applies rules $A$ given the device context $\gamma$; and the Rendering Engine lays out the adapted subgraph. On context change, the pruning and adaptation stages repeat automatically.

An illustrative example in the description uses a paragraph resource with a full-text selector and a summary selector, where mobile context renders only the summary while desktop context renders the full selection. The significance of this variant lies in its treatment of fluidity as a property of the underlying hypermedia graph rather than a mere front-end rendering tactic. Distribution, transclusion, provenance preservation, and fine-grained rights are part of the document model itself.

3. FlexDoc as adaptive content-and-layout optimization

In "FlexDoc: Flexible Document Adaptation through Optimizing both Content and Layout" (Jiang et al., 2024), FlexDoc is a framework for creating and consuming documents that adapt to different devices, author, and viewer preferences and interactions. The authoring side is organized around a tabstop-based editor. Authors import a document, place vertical and horizontal tabstops on a canvas, assign document elements to areas defined by tabstops, and provide multiple content alternatives for each element, such as full-length text and shortened summaries or image versions retargeted at different aspect ratios. Alternatives are ranked by the author, as are multiple layout templates. The editor exports a JSON template containing tabstop positions, alternatives and their ranks, the ranked template list, and default author preference weights.

The consumer side loads this JSON together with static assets and monitors device or window size, viewer-supplied preference sliders, and direct viewer interactions such as zoom-in, zoom-out, pin, switch element, or switch template. Whenever these inputs change, the engine re-runs a joint discrete-plus-continuous optimizer. For $N$ document elements, the continuous decision variables are

$s=(r,\sigma)$0

while the discrete decisions are the layout template $s=(r,\sigma)$1 and the content alternative $s=(r,\sigma)$2 for each element. The optimization objective is

$s=(r,\sigma)$3

with

$s=(r,\sigma)$4

and

$s=(r,\sigma)$5

The description specifies that exactly one of author, viewer, or interaction has weight $s=(r,\sigma)$6 for each element. The image term uses preferred image dimensions and an aspect-ratio-preservation term; the text term combines a preferred-font-size penalty with a subtraction of $s=(r,\sigma)$7 when a shortened summary is chosen; and the alignment term penalizes departures from specified alignments, such as common row centers.

The solver is described as alternating and branch-and-bound style. It initializes with the highest-rank template and rank-1 alternatives, then iterates through three phases: continuous optimization of positions and sizes under tabstop and geometry constraints; discrete re-ranking of element alternatives; and template switching if another template lowers the objective. Convergence is reached when no discrete choice changes and the gradient step fails to reduce $s=(r,\sigma)$8 further.

Two content-adaptation components are explicit. First, the framework adopts a fine-tuned BERT-based extractive summarizer, attributed in the description to Miller 2019, generating summaries at multiple compression ratios such as 25%, 50%, and 75% of original length. Second, it includes seam carving, attributed to Avidan & Shamir 2007, as a plugin for image retargeting; a small set of retargeted image variants is precomputed and treated as alternatives.

The reported applications include news front pages, scientific papers, and academic personal websites. The evaluation description includes qualitative user feedback, task completion time, and re-optimization latency, as well as a within-subject user study with 13 participants, of whom 10 were professional UI designers and 3 were graduate students. Reported findings include positive reactions to dynamic summarization and improved readability on mobile, as well as reduced manual breakpoint coding. A stated pain point was the lack of live preview of all possible sizes, and tabstop editing was described as somewhat technical by less-experienced designers. Runtime is described as a few tens of milliseconds for the continuous solve for approximately 20 elements, less than 1 ms for discrete sweeps, and total re-optimization under 0.5 s on commodity laptop hardware.

4. FlexDoc as synthetic document generation for document understanding

In "FlexDoc: Parameterized Sampling for Diverse Multilingual Synthetic Documents for Training Document Understanding Models" (Dua et al., 2 Oct 2025), the term denotes a scalable synthetic data generation framework for enterprise document AI rather than a rendering or authoring system. The problem statement is operational: developing document understanding models at enterprise scale requires large, diverse, and well-annotated datasets spanning a wide range of document types, but collecting such data is prohibitively expensive due to privacy constraints, legal restrictions, and the sheer volume of manual annotation needed, with costs that can scale into millions of dollars.

The framework combines Stochastic Schemas and Parameterized Sampling to produce realistic, multilingual semi-structured documents with rich annotations. The abstract states that FlexDoc probabilistically models layout patterns, visual structure, and content variability, enabling the controlled generation of diverse document variants at scale. Because only the abstract-level description is available here, the internal form of the stochastic schema and the exact sampling procedure are not further specified.

