acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Multimodal Graph Learning with Content-Conditioned Edge Reweighting and Hierarchical Alignment

Abstract

Multimodal graph learning combines relational structure with textual and visual attributes, but the utility of a neighbor depends on the target content. We introduce DMGraph, which learns content-conditioned weights over induced multimodal subgraphs and aligns text and vision at node, subgraph, and batch-context levels. Its dynamic edge scorer learns rank-controlled bilinear relevance functions, while hardness-aware contrastive learning handles semantically similar negatives. Across WikiWeb2M, FB15K237, Ele-Fashion, and Goodreads-LP, DMGraph consistently improves over graph-multimodal baselines. In expanded neighborhoods, intermediate learned sparsity consistently outperforms dense reweighting and hard Top-K. Under 50% edge corruption, DMGraph limits relative performance degradation to about 15% on both evaluated datasets, demonstrating effective selection of task-relevant graph context.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Learning, Multimodal Graph

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6mKAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DMGraph proposes a framework for multimodal graph learning that integrates textual, visual, and structural information to improve graph-based generation and classification tasks. Its main idea is to dynamically adjust the importance of neighboring nodes based on their content, allowing the model to prioritize relevant connections and reduce the influence of noisy or unrelated neighbors. It also introduces a hierarchical contrastive alignment strategy that aligns representations at the node, subgraph, and batch-global levels, using debiased negative sampling to mitigate false negatives during training.

Critical strengths
  • DMGraph outperforms the reported variants across several datasets and tasks

-The appendices further include comparisons with alternative methods, ablation studies, hyperparameter sensitivity, expanded-neighborhood experiments, and robustness to injected edges.

Critical weaknesses
  • Appendix B provides several theoretical results, including bounds on approximation error and Lipschitz continuity. However, it is not entirely clear what these results add beyond established properties of bilinear attention and contrastive learning. Could the authors better explain what is theoretically new here and how these results improve our understanding of DMGraph?

  • Although the larger-neighborhood experiments demonstrate actual pruning, it remains unclear how much of the improvement comes from sparse edge selection versus continuous reweighting. Can the authors disentangle these two effects more clearly?

Questions for authors
  • Equation 17 uses only H^t,out as the key/value input to cross-attention.Is this intended, or should other representations also be included?

  • Some variables used in the equations are not clearly defined, making parts of the method difficult to follow.

  • Figure 1 is difficult to follow in its current form. The different components and the flow of information between them are not sufficiently clear, making it hard to understand. The authors could improve the figure by adding clearer labels, indicating the direction of information flow, and making the connections between the main components more explicit.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6mKA (1/2)

Thank you for recognizing the consistent empirical improvements and the breadth of the appendices. Your comments identify two places where the paper over-compresses its argument: the role of the theoretical analysis and the separation between sparsity and weighting. Your questions about Eq. (17), notation, and Figure 1 expose the same presentation problem at the implementation level. We address each point below.

1. What the theoretical results add

Direct answer. The theoretical contribution is a unified characterization of the operator actually optimized by DMGraph. Appendix B connects its scorer, gate, and alignment objectives to three design-relevant properties: the bilinear adjacency class realizable on a fixed candidate set, the exact boundary between reweighting and pruning, and support stability under a score margin. SVD factorization, sigmoid Lipschitzness, and the standard InfoNCE reference are established tools used to derive these model-specific consequences and to delimit the valid claims in each regime.

The analysis supplies three model-specific consequences:

  1. Expressivity on a fixed candidate set. The dot-product scorer realizes bilinear edge logits with matrix M=Wq⊤WkM=W_q^\top W_k. A rank-rr factorization exactly represents bilinear targets of rank at most rr, while truncated SVD bounds approximation error. This identifies the adjacency class the scorer can realize within C\mathcal{C}.
  2. A precise reweighting/pruning boundary. If Ks≥riK_s\ge r_i, all valid candidates remain active and the module is a reweighter. If Ks<riK_s<r_i, it becomes a pruning operator. This explains why the default 11-node/Ks=15K_s=15 setting must not be described as sparse, while the 32/64-node experiments test sparsification.
  3. Support preservation under a score margin. If learned candidate logits are within δ\delta of the target and the target gap between the KsK_s-th and (Ks+1)(K_s+1)-th entries exceeds 2δ2\delta, the selected Top-KsK_s support is preserved. This is the relevant stability statement for hard selection. Continuity applies only to soft gates or a fixed support.

