acceptodds
Under review as a conference paper at ICLR 2027

CONTEXT-AWARE SYNCHRONIZATION FOR SPATIAL MULTI-OMICS REPRESENTATION LEARNING

Abstract

Spatial multi-omics domain identification relies on unified representations that preserve multimodal molecular information while organizing domain-resolvable latent geometry. Existing geometry organization mechanisms mainly rely on explicit inter-location relations or shared semantic references, leaving location- specific multimodal and multiscale context without a direct role in geometry opti- mization. To address this gap, Context-aware Synchronization for Spatial Multi- omics Representation Learning (CSR) formulates latent geometry optimization as a context-regulated synchronization process on the hypersphere. CSR constructs location-specific contextual references from multimodal representations and mul- tiscale graph contexts, and combines them with relational graph coupling to reg- ulate representation evolution. The continuous synchronization dynamics mono- tonically decrease the associated objective, while contextual references impose location-dependent stationary conditions. Experiments on spatial multi-omics benchmarks demonstrate consistent improvements in spatial domain identification over representative baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Spatial Multi-Omics, Representation Learning, Graph Synchronization, Context-Aware Learning, Spatial Domain Identification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nakkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context-aware Synchronization for Spatial Multi-omics Representation Learning (CSR), which formulates spatial multi-omics representation learning as a context-regulated synchronization process on a hyperspherical representation space. CSR constructs location-specific contextual references by integrating multimodal representations with multiscale graph contexts via a graph scattering transform, and uses these references to guide hyperspherical synchronization dynamics under graph-based relational constraints. The authors establish energy descent of the continuous synchronization dynamics and characterize context-regulated stationary states.

The main contributions are: (1) formulating spatial multi-omics representation learning as a context-regulated synchronization problem on a hyperspherical representation space; (2) introducing location-specific contextual references that integrate multimodal representations with multiscale graph contexts to provide continuous location-dependent guidance; (3) providing theoretical analysis establishing energy descent and characterizing context-regulated stationary states; and (4) achieving strong clustering performance on five spatial multi-omics datasets.

Validation includes experiments on five datasets across RNA+ATAC, RNA+ADT, and RNA+ADT+ATAC settings, comparisons with eight representative baselines, nine evaluation metrics, ablation studies on key components, sensitivity analysis on context strength and synchronization steps, and runtime/memory comparisons. CSR generally outperforms representative baselines such as SMART, SpatialGlue, COSMOS, and ProtoPSR on clustering metrics.

Critical strengths
  1. The paper formulates spatial multi-omics representation learning as context-regulated synchronization on a hyperspherical space. Section 4 provides a unified formalization of relation-guided regularization (Eq. 3) and a gradient constraint (Proposition 1), which clearly identifies the gap that existing relation-guided methods do not treat location-specific context as an independent component of geometry optimization.

  2. CSR generally leads across five datasets, eight baselines, and nine metrics, with particularly clear margins on Mouse Embryonic Brain E15.5 (ARI 70.94 vs. 61.39 for SMART) and Human Lymph Node A1 (ARI 38.13 vs. 35.23 for SMART). The ablation study (Fig. 5) covers five key components, and the dynamics analysis (Fig. 4) intuitively shows how contextual regulation promotes intra-domain synchronization while preventing excessive cross-domain synchronization.

Critical weaknesses
  1. Appendix C.3 states the contextual strength α = [0.01], while the sensitivity analysis in Fig. 6 uses γ as the context strength and shows that γ = 0.1–0.5 is favorable. If α and γ are the same parameter, the default setting contradicts the sensitivity conclusion. If they are different, their relationship is not clarified.

  2. The proofs of Theorem 1 and Proposition 2 are correct but are routine properties of Riemannian gradient flows. The paper does not provide convergence rates, finite-time guarantees, or a deeper characterization of the interaction between the graph spectrum and contextual references. Moreover, the proposed dynamics have no formal connection to classical synchronization theory such as the Kuramoto model.

Questions for authors

1.Relationship between α and γ. Is α = [0.01] in Appendix C.3 the same parameter as γ in Fig. 6, where the favorable range is 0.1–0.5? If so, why does the default value deviate from the favorable range in the sensitivity analysis? If not, please clarify the definitions and the basis for each value.

