acceptodds
Under review as a conference paper at ICLR 2027

HERO: Preserving Structure and Semantics in Heterogeneous Continual Graph Learning

22 views

Abstract

Machine learning on heterogeneous graphs has advanced rapidly, driven by the diverse entities, relations, and semantics inherent in real-world data. However, most existing studies assume static graphs, whereas real-world graph data are often continuously updated. Continual learning in this setting is particularly challenging because historical knowledge is encoded not only in target-node representations, but also in their heterogeneous structural contexts and relation-dependent semantic responses. To this end, we introduce HERO, a heterogeneous continual graph learning framework that preserves historical knowledge at both structural and semantic levels while adapting to incoming tasks. We analyze experience replay as an approximation to the unavailable historical gradient and decompose its error into target-selection and context-reconstruction terms. This analysis shows that replaying target nodes alone is generally insufficient for relation-sensitive HGNNs, as removing typed neighborhood context can alter both representations and historical gradients. Motivated by this, HERO employs DiSCo (Diversity Sampling with semantic Consistency) to select representative target nodes and reconstruct compact heterogeneous contexts through relation-aware multi-type neighbor expansion. To further preserve knowledge that cannot be captured by replayed subgraphs alone, HERO aligns task-conditioned predictions and heterogeneous semantic responses through multi-level knowledge distillation. A lightweight look-ahead adaptation step is additionally used to improve plasticity on incoming tasks. Experiments on four datasets with three HGNN backbones show that HERO achieves the best or tied-best average performance among strong continual graph learning baselines while maintaining competitive forgetting. Further analyses validate the roles of structural context preservation and semantic alignment in balancing continual adaptation and retention. Our code is available at https://anonymous.4open.science/r/HCGL-B37E/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Incremental Learning, Stability-Plasticity Trade-off, Semantics Preserving

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_RtTcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies heterogeneous continual graph learning. The authors argue that standard continual learning methods are insufficient for HGNNs because historical knowledge is encoded not only in target-node representations but also in their typed neighborhood contexts and relation-specific semantic interactions. To address this, they propose HERO. The authors provide theoretical analysis decomposing replay error into target-selection and context-reconstruction terms, and evaluate HERO on four benchmarks with three HGNN backbones.

Critical strengths

S1. The paper clearly identifies why standard replay and distillation fall short for HGNNs: preserving isolated target nodes does not reconstruct the typed neighborhoods and relation-specific message passing that produced historical predictions.

S2. HERO integrates three complementary mechanisms, DiSCo, HKD, and meta-learning-based adaptatiom, into a unified pipeline. The design is systematically ablated, and the interaction between retention (replay + distillation) and plasticity (meta-adaptation) is well-structured.

S3. The experiments span four heterogeneous benchmarks and three representative HGNN backbones. The authors evaluate under both standard and reordered task sequences, report multiple metrics (AP, AF, FaG), and compare against seven strong baselines from the continual graph learning literature.

Critical weaknesses

W1. HERO is evaluated under Task Incremental Learning, where the task identity is provided at inference. This is a strong assumption that sidesteps the core challenge of distinguishing among all previously seen classes without explicit task boundaries. The more realistic and widely adopted Class-IL setting is not considered.

W2. Despite claiming competitive performance, HERO falls substantially behind the Joint Train upper bound on more complex graphs. The ablation study (Table 2) reveals that removing the gradient-based meta-learning module (G-MM) causes negligible degradation on DBLP (−1.0 AP) and Yelp (−0.3 AP), with only modest increases in forgetting.

W3. While Propositions 1 and 2 decompose replay error into selection and context terms, the analysis is largely standard: Proposition 1 is a routine smoothness bound on gradient updates, and Proposition 2 is an almost tautological observation that relation-sensitive HGNNs depend on relations. The theory serves as post-hoc motivation rather than a prescriptive foundation.

W4. HKD's semantic-level alignment relies on metapath-aware or relation-aware attention coefficients, which are intrinsic to attention-based HGNNs (e.g., HAN, HGT). For backbones without explicit attention mechanisms, such as RGCN, the paper proposes an ad-hoc surrogate attention based on node similarity (Appendix A.3), which is not part of the original architecture.

Questions for authors

(1) evaluate under Class-IL without task identifiers,

(2) justify or streamline the meta-learning component given its inconsistent benefits,

(3) strengthen the theoretical analysis with more actionable bounds,

(4) validate HKD on non-attention HGNNs or reframe it as an architecture-specific technique.

5 comments

We thank the reviewer for the constructive suggestions. We address all questions point by point below.

W1/Q1. Evaluation under Class-Incremental setting (Class-IL). To further evaluate HERO in a more realistic and challenging scenario, we conducted additional experiments under Class-IL where task boundaries are unavailable at inference. As shown below, HERO achieves the best AP across all four datasets, outperforming the strongest baselines up to 6.8 percentage points. These results further demonstrate HERO's effectiveness in preserving historical knowledge and adapting to new classes without task boundaries.