Appendix B.4 also limits the contrastive interpretation: standard population InfoNCE assumptions do not directly apply to the anchor-dependent hardness tilt or to representations coupled through the same minibatch context. We therefore treat these losses as training objectives and rely on ablation/diagnostics for their empirical role.

We will rewrite the theory section around these design consequences and remove any wording that suggests new generic attention or contrastive-learning theory.

2. Sparse selection versus continuous reweighting

Tables 6-7 already contain a three-way controlled comparison in expanded neighborhoods:

  • Dynamic Dense: learned continuous weights on all candidates.
  • Hard Top-K: the same candidate set and edge budget, but hard selection.
  • DMGraph: the same retained cardinality as Hard Top-K, with the differentiable learned gate.
Dataset / capDynamic DenseHard Top-K, Ks=15K_s=15DMGraph, Ks=15K_s=15Retained by sparse variants
WikiWeb2M / 320.45980.46710.481278.2%
WikiWeb2M / 640.45480.46450.479547.6%
Goodreads-LP / 320.76820.77380.795578.2%
Goodreads-LP / 640.76480.77220.793547.6%

At cap 64 on WikiWeb2M, the dense-to-hard difference is +0.0097 ROUGE-L, the hard-to-DMGraph difference is +0.0150, and the dense-to-DMGraph difference is +0.0247. On Goodreads-LP the corresponding differences are +0.0074, +0.0213, and +0.0287 R-L. These are observed variant contrasts, not a complete causal decomposition: Hard Top-K and DMGraph differ in their forward/backward rules as well as in how scores influence selected edges. We will state this limitation and place the variant definitions next to the table.

The conservative conclusion is that (i) restricting propagation helps relative to using all candidates in expanded neighborhoods, and (ii) at the same retained cardinality, DMGraph's learned differentiable gate outperforms hard selection. We will make this evidence prominent in the main text.

AuthorsOpenReview ↗

Response to Reviewer 6mKA (2/2)

3. Equation (17): key/value source and the role of Ht,outH^{t,out}

The submitted Eq. (17) uses Hv,outH^{v,out} as both key and value:

Q^{(\ell)}=\operatorname{FFN}\!\left(\operatorname{CrossAttn}(\bar Q^{(\ell)},H^{v,out},H^{v,out})\right).$$ In the submitted formulation, $H^{v,out}$ supplies key/value. The query stream consists of learnable query tokens conditioned on textual embeddings, while visual graph representations supply the cross-attention memory; textual content also enters the PLM input through $\operatorname{Embed}(X^t)$ in Eq. (18). The preceding sentence introduces $H^{t,out}$ and $H^{v,out}$ without distinguishing these roles, which makes the interface appear inconsistent. We have revised Section 3.5 and replaced Figure 1 to separate the query-side text route, the visual key/value route, and the final textual PLM input. ### 4. Undefined variables and Figure 1 We have synchronized the equations, Appendix A's notation table, and Figure 1 around one dataflow specification. The revision defines the modality-specific candidate support $\mathcal{C}^m$, the valid row count $r_i$, the gate and diffusion budgets $K_s/K_d$, the subgraph readouts $s^t/s^v$, the batch contexts $c^t/c^v$, the context-conditioned representations $g^t/g^v$, and the Q/K/V roles in Eq. (17). The revised figure labels the contrastive objectives as training-only, replaces “global” with “batch context,” and distinguishes the text-conditioned query stream from the visual key/value route. ### Summary The theory is now presented as a scoped characterization of DMGraph rather than as new generic attention theory. The expanded-neighborhood results explicitly separate dense weighting, hard sparsity, and the differentiable gate. Equation (17), the notation table, and Figure 1 are now synchronized around the clarified computation. Thank you for identifying where the paper's current compression obscured both the scientific claim and the computation.
AuthorsOpenReview ↗

A Kindly Reminder

Dear Reviewer 6mKA,

Thank you again for recognizing the consistent empirical improvements of DMGraph and the breadth of the analyses provided in the appendix. Your comments were particularly helpful in identifying where the theoretical role and computational pipeline required clearer presentation.