The reported empirical setting is Key Information Extraction. The abstract states that FlexDoc-generated data improves the absolute F1 Score by up to 11% when used to augment real datasets, while reducing annotation effort by over 90% compared to traditional hard-template methods. It also states that the solution is in active deployment and has accelerated the development of enterprise-grade document understanding models while significantly reducing data acquisition and annotation costs.

This variant of FlexDoc shifts the meaning of flexibility from adaptive presentation to controllable data diversity. The operative object is not the final document delivered to a reader, but the synthetic corpus used to train downstream models.

5. Adjacent formalisms and “FlexDoc-style” systems

The broader literature includes systems that do not carry the exact FlexDoc name in their titles but are presented in the supplied descriptions as satisfying or exemplifying comparable goals. "S2Doc -- Spatial-Semantic Document Format" (Kempf et al., 2 Nov 2025) introduces a flexible data structure for modeling documents and tables that combines both spatial and semantic information in a single format. Its abstract emphasizes the absence of a common understanding of how to model documents and tables, the incompatibility of existing approaches, and the tendency of most data models to privilege either spatial or semantic structure. S2Doc addresses this by supporting both within one format, being easily extendable to new tasks, and supporting most modeling approaches for documents and tables, including multi-page documents. In the supplied technical description, this is operationalized through a top-level Document object, Page and Space objects, Elements with Regions and free-form data_dict, a directed acyclic ReferenceGraph, and a SemanticKnowledgeGraph plus SemanticReferenceGraph. The same description concludes that S2Doc fulfills the requirements of a “FlexDoc” solution for modern document understanding.

"Towards Flexible Multi-modal Document Models" (Inoue et al., 2023) presents FlexDM, a unified model for vector-graphic documents represented as unordered sets of multi-modal elements. Each element contains categorical variables, numerical variables, and dense feature vectors, with training based on masked-field prediction over fields such as element type, position, styling attributes, image, and text. The model uses an encoder–transformer–decoder architecture, in-domain pre-training with 15% random masking, and explicit multi-task learning. The description states that the resulting single model can solve a multitude of different design tasks and that FlexDM exemplifies “FlexDoc” by specializing to many design-automation tasks through masking patterns without adding per-task parameters.

Taken together, these adjacent systems indicate that flexibility in document research spans at least four axes: spatial-semantic representation, multi-modal field completion, adaptive human-facing rendering, and synthetic data generation. This suggests a broader research landscape in which “FlexDoc” functions less as a single artifact than as a family of approaches for controllable variability in document structure, semantics, presentation, and training data.

6. Conceptual significance and recurrent misconceptions

A common misconception is that FlexDoc refers to a standardized document format. The arXiv record instead applies the name to multiple, technically distinct systems: an RSL-based fluid-document metamodel and editor, an optimization-based adaptive rendering framework, and a synthetic-document generation pipeline for document understanding (Tayeh et al., 2021, Jiang et al., 2024, Dua et al., 2 Oct 2025). The ambiguity is not merely terminological; it corresponds to different units of analysis, ranging from hypermedia entities and selectors, to tabstop-constrained layout variables, to probabilistically generated training instances.

A second misconception is that flexibility denotes only responsive reflow. In the RSL-based variant, flexibility includes distribution, user rights management, transclusion, and context-driven rewriting. In the optimization-based variant, it includes ranked templates, automatic summarization, seam-carved image alternatives, and direct viewer interaction. In the synthetic-data variant, it includes controlled generation of multilingual semi-structured documents with rich annotations. These are materially different technical commitments.

The literature also points to a continuing standardization problem. S2Doc’s abstract states that there is no real common understanding of how to model documents and tables in particular, and that existing approaches often produce incompatible data structures while emphasizing only one aspect of document structure (Kempf et al., 2 Nov 2025). A plausible implication is that FlexDoc should be understood not as the endpoint of standardization but as evidence of an active design space: graph-based fluid documents, optimization-based adaptive documents, and schema-driven synthetic corpora each solve different parts of the document-technology stack.

In that sense, FlexDoc names a recurring research ambition: to move beyond static, monolithic documents toward systems in which structure, content, presentation, semantics, or annotation can be varied in a controlled way. The specific mechanism differs sharply across papers, but the underlying objective is consistent—documents should be manipulable artifacts rather than fixed outputs.