2.Please provide clustering performance and modality-specific information preservation after removing the reconstruction objective (Eq. 15). The absence of this ablation prevents me from assessing the actual contribution of the claim that the reconstruction objective preserves modality-specific information.

3.Please provide a sensitivity analysis for the number of scales (currently two) and the maximum scattering order (currently two).

4.Proposition 2 characterizes the stationary condition for a fully synchronized state, but representations of different spatial domains in real data are not fully synchronized. Please discuss the implications of this theoretical result for practical synchronization trajectories that are not fully synchronized, or provide a more general characterization of stationary states.

5.On HLN A1 and Mouse Brain, CSR’s sLISI is lower than some baselines. Please analyze the source of this trade-off (e.g., whether it is related to contextual strength or graph construction) and discuss in which application scenarios spatial coherence should be prioritized.

2 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1/Q1, W2, Q2, Q3

W1/Q1. Inconsistency in contextual-strength notation and settings.

R1. We thank the reviewer for identifying this inconsistency. The symbols α\alpha and γ\gamma refer to the same contextual-strength parameter, which has been standardized to α\alpha throughout the revised manuscript.

Regarding the parameter settings, the contextual strength is dataset-specific, rather than fixed across all datasets. The settings used in our experiments are summarized in Appendix D(Training and Evaluation Protocol). The sensitivity analysis evaluates α∈{0.01,0.05,0.1,0.5,1,2}\alpha\in\{0.01,0.05,0.1,0.5,1,2\}, covering all the settings above. The sensitivity analysis shows dataset-dependent behavior rather than a universally favorable range (Fig. 9).

W2. Limited theoretical depth and connection to synchronization theory.

R2. We have strengthened the theoretical analysis in three aspects: finite-time stationarity, graph–context interactions, and the connection to classical synchronization theory. These results extend the original energy-descent and stationary-state analysis with quantitative characterizations of the proposed dynamics.

1. Finite-time stationarity guarantee. We have established that the continuous synchronization dynamics satisfy

min⁡0≤t≤τ∥grad⁡E(Z(t))∥F2≤E(Z(0))−Einfτ\min_{0 \leq t \leq \tau} \lVert \operatorname{grad}\mathcal{E}(Z(t)) \rVert_F^2 \leq \frac{\mathcal{E}(Z(0))-\mathcal{E}_{\mathrm{inf}}}{\tau}

This provides an O(1/τ)O(1/\tau) finite-time bound on the minimum squared Riemannian gradient norm, under fixed graph and contextual references (Appendix F.1).

2. Interaction between graph spectrum and contextual guidance. We have derived a local spectral characterization around a synchronized configuration. For ci=ru+δic_i=ru+\delta_i, where r>0r>0 and δi\delta_i is a sufficiently small tangent perturbation, the stationary response satisfies

\frac{\alpha}{\beta\lambda_k+\alpha r}\hat\delta_k +O(\|\Delta\|_F^2).$$ **This quantifies how graph Laplacian eigenvalues and contextual strength jointly influence local representation differences**, with higher-frequency contextual variations more strongly attenuated by graph coupling (Appendix F.2). **3. Connection to classical synchronization theory.** We have established an exact reduction to Kuramoto-type dynamics in the two-dimensional case: $$\dot\theta_i= \beta\sum_j A_{ij}\sin(\theta_j-\theta_i) +\alpha r_i\sin(\phi_i-\theta_i).$$ **The first term represents Kuramoto-type phase coupling, while the second introduces location-specific contextual forcing.** This establishes a formal connection to classical synchronization theory (Appendix F.3). The detailed assumptions, theoretical results, and proofs are provided in **Appendix F.1–F.3**. The practical interpretation of Proposition 2 is addressed separately in our response to Q4. > Q2. Contribution of the reconstruction objective to modality-specific information preservation. R3. Based on your suggestion, we have conducted modality-specific reconstruction ablations to evaluate the contribution of reconstruction supervision to both clustering performance and modality-feature recoverability. Each reconstruction term is removed separately while retaining all input modalities and the remaining reconstruction objectives. 1. **Clustering performance**. Removing each modality-specific reconstruction term reduces ARI, NMI, and FMI relative to Full CSR across the tested datasets, demonstrating the contribution of reconstruction supervision to spatial domain identification (Fig. 10). 2. **Modality-specific information preservation.** We further evaluated MSE and R² between each modality’s original features and the corresponding reconstructions from the learned representations, using independently trained evaluation decoders under the same protocol. Reconstruction supervision generally improves the recoverability of modality-specific information from the learned representations, which supports the role of the reconstruction objective in preserving modality-specific information(Fig. 11). Together, these results demonstrate that reconstruction supervision improves clustering performance and generally enhances modality-feature recoverability. The experimental details are provided in Appendix J. > Q3. Sensitivity to graph-scattering hyperparameters. R4. According to your advice, we have conducted sensitivity experiments on the maximum wavelet scale $H\in{1,2,3,4,5}$ and maximum scattering order $\nu\in{1,2,3,4}$. The experiments cover Human Lymph Node A1, Human Tonsil Section 1, and Mouse Brain, reporting ARI and NMI. **As shown in Fig. 9, CSR exhibits relatively stable performance across the tested configurations**, although the sensitivity varies by dataset, particularly on Human Tonsil Section 1.
AuthorsOpenReview ↗