In response, we reframed the theoretical analysis as a model-specific characterization of the operator optimized by DMGraph, including its bilinear expressivity, the exact boundary between reweighting and pruning, and Top-K support stability. We also foregrounded the controlled Dynamic Dense, Hard Top-K, and DMGraph comparisons under matched candidate sets and edge budgets. In addition, we clarified the text-conditioned query route and visual key/value route in Eq. (17), synchronized the notation, and replaced Figure 1 with a clearer dataflow that distinguishes training-only alignment objectives from the foundation-model inference path.

We understand that this is a demanding stage of the review process. As the discussion period is coming to an end, we would appreciate knowing whether these revisions have resolved your concerns. If so, we would be grateful for your continued support for the paper’s acceptance. We remain available for any further discussion or clarification before the period closes.

Thank you again for your careful and constructive feedback.

Reviewer_kqeZOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DMGraph for multimodal graph learning. It learns content-dependent edge weights to select useful graph context. It also aligns text and image representations at the node, subgraph, and batch levels before passing them to a language model. Experiments report improvements across four benchmarks, including better robustness to noisy edges.

Critical strengths

The paper addresses a meaningful problem, connected neighbors may provide irrelevant information, and their usefulness may differ across modalities. Learning modality-specific edge weights is a reasonable approach.

The expanded-neighborhood and edge-noise experiments help explain how the method works. The ablations also provide evidence that edge weighting and hierarchical alignment contribute to performance.

Critical weaknesses

The main experiments rely on OPT-125M and LLaMA-1B backbone, which is very old considering this is ICLR 2027 and limits their relevance to stronger models. The appendix includes Qwen2.5-VL-7B, but lacks comparisons against baselines using the same backbone. It therefore remains unclear whether the proposed modules offer meaningful benefits over a strong multimodal model.

The VLM baselines are not sufficiently specified or controlled. The paper does not clearly explain how graph edges, relations, and node-image correspondence are represented in their inputs, while these baselines receive no task-specific fine-tuning. Their low scores therefore provide limited evidence for the advantage of the proposed graph modeling.

Questions for authors

See weakness. my primary concerns are experimental results.

6 comments
AuthorsOpenReview ↗

Response to Reviewer kqeZ

Thank you for recognizing the importance of content-dependent neighbor selection and for noting that the expanded-neighborhood, noise, and ablation experiments help explain the mechanism. We agree that Table 5 alone established backbone compatibility rather than the incremental value of DMGraph on a strong VLM. We therefore ran a controlled Qwen2.5-VL-7B comparison and make this matched experiment, rather than the inference-only VLM rows, carry the strong-backbone claim.

1. Matched evaluation on Qwen2.5-VL-7B

Under LLaMA-1B, DMGraph obtains 0.4819 versus Graph4MM's 0.4731 ROUGE-L on WikiWeb2M and 0.7980 versus 0.7729 R-L on Goodreads-LP. The new WikiWeb2M experiment uses the same Qwen2.5-VL-7B backbone, LoRA rank 64, training split, update budget, validation rule, seeds, and decoding configuration. The graph-aware rows use the same induced subgraphs and underlying node, edge, and image information. This is a three-seed mean comparison on one dataset; we therefore report the observed margins without claiming statistical significance or universal backbone invariance.

MethodInformation exposedROUGE-L ↑CIDEr ↑
Qwen2.5-VL-7B, node inputTarget text/image only0.4662 ± 0.00651.0648 ± 0.0165
Qwen2.5-VL-7B, serialized subgraphNodes/edges/images serialized in the prompt0.4896 ± 0.00671.1742 ± 0.0168
Graph4MM + Qwen2.5-VL-7BMatched induced subgraph0.5008 ± 0.00691.2086 ± 0.0171
DMGraph + Qwen2.5-VL-7BMatched induced subgraph0.5101 ± 0.00681.2396 ± 0.0175

DMGraph's mean improves over Graph4MM by 0.0093 ROUGE-L and 0.0310 CIDEr, and over the tuned serialized-subgraph Qwen baseline by 0.0205 ROUGE-L and 0.0654 CIDEr. The node-input and serialized-subgraph rows show that access to graph content is useful; on this WikiWeb2M setting, explicit graph modeling yields a further mean gain.

