Unified Data Representation (UDR)

• Unified Data Representation (UDR) is a framework that converts heterogeneous data into a single format while preserving key task-relevant details and modality-specific nuances.
• It employs varied forms—such as probabilistic matrices, shared latent spaces, and heterogeneous graphs—to integrate data from diverse sources.
• UDR enhances data transfer, retrieval, and reasoning by balancing common representation with the preservation of essential structural and semantic differences.

Unified Data Representation (UDR) denotes a family of theories, schemas, and learning frameworks that convert heterogeneous data into a single representation space or structure while preserving task-relevant information. Across the literature, the unified object may be a probabilistic matrix, a shared latent space, a heterogeneous knowledge graph, a discrete code space, a tensor factorization framework, a table-like serialization, or a modular directory of derived artifacts. What remains constant is the attempt to replace fragmented, modality-specific, database-specific, or task-specific encodings with one common substrate that supports transfer, retrieval, prediction, reasoning, compression, or visualization (Kovács et al., 2014, Liu et al., 29 May 2026, Belluomo et al., 2024, Lyu et al., 2024).

1. Historical and conceptual foundations

One of the earliest explicit formulations of UDR is information-theoretic. In network analysis, the representation problem is posed as finding a matrix $B$ that is hardest to distinguish from the original data matrix $A$, with quality measured by relative entropy $D(A\|B)$; the optimal representation is $B^*=\arg\min_B D(A\|B)$ (Kovács et al., 2014). In that formulation, network visualization, one-dimensional ordering, and coarse-graining are not separate tasks but different instances of the same optimization principle.

Later work generalizes this unifying idea from matrices to heterogeneous data ecosystems. In heterogeneous recommendation, different item types are mapped from type-specific feature spaces into a shared latent space while preserving topology, so that one recommender can exploit attributes from all types jointly (Lu et al., 2022). In structured data-to-text generation, tables, knowledge graph triples, and meaning representations are converted into one table-like serialization with shared tags such as <table>, <row>, <cell>, and <col_header> (Li et al., 2023). In routing, UDR explicitly “replaces problem enumeration with data unification,” encoding many VRP variants through one node-level schema rather than a fixed taxonomy of hand-labeled tasks (Zhou et al., 27 Sep 2025).

A recurring conceptual distinction in this literature is that unification does not imply homogenization. Several papers define UDR precisely by what must remain distinct inside the unified form: topology in recommendation (Lu et al., 2022), modality-specific residuals in multimodal discrete learning (Huang et al., 2024), species-related biological structure in cross-species transcriptomics (Nishio et al., 10 Jul 2025), and structural versus semantic information in RTL design (Liu et al., 29 May 2026). UDR is therefore best understood as controlled commonality rather than indiscriminate merging.

2. Canonical representational forms

Modern UDR systems recur around a small number of representational patterns. Shared latent spaces are common when the goal is cross-modal or cross-domain transfer. UniRTL learns a single embedding space in which RTL code and its corresponding control/data flow graph (CDFG) are jointly encoded and aligned (Liu et al., 29 May 2026). UniBind replaces image-centered alignment with a modality-agnostic, language-anchored representation space for seven modalities (Lyu et al., 2024). Mix-Geneformer uses one Transformer backbone and one mixed corpus to embed human and mouse scRNA-seq cells into a shared latent space (Nishio et al., 10 Jul 2025). USRA seeks a latent state representation that is invariant to visual nuisance factors across RL domains (Hearn et al., 2022).

Graph-structured UDRs are used when heterogeneous entities and relations must remain explicit. Precision oncology is represented as a single 4-partite knowledge graph

$$H=(V,E), \qquad V = Pa \cup Mu \cup Di \cup Dr, \qquad E = E_G \cup E_R \cup E_M,$$

which unifies patient mutations, medical records, and external medical knowledge (Belluomo et al., 2024). UniHR converts hyper-relational, temporal, and nested facts into a triple-based hierarchical graph with atomic nodes, relation nodes, and fact nodes (Liu et al., 2024). In BIM, semantic properties become node features, while topological and local spatial relations become typed edges and edge features in one multi-relational network (Han et al., 8 May 2025).

Discrete or sequence-based UDRs appear when heterogeneous data are forced into a common symbolic interface. PixelBytes flattens text, audio, action-state signals, and pixelated images into a single token sequence processed by one sequence model family (Furfaro, 2024). SRCID uses a shared latent codebook across audio, video, and text, but argues that unified discrete representation must preserve cross-modal semantic commonality while separating modality-specific information (Huang et al., 2024). Data-to-text UDR similarly treats heterogeneous structured inputs as instances of one serialized “virtual table” language (Li et al., 2023).

Schema-based UDRs replace latent unification with standardized data organization. URDD is a modular directory of versioned JSON and YAML modules that stores derived robot information—such as DOF mappings, kinematic paths, bounds, convex hulls, convex decompositions, and link-shape abstractions—on top of a base robot specification (Klein-Seetharaman et al., 29 Dec 2025). In this sense, UDR can be a software architecture rather than a learned embedding.