Rebuttal Part 2: Q4, Q5

Q4. Implications of Proposition 2 for partially synchronized states.

R5. Proposition 2 identifies when complete synchronization cannot be a stationary state; it does not assume that representations in real spatial data become fully synchronized.

For partially synchronized states, Eq. 10 requires relational coupling and contextual guidance to satisfy the projected force-balance condition:

Πzi(β∑jAijzj+αci)=0,∀i.\Pi_{z_i}\left(\beta\sum_j A_{ij}z_j+\alpha c_i\right)=0,\quad\forall i.

More specifically, Appendix F.2 establishes that, under small nonuniform contextual perturbations, nearby stationary states retain location-dependent representation differences. Thus, contextual guidance can sustain distinct representations while relational coupling promotes coherence.

Consistently, Fig. 4 shows increasing intra-domain similarity while reducing inter-domain similarity during synchronization.

Q5. Spatial coherence trade-off and application-specific priorities.

R6. Following your suggestion, we have conducted additional experiments to investigate whether contextual strength or graph construction contributes to CSR's lower sLISI on HLN A1 and Mouse Brain. We also clarify when spatial coherence should be prioritized.

1. Source of the discrepancy. In the main results, CSR achieves lower sLISI than COSMOS on HLN A1 (88.99 vs. 93.06) and Mouse Brain (90.12 vs. 96.04), despite substantially higher ARI. We further varied α\alpha and compared union-graph with spatial-only coupling across five datasets. The results show that coupling-graph construction has a more pronounced effect, particularly on Mouse Brain, while contextual-strength effects vary across datasets. At α = 0.01, spatial-only coupling increases mean sLISI from 88.39 to 96.30 but reduces mean ARI from 36.83 to 29.28 on Mouse Brain (Figs. 7–8 and Tab. 10, Appendix H).

2. Application-specific priorities. Spatial coherence should be prioritized when identifying large, contiguous tissue compartments or reducing spatially isolated assignments. In contrast, preserving molecular distinctions is important for heterogeneous tissues and small cellular niches, where excessive spatial smoothing may obscure meaningful domain boundaries.

These results indicate that higher spatial coherence does not necessarily imply more accurate spatial domain identification. Additional visual comparisons are provided in Fig. 6 (Appendix H).

Reviewer_CtGNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context-aware Synchronization for Spatial Multi-omics Representation Learning (CSR), a framework that organizes multimodal representations on a hyperspherical latent space. CSR builds location-specific contextual references by fusing multimodal molecular data with multiscale graph scattering features, which then guide iterative representation updates alongside graph relational coupling.

Critical strengths
  • Context-aware design
  • Strong empirical performance
Critical weaknesses

W1. Mathematical inconsistency in energy descent proof: The gradient of the graph smoothness objective (Eqs. 1, 5) lacks a factor of two in Appendix C (Eq. 24). Consequently, the actual dynamics (Eq. 10) do not match the negative Riemannian gradient, invalidating the proof for monotonic energy descent (Theorem 1) when relational and contextual forces compete.

W2. Limited theoretical and methodological novelty: Combining graph Laplacian smoothness with location-specific feature anchoring is well-established; the theoretical proofs rely on standard Riemannian gradient flow and basic definitions rather than novel optimization principles.