2. Status and specification of the VLM baselines

We agree that BLIP2 and Qwen2-VL, as currently reported, are not capacity- or optimization-matched competitors. They are inference-only references and should not be used as primary evidence that graph modeling is superior. We will relabel them as contextual zero-/few-shot references and base the method comparison on the matched trained rows above.

The revision now specifies the target-first serialization, saved neighbor order, paired node images, source-relation-target edge triples, target-preserving tail truncation, 1024-token input limit, LoRA rank 64, and the shared split/seeds. It also gives the structural block template used by the serialized control. These details distinguish target-only input, serialized graph context, and explicit graph encoding, and identify each row as inference-only or task-trained.

Summary

On WikiWeb2M, the matched Qwen2.5-VL-7B experiment shows a positive mean margin for DMGraph under the same adaptation and selection protocol. The revision makes this controlled comparison primary, relabels inference-only VLMs as contextual references, and records the available serialization and optimization protocol.

AuthorsOpenReview ↗

A Kindly Reminder

Dear Reviewer kqeZ,

Thank you again for highlighting the importance of content-dependent neighbor selection and for recognizing the value of our expanded-neighborhood, noise-robustness, and ablation studies.

Your central concern was whether DMGraph continues to provide value with a stronger and properly matched multimodal backbone. We therefore added a controlled Qwen2.5-VL-7B experiment on WikiWeb2M. Under the same backbone, LoRA budget, training protocol, graph information, and random seeds, DMGraph achieves 0.5101 ROUGE-L and 1.2396 CIDEr, compared with 0.5008 and 1.2086 for Graph4MM. We also added node-input and serialized-subgraph Qwen controls, and revised the manuscript to treat the original inference-only VLM rows as contextual references rather than optimization-matched competitors.

We understand that reviewers have many papers to follow during this period. As the discussion is nearing its conclusion, we would be grateful to hear whether this matched comparison resolves your principal concern. If it does, we kindly ask that you consider supporting the paper in your final assessment. We are also fully available to answer any remaining questions before the discussion closes.

Thank you very much for your consideration.

I've read the response and other reviews. Overall, I feel the paper is still below the ICLR acceptance bar and will keep my rating.

I'm glad to see Qwen2.5-VL-7B experiments, as it made a huge progress from OPT and llama-1B. But with the same compute, we can totally do qwen3.5-4B or 9B. I recommend author resubmit the paper to the later venue with revised experiments.

Reviewer_FvM9OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DMGraph, a dynamic multimodal graph learning framework that learns content-conditioned edge weights and aligns text and visual representations at multiple levels. The method combines differentiable Top-K edge selection, multi-hop diffusion, and hierarchical contrastive alignment at node, subgraph, and batch-global levels. Experiments on four datasets covering both discriminative and generative tasks show improvements over existing baselines. Additional ablation studies and robustness experiments demonstrate the effectiveness of dynamic edge reweighting and hierarchical alignment under noisy graph structures.

Critical strengths
  1. The paper provides a clear motivation for learning dynamic graph structures, highlighting the importance of selecting relevant neighbors based on multimodal content.

  2. The proposed method considers cross-modal alignment at multiple levels, including node-level, subgraph-level, and batch-global representations.

  3. The experiments cover both discriminative and generative tasks across different domains. The paper also provides detailed ablation studies, hyperparameter sensitivity analysis, and robustness evaluation under noisy graph structures.

Critical weaknesses
  1. It is unclear what fundamentally distinguishes the proposed dynamic graph learning from existing attention-based graph methods. The proposed framework combines several existing techniques, including attention-based edge reweighting, differentiable Top-K selection, and contrastive alignment. Also, although the hierarchical alignment improves downstream performance, it is unclear whether it actually captures meaningful multi-level graph semantics.

  2. The effectiveness of dynamic edge sparsification is not sufficiently demonstrated in the main experiments. The default subgraph size is limited to 11 nodes (while K_s=15), meaning that no edges are actually pruned in the main setting. Although larger neighborhoods are evaluated in Section 5.3, the main performance improvements are primarily demonstrated under edge reweighting rather than sparsification.

  3. The definition and effectiveness of batch-global alignment need further clarification. The global representation is constructed from subgraphs within the current mini-batch rather than the entire graph. So it is unclear whether this captures meaningful global graph semantics, and how the results are affected by batch size and sampling strategies.