Domain	UDR form	Representative paper
Networks	Probabilistic matrix minimizing $D(A\|B)$	(Kovács et al., 2014)
RTL hardware	Shared Transformer space over code, text, and CDFG	(Liu et al., 29 May 2026)
Precision oncology	Single heterogeneous 4-partite knowledge graph	(Belluomo et al., 2024)
Multimodal binding	Modality-agnostic, LLM-anchored embedding space	(Lyu et al., 2024)
Cross-species scRNA-seq	Shared human–mouse Transformer latent space	(Nishio et al., 10 Jul 2025)
Robotics	Versioned modular directory of derived robot data	(Klein-Seetharaman et al., 29 Dec 2025)

3. Construction mechanisms and optimization principles

UDR methods differ most sharply in how they remove nuisance variation and how they enforce common structure. Some operate by normalization before representation learning. In HRTF modeling, cross-database differences are attributed to measurement setup variations, including loudspeaker frequency response, microphone response, and anechoic room imperfections. The proposed position-dependent normalization divides each subject’s HRTF by the average-person HRTF from the same database at the same source position and ear, thereby removing location-specific system response before learning a shared representation (Wen et al., 2023).

Others rely on explicit cross-modal alignment objectives. UniRTL first aligns functional summaries and RTL code, then incorporates graph information and performs mutual masked modeling across text, code, and graph. During text-code warm-up, 20% of tokens are masked with the 80/10/10 replacement rule; when graphs are integrated, 20% of graph nodes are masked with a learnable $[\text{MASK}]_2$ embedding and the model jointly reconstructs masked text tokens, code tokens, and graph node types (Liu et al., 29 May 2026). This enforces fine-grained rather than instance-level alignment.

Distribution alignment can also be formalized geometrically. DURation uses RKHS mean embeddings with a Gaussian RBF kernel to align transformed distributions from different item types, adds a modified CORAL loss to preserve second-order topology, and optimizes the combined objective

$L = C + \alpha A + \beta T$

for classification, alignment, and topology preservation (Lu et al., 2022). UniBind aligns each modality directly to text-derived semantic centers through contrastive learning rather than routing all modalities through image embeddings (Lyu et al., 2024). Mix-Geneformer combines Masked Language Modeling with a SimCSE-style contrastive loss to capture shared and species-specific gene patterns in one encoder (Nishio et al., 10 Jul 2025).

Several UDR frameworks emphasize disentanglement rather than simple alignment. SRCID argues that better quantization does not automatically yield better multimodal unification, and therefore replaces numerical residuals with semantic residuals. Its objective combines reconstruction, commitment, CPC, CMCM, and mutual-information terms, while minimizing mutual information between modal-general and modal-specific features and maximizing mutual information among general features across modalities (Huang et al., 2024). USRA similarly decomposes observation latents into domain-general and domain-specific factors and adds SVEA-style Q-value consistency under augmentation (Hearn et al., 2022).

A separate line of work treats the transform itself as part of the unified representation. DTR defines the recovered tensor as

$\mathcal{X}=f_\xi \circ g_\theta(\mathcal{Z}),$

where $g_\theta$ is a deep latent generative module and $A$0 is a deep transform module. The central claim is that transform-based tensor recovery depends on both transform and characterization, so both should be learned jointly rather than pairing a transform with shallow SVD-, NMF-, or QR-based latent characterization (Zhou et al., 2024).

4. Representative domain instantiations

In hardware design automation, UDR is used to fuse complementary views of the same artifact. RTL code is treated as semantically and functionally expressive, whereas the CDFG preserves complete structural information and can be faithfully converted back to code. UniRTL therefore uses a unified Transformer, initialized from CodeBERT-base-mlm, with a graph-aware tokenizer and hierarchical training strategy so that code, text, and graph become mutually predictive in one space (Liu et al., 29 May 2026).

In biomedical and clinical contexts, UDR often takes graph form because cross-source inference is central. The oncology knowledge graph integrates patient mutation profiles, diseases, treatments, treatment effectiveness, survival labels, DisGeNET disease–gene associations, and PharmGKB-style mutation–drug relations into one queryable structure. This unified graph reduces medical tasks to breadth-first search, Hamming and Jaccard similarity computations, minimum hitting set, and minimum weight hitting set (Belluomo et al., 2024). UniHR performs a comparable unification for knowledge graphs themselves by converting qualifiers, timestamps, and nested facts into one triple-based hierarchical representation (Liu et al., 2024).

In multimodal machine learning, UDR is often motivated by hub bias. UniBind argues that image-centered binding yields an unbalanced representation space among modalities, and therefore constructs an LLM-generated knowledge base, localizes class-wise embedding centers from top-50 text embeddings, and aligns image, audio, point cloud, thermal, video, and event data directly to those semantic centers (Lyu et al., 2024). PixelBytes takes the more radical position that modalities should be flattened into one sequence language of bytes, palette indices, and control states (Furfaro, 2024). SRCID, by contrast, retains a shared discrete codebook but preserves non-shared content as semantic residuals (Huang et al., 2024).