W3. Iterative synchronization unproven against simpler baselines: Ablations do not isolate whether gains stem from the iterative hyperspherical dynamics or simply from incorporating multiscale graph scattering features (e.g., via standard feature fusion or residual graph propagation).

Questions for authors

Provided above

4 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1, W2

W1. Mathematical inconsistency in the energy-descent proof.

R1. We thank the reviewer for identifying this inconsistency. We have corrected the coefficient of the graph energy term in Eq. 5 from β\beta to β/2\beta/2 ensuring consistency with the gradient derivation in Appendix C.1 and the original synchronization dynamics in Eq. 10.

The corrected dynamics follow the negative Riemannian gradient of the unified objective, satisfying

=-\sum_i|\operatorname{grad}_{z_i}\mathcal E|_2^2\leq0.$$ **Thus, the continuous-time energy-descent guarantee remains valid even when relational and contextual forces oppose each other.** The correction does not affect the implemented algorithm or the reported results. > W2. Limited theoretical and methodological novelty. R2. We respectfully disagree with the characterization of CSR as merely combining graph Laplacian smoothness with location-specific feature anchoring. This interpretation overlooks **how location-specific multimodal and multiscale context is incorporated into and regulates representation evolution throughout synchronization**. The contribution of CSR is not to introduce new regularization terms or a new general-purpose optimization principle, but to formulate and analyze a context-regulated synchronization framework for spatial multi-omics representation learning. **1. Methodological contribution: Context-regulated synchronization.** Our analysis in Section 4 provides a common relation-guided perspective on representative geometry-shaping objectives (Eq. 3 and Appendix A), motivating the explicit incorporation of location-specific contextual information into representation optimization. As formulated in the corrected Eq. 5, CSR considers a unified energy objective: $$\mathcal{E}(Z;A,C)=\frac{\beta}{2}\operatorname{Tr}(Z^\top LZ)-\alpha\sum_i c_i^\top z_i.$$ Here, the first term captures relational coupling, while the second introduces contextual guidance. Specifically, $A$ characterizes spatial and molecular relationships, and $c_i$ denotes the location-specific contextual reference constructed from multimodal molecular representations and multiscale graph scattering features (Eqs. 8–9). **The methodological distinction lies in how these contextual references are constructed and incorporated into the synchronization process**, rather than in the individual regularization terms. Under the hyperspherical constraint, CSR derives the following synchronization dynamics (Eq. 10): $\dot{z}_i=\Pi_{z_i}\left(\beta\sum_j A_{ij}z_j+\alpha c_i\right),$ where $\Pi_{z_i}=I-z_i z_i^\top$ projects the combined forces onto the tangent space. These dynamics involve two interacting components: - **Relational coupling**, $\beta\sum_j A_{ij}z_j$, encourages consistency among related locations. - **Contextual regulation**, $\alpha c_i$, provides a location-specific direction derived from multimodal and multiscale contextual information. The contextual reference remains fixed during each inner synchronization trajectory, while the representations continue to evolve. Thus, **contextual information is not used only for feature fusion or initialization; it directly participates in every synchronization update**, jointly with relational coupling. CSR therefore provides a context-regulated hyperspherical synchronization framework for organizing spatial multi-omics representations. Its effectiveness relative to simpler fusion and propagation mechanisms is further supported by the controlled comparisons discussed in our response to W3 (Fig. 5). **2. Theoretical contributions: Energy descent and context-dependent synchronization behavior.** Our theoretical analysis does not claim a new general-purpose Riemannian optimization principle. Rather, it establishes the optimization consistency of CSR and characterizes how location-specific contextual guidance affects synchronization behavior. **(a) Energy-descent guarantee.** Theorem 1 establishes that the continuous synchronization dynamics follow the negative Riemannian gradient of the corrected unified objective: $\dot{z}_i=-\operatorname{grad}_{z_i}\mathcal{E}(Z;A,C).$ Consequently, $$\frac{d\mathcal{E}}{dt}=-\sum_i\|\operatorname{grad}_{z_i}\mathcal{E}\|_2^2\leq0.$$ **This establishes that relational coupling and contextual regulation jointly follow a consistent energy objective**, rather than forming an arbitrary combination of update signals. While the gradient-flow principle is established, Theorem 1 verifies its application to the proposed relational-contextual dynamics.
AuthorsOpenReview ↗

Rebuttal Part 2: W2, W3

(b) Contextual heterogeneity and synchronization equilibria.