MethodDBLP AP (%) ↑IMDB AP (%) ↑Freebase AP (%) ↑Yelp AP (%) ↑
Finetune48.64 ± 0.1043.37 ± 0.2320.96 ± 3.867.97 ± 0.00
EWC48.74 ± 0.1643.40 ± 0.2317.83 ± 0.227.97 ± 0.00
MAS48.78 ± 0.4843.29 ± 0.2023.72 ± 2.779.55 ± 1.91
TWP48.85 ± 0.0743.02 ± 0.8424.81 ± 0.037.97 ± 0.00
ER-GNN47.93 ± 0.6443.41 ± 3.6925.24 ± 1.578.39 ± 0.33
FTF-ER55.86 ± 5.6243.44 ± 0.2024.89 ± 0.09—
DMSG54.10 ± 6.6539.23 ± 0.0620.82 ± 6.657.97 ± 0.00
MetaCLGraph54.86 ± 9.5542.42 ± 0.5619.63 ± 4.886.59 ± 1.88
HERO60.13 ± 7.8345.77 ± 6.4629.18 ± 4.2716.35 ± 0.54
Table: Performance comparison in Class-IL setting.

W2/Q2. Comparison with Joint Train and the role of G-MM. We would like to clarify that joint train runs under an offline training setting rather than the continual learning protocol followed by HERO and other baselines. Specifically, Joint Train has unrestricted access to data from all tasks and trains on them jointly, whereas continual learning methods learn tasks sequentially with limited memory budget. Therefore, as stated in Section 4(Experiment Setup), joint train serves as an approximate upper bound rather than a directly comparable cl baseline.

Regarding G-MM, we would like to clarify that G-MM plays a complementary role within HERO. While DiSCo and HKD preserve historical structural and semantic knowledge, G-MM improves adaptation to incoming tasks. Specifically, it performs look-ahead adaptation using representative nodes and evaluates the updates against the overall objective. Gains are consistent across different datasets. As shown in Table 2, incorporating G-MM consistently improves AP across all four datasets, while also reducing forgetting.

W3/Q3. Theoretical analysis and actionable bounds. We would like to clarify that our theoretical analysis provides a mechanism-level characterization of heterogeneous replay errors, directly motivating the two-stage design of DiSCo. Beyond Propositions 1 and 2, Appendix A.1 (Corollary 1) provides an explicit upper bound on context-induced replay-gradient error:

\leq \frac{L_g L_\sigma H}{|S|} \sum_{v\in S}\sum_{r\in\mathcal R} \|W_r\|_2\,\Delta_r(v;G,\widetilde G),$$ where $\Delta_r$ captures omitted relation-specific messages and changes in aggregation weights. This bound identifies concrete sources of gradient distortion and motivates preserving relation-aware heterogeneous neighborhoods during replay. To further establish the connection between the theoretical analysis and our practical design, we additionally evaluated empirical gradient-error proxies on IMDB. Stage 1 (target-selection) reduces the target-selection error by 24.9% compared with random target selection, while Stage 2 (relation-aware expansion) reduces the context-reconstruction error by 5.1% compared with target-only replay. The complete DiSCo also achieves the lowest overall approximation error. | Replay Strategy | Rel. $\epsilon_{\mathrm{sel}}$ ↓ | Rel. $\epsilon_{\mathrm{ctx}}$ ↓ | Rel. $\epsilon_B$ ↓ | | :--- | :---: | :---: | :---: | | CM + Target-only | 0.776 ± 0.040 | 1.027 ± 0.244 | 1.455 ± 0.246 | | Random + DiSCo Expansion | 1.004 ± 0.186 | 0.982 ± 0.247 | 1.403 ± 0.160 | | **DiSCo (Ours)** | **0.754 ± 0.074** | **0.975 ± 0.161** | **1.366 ± 0.158** | *Table: Empirical gradient-error proxy comparison. Lower is better.* Overall, the theoretical bound and empirical results provide a more direct evidence. **W4/Q4. Generalizability of HKD to HGNNs without attention mechanisms.** We would like to clarify that HKD is not restricted to attention-based HGNN architectures. As introduced in **Appendix A.3**, for HGNNs without explicit attention mechanisms, semantic responses can be constructed from hidden representations and aligned between teacher and student models, enabling semantic-level distillation beyond attention-based architectures. To further validate this generalizability, we conducted additional experiments using RGCN[1], a non-attention HGNN backbone. As shown below, HERO achieves the highest AP across all three datasets, outperforming strong baselines. | Method | DBLP AP (%) ↑ | IMDB AP (%) ↑ | Freebase AP (%) ↑ | | :--- | :---: | :---: | :---: | | Finetune | 74.77 ± 1.10 | 74.64 ± 1.42 | 64.62 ± 0.65 | | EWC | 76.05 ± 1.72 | 74.77 ± 2.68 | 65.53 ± 0.27 | | MAS | 80.50 ± 1.00 | 76.54 ± 1.10 | 64.47 ± 0.56 | | TWP | 80.41 ± 2.66 | 77.17 ± 1.92 | 66.95 ± 2.39 | | ER-GNN | 80.91 ± 1.40 | 76.73 ± 1.46 | 63.80 ± 0.57 | | FTF-ER | 79.65 ± 1.42 | 72.89 ± 3.77 | 64.20 ± 0.33 | | DMSG | 71.45 ± 0.82 | 71.63 ± 8.51 | 62.79 ± 1.71 | | MetaCLGraph | 79.94 ± 0.91 | 76.53 ± 0.74 | 66.73 ± 2.96 | | **HERO** | **82.18 ± 1.25** | **77.81 ± 1.10** | **69.20 ± 1.39** | *Table: Performance comparison using RGCN as the backbone.* Together with the extension described in Appendix A.3, these results provide further evidence that HERO can generalize beyond attention-based HGNNs. [1] Modeling Relational Data with Graph Convolutional Networks