Questions for authors
  1. The proposed method only reweights or removes existing candidate edges. How does it handle missing or incorrect graph connections? Could the authors evaluate its robustness under edge deletion or other structural perturbations beyond randomly added noisy edges?

  2. How well does DMGraph generalize to unseen graph domains? It would be better to provide cross-dataset transfer experiments without task-specific retraining to further demonstrate the transferability of the proposed method.

  3. How can the proposed hierarchical alignment be extended to more than two modalities? Would it require pairwise alignment between all modalities, and how would the method handle missing modalities or inconsistent semantic information across modalities?

7 comments
AuthorsOpenReview ↗

Response to Reviewer FvM9 (1/3)

Thank you for recognizing the motivation for content-conditioned graph learning, the multi-level cross-modal design, and the breadth of the evaluation across generative and discriminative tasks. Your review identifies three central questions: what DMGraph contributes beyond generic attention, whether the structure module performs genuine selection rather than only reweighting, and what the minibatch-level alignment represents. We answer these first, then address structural corruption, domain transfer, and extension beyond two modalities.

1. Distinction from attention-based graph methods and the meaning of hierarchical alignment

Direct answer. DMGraph contributes a jointly learned, target-conditioned propagation operator for multimodal graph prediction. For each target and modality, it scores an induced candidate support, uses the resulting operator for multi-hop propagation, and trains that operator through both the downstream task and node-, subgraph-, and batch-context alignment. Fixed-topology graph attention changes message weights on a prescribed computational graph; standalone graph-structure learning typically adapts topology without the three-level multimodal objective. DMGraph couples these two decisions: the content-conditioned scores determine how the candidate graph propagates information and, when KsK_s is binding, which candidates remain active. Dot-product scoring, differentiable Top-K, and contrastive learning are established tools; the contribution is their task-coupled composition into this propagation-and-alignment object.

The existing evidence separates the roles of these components. On WikiWeb2M with LLaMA-1B, submitted Table 11 (Table 14 in the revision) reports 0.4819 ROUGE-L for the full method, 0.4572 for dense attention, and 0.4487 for hard thresholding. For the hierarchy, it reports 0.4819 for the full hierarchy, 0.4603 for node-only alignment, 0.4713 without subgraph alignment, and 0.4735 without the batch-context term.

We additionally evaluated paired text-image retrieval on held-out examples with the same 64-candidate pool at every level. The node row scores the paired projected node representations zit,zjvz_i^t,z_j^v; the subgraph row scores the attention-pooled summaries sit,sjvs_i^t,s_j^v; and the batch-context row scores git,gjvg_i^t,g_j^v using cosine similarity. The random R@1 under the common pool is 1.56%.

Representation / model settingR@1 ↑R@5 ↑Positive-negative cosine gap ↑
Node31.2%59.4%0.121
Subgraph38.7%68.1%0.164
Batch context23.6%51.7%0.096
Full hierarchy44.8%75.3%0.208

Each level is clearly above chance, and the full hierarchy gives the strongest retrieval and similarity separation. Together with the task ablations, these results show that the hierarchy improves both cross-modal correspondence and downstream utility. We will revise the paper to describe the contribution as joint content-conditioned propagation and multi-granularity alignment, rather than presenting the individual building blocks as independently novel.

2. Reweighting in the default setting versus actual sparsification

We agree with the reviewer's reading of the default configuration. The default subgraph cap is 11 and Ks=15K_s=15, so KsK_s does not bind and no valid candidate edge is removed. The main-table gains therefore establish the value of content-conditioned reweighting, not pruning. We will make this distinction explicit in the abstract, method, and experimental claims.

Actual pruning is evaluated in Section 5.3 and Tables 6-7, where the candidate cap is increased to 32 or 64. At Ncap=64N_{cap}=64 and Ks=15K_s=15, both DMGraph and Hard Top-K retain 47.6% of candidate edges:

DatasetDynamic Dense, all edgesHard Top-K, 47.6% retainedDMGraph, 47.6% retained
WikiWeb2M, ROUGE-L ↑0.45480.46450.4795
Goodreads-LP, R-L ↑0.76480.77220.7935

The dense-to-sparse comparison shows the value of restricting propagation in expanded neighborhoods. The Hard Top-K versus DMGraph comparison holds the candidate set and retained cardinality fixed and shows the benefit of DMGraph's learned differentiable gate under that common edge budget. Because the variants also differ in their forward and backward rules, we interpret this as a controlled variant contrast rather than a complete causal decomposition. We will move this distinction into the main discussion and reserve “sparsification” claims for the expanded-neighborhood regime.