In scientific data integration, unification may be driven by cross-database or cross-species mismatch. The HRTF work shows that unified learning across databases is undermined by measurement-induced spectral shifts unless they are normalized away first (Wen et al., 2023). Mix-Geneformer integrates approximately 50 million human and mouse cells in one Transformer-based corpus, using rank-value encoding to emphasize high-variance genes and enabling direct comparison of cell states across species (Nishio et al., 10 Jul 2025).

In engineering and infrastructure modeling, UDR becomes a standardized graph or directory. The BIM network representation turns IFC-derived components into nodes, semantic properties into node features, and host, connection, touch-floor, and local spatial relationships into typed edges with geometric descriptors such as angle, shortest distance vector coordinates, shortest distance, and angle with the horizontal plane (Han et al., 8 May 2025). URDD similarly stores derived robot data once, in reusable modules, rather than recomputing the same kinematic and geometric information independently in every simulator or planner (Klein-Seetharaman et al., 29 Dec 2025).

5. Empirical behavior and demonstrated advantages

Across domains, empirical results generally support the claim that unified representations improve transfer, robustness, or data efficiency when compared with unimodal, fragmented, or task-specific alternatives. In HRTF reconstruction, position-dependent normalization improves Log-Spectral Distortion in all five cross-database experiments, with representative changes such as $A$1 dB and $A$2 dB (Wen et al., 2023). In heterogeneous recommendation, DURation reports improvements of 4.1% to 34.9% by AUC score and 3.7% lift by online CTR (Lu et al., 2022). In RL domain adaptation, USRA achieves higher sample efficiency and 14.3% better domain adaptation performance than the best baseline on the Walker benchmark (Hearn et al., 2022).

The same pattern appears in multimodal and sequence modeling. UniBind reports an average zero-shot recognition gain of 6.36%, a 6.75% gain on ImageNet in the multi-modal fine-tuning setting, and a 90% reduction in learnable parameters (Lyu et al., 2024). In structured data-to-text generation, unified serialization yields a 66% relative zero-shot BLEU improvement when transferring from ToTTo to DART (Li et al., 2023). URS reports a single model that can handle more than 100 VRP variants, including more than 90 unseen variants, without any fine-tuning (Zhou et al., 27 Sep 2025).

Domain-specific studies also show that the benefits are not confined to generic transfer benchmarks. UniRTL is best on all metrics for post-synthesis area and delay prediction without netlist information, remains best with netlist information, and achieves the strongest results in natural language code search and functionality equivalence checking; its ablations show that removing graph or code harms performance, and that direct concatenation underperforms explicit mutual masked modeling (Liu et al., 29 May 2026). Mix-Geneformer reaches 95.8% accuracy on mouse kidney data versus 94.9% from the best existing model and identifies perturbation targets such as Umod, HSPB1, and S100B that are described as consistent with prior in vivo findings (Nishio et al., 10 Jul 2025). These results support a broader interpretation of UDR as an intervention on representation geometry rather than merely on input formatting.

6. Misconceptions, limitations, and open problems

A common misconception is that UDR is equivalent to concatenating heterogeneous inputs. Multiple papers reject this directly. UniRTL includes a direct-combine baseline that concatenates text-code and graph features without explicit alignment; it improves over unimodal baselines but still underperforms the aligned model, indicating that multimodality alone is not sufficient (Liu et al., 29 May 2026). SRCID reports that RVQ and FSQ improve unimodal reconstruction-like metrics yet do not improve cross-modal tasks, which argues against identifying better unification with more precise quantization (Huang et al., 2024).

Another misconception is that a unified representation should erase all domain-specific structure. The HRTF study shows the opposite: cross-database unification fails unless database-specific nuisance variation is removed in a physically motivated, position-dependent way (Wen et al., 2023). Mix-Geneformer explicitly aims to capture both shared and species-specific gene patterns, and its authors note variability in zero-shot transfer as well as substantial computational cost, including 8 A100 GPUs and roughly 3 days for fine-tuning and classification experiments (Nishio et al., 10 Jul 2025). In multimodal discrete learning, semantic residuals are treated as necessary because modalities often contain only partial overlap (Huang et al., 2024).

The literature also exposes methodological limits. Some UDR proposals are strongly empirical but domain-bound, while others are conceptually rich but only qualitatively validated. The BIM representation paper, for example, provides no formal benchmark dataset, no numerical metrics, and no ablation study (Han et al., 8 May 2025). URDD shows the opposite tradeoff: richer unified representation yields larger disk footprint, since URDDs with meshes are much larger than corresponding URDFs with meshes because they package precomputed derived data (Klein-Seetharaman et al., 29 Dec 2025). These cases make clear that unification can shift complexity rather than eliminate it.

A plausible implication is that future UDR systems will continue to combine three elements that already recur across the literature: explicit nuisance modeling before unification, architectures that preserve typed structure inside the unified form, and task objectives that separate shared from non-shared information. The published record suggests that UDR is most successful when it is treated not as a universal encoding trick, but as a principled compromise between common representation and preserved heterogeneity.