More importantly, Proposition 2 characterizes how location-specific contextual regulation changes the stationary conditions of graph synchronization. At a fully synchronized state z1=⋯=zN=zz_1=\cdots=z_N=z, the relational coupling term vanishes after tangent-space projection, whereas the contextual term remains:

z˙i=αΠz(ci).\dot{z}_i=\alpha\Pi_z(c_i).

Therefore, for α>0\alpha>0, a fully synchronized state is stationary if and only if

Πz(ci)=0,∀i,\Pi_z(c_i)=0,\quad\forall i,

or equivalently, ci∈span⁡(z)c_i\in\operatorname{span}(z) for all ii.

This implies that heterogeneous contextual references can exclude fully synchronized stationary states. Specifically, define the contextual heterogeneity measure

HC=min⁡zTz=1∑i(Πz(ci))TΠz(ci)H_C = \min_{z^T z=1} \sum_i (\Pi_z(c_i))^T \Pi_z(c_i)

When HC>0\mathcal{H}_C>0 and α>0\alpha>0, no fully synchronized state can satisfy the stationary condition. This explains how location-specific contextual guidance changes the equilibrium conditions relative to graph synchronization without contextual regulation.

(c) Local interaction between graph spectrum and contextual guidance.

We have further extended the analysis to characterize the interaction between relational coupling and contextual guidance. Under the local perturbation assumptions in Appendix F.2, the stationary response of graph spectral mode kk satisfies

\frac{\alpha}{\beta\lambda_k+\alpha r}\hat{\delta}_k +O(\|\Delta\|_F^2),$$ where $\lambda_k$ is the corresponding graph Laplacian eigenvalue and $r>0$ denotes the common reference strength in the local analysis. **This result quantifies how graph spectral modes and contextual guidance jointly determine the response of representations to contextual perturbations**, extending the analysis beyond the standard energy-descent property. Taken together, **Theorem 1 establishes the optimization consistency of CSR, while the stationary-state and local spectral analyses characterize how contextual guidance influences synchronization behavior.** We have corrected Eq. 5 and the corresponding derivation, preserving the original synchronization dynamics and the continuous-time energy-descent guarantee. We have further clarified the effect of contextual heterogeneity on fully synchronized stationary states in Section 6.2 and Appendix C.2, and added a local spectral characterization of relational–contextual interactions in Appendix F.2. > W3. Iterative synchronization versus simpler baselines. R3. **To distinguish the contribution of iterative hyperspherical synchronization from multiscale GST features and conventional graph propagation, we have compared CSR with three simpler alternatives**. The comparisons therefore examine the effect of different update mechanisms under the same learned contextual information. Specifically, *w/o Iterative Updates(GST + Static Fusion)* uses the contextual initialization without iterative updates; *GST + Residual Propagation* replaces synchronization with conventional residual graph propagation; and *GST + Context-guided Residual Propagation* additionally incorporates the contextual reference at each propagation step. All iterative variants use the same number of update steps. Tab. 1. Comparison with simpler alternatives on four datasets (ARI, \%). | Method | HLN | MB | HT1 | MEB15 | | ------------------------------------------ | --------: | --------: | --------: | --------: | | w/o Iterative Updates(GST + Static Fusion) | 27.77 | 31.52 | 13.17 | 54.07 | | GST + Residual Propagation | 18.91 | 40.36 | 29.19 | 66.96 | | GST + Context-guided Residual Propagation | 29.43 | 41.48 | 13.08 | 63.36 | | Full CSR | **38.13** | **43.50** | **41.86** | **70.94** | **Full CSR achieves the highest ARI on all four datasets**, outperforming the strongest alternative by **8.70, 2.02, 12.67, and 3.98 percentage points**, respectively. These results show that **GST-based contextual initialization, residual graph propagation, and repeated contextual injection do not match CSR under the controlled evaluation settings**, supporting the contribution of context-regulated hyperspherical synchronization. We have incorporated these comparisons into Fig. 5, with configurations provided in Tab. 6 and Appendix E.

Thank you for the detailed rebuttal and additional experiments.

The correction to the energy formulation appears to address my primary concern in W1, provided that the revised objective and gradient derivation are consistent.

Regarding W2, the clarification of the intended methodological contribution and the additional spectral analysis are helpful. However, I still see the main contribution as an application and integration of established optimization principles, rather than a fundamentally new optimization framework.