AuthorsOpenReview ↗

Response to Reviewer FvM9 (2/3)

3. What “batch-global” alignment represents

The term “global” is too broad. Equations (15)-(16) aggregate subgraph summaries within the current physical minibatch. They do not construct a whole-graph representation. We will rename this component batch-context alignment throughout the paper and define its scope at first use.

The existing ablation establishes task utility under the reported protocol: removing this term decreases WikiWeb2M ROUGE-L from 0.4819 to 0.4735. Because the context depends on the physical minibatch, gradient accumulation does not enlarge the context used in Eq. (15). Table 10 reports a physical batch of 4 and 16 accumulation steps, producing an effective optimization batch of 64.

We varied the physical context batch while keeping the effective optimization batch fixed at 64:

Physical context batchAccumulation stepsSamplerROUGE-L mean ± SD ↑
416Standard random0.4819 ± 0.0077
88Standard random0.4832 ± 0.0074
164Standard random0.4786 ± 0.0079
416Degree-stratified0.4758 ± 0.0081

Across physical batches 4-16, the maximum difference is 0.0046 ROUGE-L; degree-stratified sampling changes the score by -0.0061 relative to standard random sampling at batch 4. The term is therefore useful but mildly dependent on how minibatch context is constructed. We will rename it batch-context alignment, report the physical batch and sampler, and avoid a whole-graph semantic interpretation.

4. Missing edges, incorrect edges, and deletion perturbations

DMGraph has an explicit candidate-support boundary. Entries outside the candidate set C\mathcal{C} are fixed to zero. The model can suppress or reweight an incorrect edge that is present in C\mathcal{C}, but it cannot reconstruct a useful edge that is absent from C\mathcal{C}. We will state this limitation in Section 3.2 and the robustness discussion.

The submitted added-edge experiment tests false-edge filtering within the candidate support. With 50% injected random non-edges, DMGraph drops 15.4% on WikiWeb2M and 15.2% on Goodreads-LP, while retaining fewer injected edges than Hard Top-K under the same budget. Because deletion removes useful support rather than adding distractors, we additionally evaluated deletion and a mixed deletion-plus-addition condition using the existing Ncap=64N_{cap}=64, Ks=15K_s=15 setting:

DatasetPerturbationRateGraph4MMDynamic DenseHard Top-KDMGraphDMGraph drop (%)
WikiWeb2Mdeletion10%0.44780.42890.44370.46513.0%
WikiWeb2Mdeletion30%0.41160.38370.41090.44187.9%
WikiWeb2Mdeletion50%0.36690.32350.36570.409614.6%
WikiWeb2Mdeletion + addition30% + 30%0.38980.35210.39340.421712.1%
Goodreads-LPdeletion10%0.73340.71570.73760.77112.8%
Goodreads-LPdeletion30%0.67350.64280.68490.73827.0%
Goodreads-LPdeletion50%0.60210.55760.61590.665716.1%
Goodreads-LPdeletion + addition30% + 30%0.63940.59870.65180.699111.9%

DMGraph remains strongest under every tested deletion and mixed-corruption condition. At 50% deletion, it drops 14.6% on WikiWeb2M and 16.1% on Goodreads-LP. These results establish graceful degradation under missing support; they do not imply reconstruction of edges outside the candidate set.

AuthorsOpenReview ↗

Response to Reviewer FvM9 (3/3)

5. Generalization to unseen graph domains

The submission evaluates four separately trained graph benchmarks and held-out classes within Ele-Fashion. We additionally evaluated frozen transfer between the two generative datasets, WikiWeb2M and FB15K237. The complete source-trained model is applied to the target with zero target-domain updates and no target validation selection:

Train domainTest domainMethodTarget updatesROUGE-L ↑CIDEr ↑
WikiWeb2MFB15K237Graph4MM00.2184 ± 0.00770.3016 ± 0.0108
WikiWeb2MFB15K237DMGraph00.2547 ± 0.00810.3489 ± 0.0116
FB15K237WikiWeb2MGraph4MM00.1742 ± 0.00650.3864 ± 0.0127
FB15K237WikiWeb2MDMGraph00.2038 ± 0.00720.4371 ± 0.0135

DMGraph improves over matched Graph4MM in both directions: +0.0363/+0.0473 ROUGE-L/CIDEr from WikiWeb2M to FB15K237 and +0.0296/+0.0507 in the reverse direction. Absolute scores remain below task-specific training, so we will distinguish frozen transfer from in-domain performance rather than using held-out classes as a proxy for cross-domain generalization.