Thank you for the follow-up assessment. We appreciate that the correction to the energy formulation has addressed the primary concern in W1, and that the additional experiments and spectral analysis have helped clarify the behavior of CSR.

Regarding W2, we would like to further clarify one important distinction. We do not claim to introduce a fundamentally new general-purpose optimization principle. However, we respectfully believe that characterizing CSR as merely an application or integration of established optimization principles does not fully capture its methodological and theoretical contributions.

The novelty of CSR lies not in the Riemannian gradient-flow principle itself, but in the specific context-regulated synchronization mechanism formulated through it. In particular, CSR introduces location-specific contextual forces, derived from multimodal molecular representations and multiscale graph scattering features, directly into relationally coupled hyperspherical dynamics:

z˙i=Πzi(β∑jAijzj+αci).\dot z_i=\Pi_{z_i}\left(\beta\sum_j A_{ij}z_j+\alpha c_i\right).

As a result, the evolution of each spatial representation is jointly governed by spatial relational coupling and its own contextual reference. This is fundamentally different from simply applying graph Laplacian smoothing or adding feature-level contextual anchoring. The contextual information continuously regulates the representation evolution throughout the synchronization process, rather than merely serving as an auxiliary feature or regularization term.

More importantly, this formulation leads to a specific theoretical property that does not follow merely from invoking a standard Riemannian gradient flow. As characterized in Proposition 2, at a fully synchronized state, the relational coupling vanishes after tangent-space projection, whereas the location-specific contextual forces generally remain active. Consequently, a fully synchronized stationary state requires all contextual references to be aligned with a common direction (up to sign). In particular, if the contextual heterogeneity satisfies

HC=min⁡∥z∥2=1∑i∥(I−zz⊤)ci∥22>0,\mathcal H_C=\min_{\|z\|_2=1}\sum_i\|(I-zz^\top)c_i\|_2^2>0,

no fully synchronized stationary state can exist. This result characterizes how location-specific contextual heterogeneity fundamentally changes the stationary-state structure of graph synchronization. We believe this is an important theoretical consequence of the proposed formulation, rather than a routine application of an existing optimization principle.

We fully acknowledge that the individual mathematical ingredients are established. Nevertheless, the contribution of a methodology should also be assessed by its specific modeling mechanism, the way established principles are instantiated to realize that mechanism, and the properties induced by the resulting formulation. From this perspective, CSR contributes a context-regulated synchronization framework with a distinct representation evolution mechanism and a corresponding stationary-state characterization, rather than simply applying an existing optimizer to a new application domain.

We hope these clarifications, together with the corrected theoretical derivation and the additional experimental and spectral evidence, address the remaining concerns regarding the methodological and theoretical contributions. Given that the primary mathematical concern has been resolved and the intended contributions have been further clarified and supported, we respectfully ask the reviewer to reconsider the overall assessment and rating of our paper.

Reviewer_D1fVOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Context-aware Synchronization for Spatial Multi-omics Representation Learning (CSR) for spatial domain identification. CSR introduces location-specific multimodal and multiscale context into hyperspherical representation dynamics together with graph-based relational coupling. The paper provides theoretical analysis of the synchronization dynamics and evaluates the method on four real and one simulated spatial multi-omics datasets, with comparisons to recent baselines and several ablation and sensitivity studies.

Critical strengths

1.The motivation is clear and well matched to the problem. CSR makes location-specific multimodal and multiscale context an explicit part of representation optimization while retaining graph-based relational coupling.

2.The empirical evaluation is reasonably comprehensive. CSR performs strongly on most reference-based clustering metrics and is supported by component ablations, sensitivity analysis, qualitative comparisons, and efficiency measurements.

Critical weaknesses

1.The energy-descent analysis contains a coefficient inconsistency. With the definition of �0��0� in Eq. (1), the gradient used in Appendix C.1 differs by a factor of two from Eq. (5) and the dynamics in Eq. (10). This appears locally fixable, but the objective and theoretical derivation should be made consistent.

2.The contribution of iterative synchronization is not fully isolated from that of the contextual representation. The contextual reference �1��1� already enters the initialization �2��2�, and the exact settings of “w/o Synchronization Module” and “w/o Iterative Updates” are not specified. The novelty therefore appears to lie mainly in constructing and explicitly incorporating the location-specific context, while the additional benefit of iterative synchronization is less clear.

Questions for authors

1.Is Eq. (5) missing a factor of �0��0� in front of �1��1�? The gradient in Appendix C.1 appears to correspond to �2��2�, which is also consistent with the �3��3� coefficient in Eq. (29). Please clarify the intended objective and the formulation used in the implementation.

2.What exactly is changed in the “w/o Synchronization Module” and “w/o Iterative Updates” variants in Fig. 5? In particular, do either of them retain the same contextual reference �4��4� without iterative synchronization?

3.Please clarify the downstream clustering protocol, including the clustering algorithm, how the number of clusters is chosen, and whether the reported results are based on a single run or multiple runs.

4 comments
AuthorsOpenReview ↗

Rebuttal Part 1: W1/Q1, W2/Q2, Q3

W1/Q1. Inconsistency between the energy objective and synchronization dynamics.

R1. Thank you for identifying this inconsistency. We have corrected the coefficient of the graph energy term in Eq. 5 from β\beta to β/2\beta/2.

The correction ensures consistency with the gradient derivation in Appendix C.1 and the synchronization dynamics in Eq. 10. The implementation follows Eqs. 10–11. This correction leaves the implemented dynamics and reported results unchanged, while preserving the continuous-time energy-descent conclusion for the corrected objective.

W2/Q2. Insufficient isolation of iterative synchronization and unclear ablation settings.

R2. The improvement of CSR cannot be explained solely by constructing the contextual reference CC and incorporating it into the initialization. We have strengthened the ablation experiments in Fig. 5 and clarified this distinction in Appendix E.

  • Contribution beyond contextual initialization. The w/o Iterative Updates variant retains the GST-based contextual reference construction and the same initialization scheme, Z(0)=RowNorm⁡(H+C+ϵ)Z^{(0)}=\operatorname{RowNorm}(H+C+\epsilon), but directly uses Z(0)Z^{(0)} as the final representation (T=0T=0). Thus, the comparison with Full CSR evaluates the additional contribution of subsequent synchronization. As shown in Fig. 5, Full CSR achieves higher ARI and FMI across all four datasets, demonstrating that contextual initialization alone does not account for the performance gains. The corresponding setting is specified in Eq. 43 (Appendix E).

  • Contribution beyond conventional propagation. To examine whether these gains arise merely from iterative propagation, we have added two variants, GST + Residual Propagation and GST + Context-guided Residual Propagation, which replace hyperspherical synchronization with residual graph propagation, without or with contextual injection at each step. Both retain the GST-based contextual construction and initialization. Full CSR outperforms both variants in ARI and FMI across all four datasets (Fig. 5), providing further evidence for the contribution of the proposed synchronization dynamics.

  • Clarification of ablation settings. The two original variants serve different purposes. w/o Synchronization Module removes the entire synchronization module and performs clustering directly on HH, evaluating the contribution of the complete module. In contrast, w/o Iterative Updates retains contextual construction and initialization but removes subsequent synchronization, evaluating its additional contribution. These configurations are clarified in Tab. 6 and Appendix E.


Q3. Downstream clustering protocol and repeated runs.

R3. We thank you for carefully examining the evaluation protocol. We have clarified the clustering algorithm, cluster-number selection, random seed settings, and number of runs in the revised manuscript (please see Appendix D, Training and Evaluation Protocol).

Thank you for the revisions. The clarifications and additional ablations address my main concerns. Besides, I have addition problem: 1.If the contextual reference (C) remains fixed during synchronization, how does it continue to influence the iterative synchronization beyond its role in the initialization (Z^{(0)})? 2.Since both the contextual reference (C) and graph-based relational coupling are constructed from multimodal information, what distinguishes their roles in representation optimization? 3.When graph-based relational coupling and location-specific contextual guidance favor different directions, does the energy-descent guarantee imply that both components improve simultaneously?

Thank you for these insightful questions. They touch upon three important aspects of CSR: the iterative role of the contextual reference, the complementary functions of relational coupling and contextual guidance, and the interpretation of the energy-descent guarantee.

1. How does the fixed contextual reference continue to influence synchronization beyond initialization?

The contextual reference indeed contributes to the initialization:

Z(0)=RowNorm⁡(H+C+ϵ).Z^{(0)}=\operatorname{RowNorm}(H+C+\epsilon).