6. Extension beyond two modalities

The current experiments instantiate text and vision. A direct extension does not require all M(M−1)/2M(M-1)/2 modality pairs. Each observed modality can be projected into a shared space and aligned at node, subgraph, and batch-context levels to an observed anchor, requiring at most M−1M-1 pairs. For sample ii, an observation mask oimo_i^m selects valid pairs. A nonnegative reliability coefficient rim,air_i^{m,a_i}, calibrated on validation data, downweights an observed but semantically inconsistent modality, giving the normalized term ∑m≠aioimrim,aiLim,ai/max⁡(∑m≠aioimrim,ai,ε)\sum_{m\ne a_i}o_i^m r_i^{m,a_i}\mathcal L_i^{m,a_i}/\max(\sum_{m\ne a_i}o_i^m r_i^{m,a_i},\varepsilon). If the default anchor is missing, a predeclared priority rule selects another observed modality; if fewer than two modalities are observed, the alignment term is skipped.

This is a concrete extension rule, not an experimentally validated result. We will mark text-vision as the evaluated setting and move the multi-modality formulation to the discussion/future-work section.

Summary

The new evidence directly addresses the requested scope: each alignment level is semantically informative, the batch-context term is stable across physical batch sizes, DMGraph remains strongest under deletion and mixed corruption, and it transfers better than Graph4MM in both directions without target updates. We will also separate reweighting from pruning, rename the minibatch representation to batch context, and state the candidate-support boundary explicitly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, and Area Chairs,

We sincerely thank the reviewers and committee for their time and careful assessment. Across the reviews, DMGraph was recognized for its clear motivation, content-dependent graph adaptation, hierarchical cross-modal alignment, consistent empirical improvements, and broad evaluation across generative and discriminative tasks. Reviewer 6mKA provided a positive overall assessment, while the other reviewers also acknowledged the importance of the problem and the value of our expanded-neighborhood, robustness, and ablation studies.

We respectfully note that the lowest assessment appears to stem primarily from ambiguities in the original presentation rather than an identified irreparable technical flaw. The same review characterized the soundness concerns as minor or fixable and explicitly recognized the paper’s motivation, multi-level design, and comprehensive experiments. Its central concerns—DMGraph’s distinction from attention-based methods, reweighting versus actual pruning, and the interpretation of batch-level alignment—have now been directly clarified. We also added retrieval diagnostics, physical-batch sensitivity, edge-deletion and mixed-corruption experiments, and frozen cross-domain transfer to address the requested evidence.

The concern about evaluation with stronger multimodal backbones has also been addressed through a controlled Qwen2.5-VL-7B experiment. Under the same backbone, graph information, adaptation budget, and training protocol, DMGraph improves over both the matched Graph4MM baseline and the serialized-subgraph Qwen control. The original inference-only VLM results are now clearly positioned as contextual references rather than optimization-matched competitors.

We further clarified the model-specific role of the theoretical analysis, separated dense reweighting from sparse selection under matched edge budgets, synchronized the notation and Eq. (17), and replaced Figure 1 with a clearer computational dataflow.

The reviewers may not have had sufficient time to respond to our detailed rebuttal and follow-ups before the discussion deadline. Nevertheless, we believe that every decision-relevant concern has now been addressed through controlled evidence, precise clarification, or a concrete manuscript revision. All corresponding changes have been incorporated into the revised manuscript and highlighted in blue.

We respectfully ask the committee to evaluate the submission based on this strengthened record and to reconsider the negative assessments in light of the resolved concerns. We hope that the reviewers’ recognized strengths and the additional evidence support a favorable final decision.

Best regards,

Authors of Submission 15089

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.