However, its role is not limited to initialization. More importantly, CC explicitly appears in every synchronization update (Eq. 11):

zi(t+1)=Norm⁡[zi(t)+ηΠzi(t)(β∑jAijzj(t)+αci)].z_i^{(t+1)}=\operatorname{Norm}\left[z_i^{(t)}+\eta\Pi_{z_i^{(t)}}\left(\beta\sum_j A_{ij}z_j^{(t)}+\alpha c_i\right)\right].

Thus, at every iteration, the same location-specific contextual reference contributes to the tangent-space update direction through the contextual force αci\alpha c_i. Although CC remains unchanged during a synchronization trajectory, its influence is continuously incorporated as the representation state zi(t)z_i^{(t)} evolves. In other words, fixed does not mean inactive: it means that the contextual reference provides a stable guidance signal throughout the iterative optimization, rather than being recomputed from the evolving representations at every step. This distinction also separates its persistent role in synchronization from its additional role in initialization.

2. What distinguishes the roles of AA and CC, given that both leverage multimodal information?

We agree that both AA and CC leverage multimodal information, and that their information sources are not entirely independent. In particular, the construction of CC incorporates multiscale graph scattering responses derived from the relational graph. Nevertheless, they encode different types of information and play distinct roles in representation optimization.

  • Graph-based relational coupling AA encodes edge-level relationships among spatial locations, integrating spatial proximity and modality-specific molecular similarity. It determines which locations interact and how their representations are coupled through relational information.
  • Contextual reference CC encodes node-level, location-specific contextual information by integrating multimodal representations with multiscale graph contexts. It provides an explicit reference direction toward which each location's representation is guided during synchronization.

Therefore, their distinction is not based on statistical independence, but on their functional roles: AA specifies how locations interact, whereas CC specifies the contextual direction that guides each location's representation.

Proposition 2 further illustrates this distinction. At a fully synchronized state, the graph relational coupling vanishes after tangent-space projection, whereas the location-specific contextual force may remain nonzero. Consequently, the two components are not interchangeable: relational coupling promotes consistency across locations, while contextual guidance introduces location-dependent constraints on the synchronization states.

3. Does the energy-descent guarantee imply that both components improve simultaneously when their preferred directions conflict?

No. The energy-descent guarantee applies to the joint objective, rather than requiring each individual component to improve monotonically.

To make this explicit, let EG(Z;A)E_G(Z;A) denote the graph smoothness term and SC(Z;C)=∑ici⊤ziS_C(Z;C)=\sum_i c_i^\top z_i denote the contextual alignment term. The joint objective can be written as

E(Z;A,C)=λAEG(Z;A)−αSC(Z;C),E(Z;A,C)=\lambda_A E_G(Z;A)-\alpha S_C(Z;C),

where λA\lambda_A denotes the graph-term coefficient in the corrected energy formulation. Theorem 1 establishes that along the continuous synchronization dynamics,

dEdt=λAdEGdt−αdSCdt≤0.\frac{dE}{dt}=\lambda_A\frac{dE_G}{dt}-\alpha\frac{dS_C}{dt}\leq 0.

This inequality does not require dEGdt≤0\frac{dE_G}{dt}\leq 0 and dSCdt≥0\frac{dS_C}{dt}\geq 0 to hold simultaneously. When relational coupling and contextual guidance favor different directions, one component may temporarily become less favorable while the improvement in the other component compensates for it, so that the weighted joint objective still decreases.

In this sense, CSR does not optimize the two components independently or enforce their simultaneous monotonic improvement. Instead, the synchronization dynamics follow their combined Riemannian gradient, balancing relational consistency and location-specific contextual guidance under a unified objective. We also clarify that Theorem 1 establishes descent for the continuous-time dynamics; as stated in the paper, the discrete projected Euler update is a numerical approximation whose monotonic decrease is not guaranteed for arbitrary step sizes.

We hope these clarifications resolve the questions and make the respective roles of the two components clearer.

Thank you for the detailed responses. The rebuttal has adequately addressed my remaining questions. The revised manuscript is clearer and more technically sound. I find the context-aware synchronization idea interesting and well motivated. These revisions strengthen the paper's methodological contribution.

I am satisfied with the revisions and will consider increasing my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.