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback. We have carefully addressed your comments in our point-by-point rebuttal and provided additional experimental results to further support our contributions.

We would greatly appreciate it if you could take a look at our responses. We would be happy to provide further clarification or address any remaining questions.

Thanks again for your time.

Reviewer_zHjvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HERO, a framework for continual learning on heterogeneous graphs. It aims to reduce catastrophic forgetting while adapting to new tasks. The method has three main components: DiSCo, which selects representative nodes and reconstructs their heterogeneous neighborhoods; HKD, which preserves historical knowledge through prediction- and semantic-level distillation; and a lightweight meta-adaptation module to improve learning on new tasks. Experiments on four datasets with three HGNN backbones show that HERO achieves competitive performance and generally reduces forgetting compared with existing baselines. The paper also provides an optimization-based analysis to explain why preserving structural context is important.

Critical strengths
  1. The paper addresses an important problem and considers both structural and semantic knowledge preservation.
  2. Experiments across multiple datasets and backbones support the effectiveness of the proposed method.
Critical weaknesses
  1. The method combines several components, making its overall complexity relatively high.
  2. The theoretical analysis relies on assumptions that may limit its practical implications.
  3. Sampling strategy may introduce additional computational and memory overhead.
  4. Evaluation is limited to four datasets, and broader experiments would strengthen the claims.
  5. The paper could provide a clearer comparison with recent heterogeneous graph continual learning methods and better explain the generalizability of the approach.
Questions for authors
  1. The method combines several components, making its overall complexity relatively high.
  2. The theoretical analysis relies on assumptions that may limit its practical implications.
  3. Sampling strategy may introduce additional computational and memory overhead.
  4. Evaluation is limited to four datasets, and broader experiments would strengthen the claims.
  5. The paper could provide a clearer comparison with recent heterogeneous graph continual learning methods and better explain the generalizability of the approach.
4 comments

We thank the reviewer for the constructive suggestions.

W1/Q1. Framework complexity. We would like to clarify that HERO's components serve distinct and complementary roles. Specifically, DiSCo preserves heterogeneous structural contexts, HKD aligns historical predictions and semantic responses, and G-MM improves adaptation to incoming tasks. Their contributions are supported by the ablation results in Table 2, where removing any component degrades performance across all four datasets.

Importantly, HERO maintains competitive computational and memory efficiency. As reported in Appendix C (Table 4), HERO achieves the lowest average and peak GPU memory usage among the compared replay-based methods. Table 6 further shows that HERO is trained faster than FTF-ER and MetaCLGraph across all four datasets. Additionally, compact replay contexts and a fixed-size teacher window keep historical storage bounded. These results demonstrate that HERO's complementary design achieves strong performance with practical costs.

W2/Q2. Practical implications of theoretical assumptions. We would like to clarify that the smoothness and Lipschitz assumptions in Section 3.1 and Appendix A.1 are standard conditions used to characterize replay-gradient approximation errors.

Importantly, Corollary 1 (Appendix A.1) establishes an explicit connection between context-induced gradient error and relation-specific message discrepancies, motivating DiSCo's design. Our additional empirical analysis further supports this connection. Stage 1 (target selection) reduces the target-selection error by 24.9% compared with random selection, while Stage 2 (relation-aware expansion) achieves a 5.1% lower context-reconstruction error than target-only replay. These results demonstrate the practical relevance of our theoretical analysis and its connection to DiSCo's design.

Replay StrategyRel. ϵsel\epsilon_{\mathrm{sel}} ↓Rel. ϵctx\epsilon_{\mathrm{ctx}} ↓Rel. ϵB\epsilon_B ↓
CM + Target-only0.776 ± 0.0401.027 ± 0.2441.455 ± 0.246
Random + DiSCo Expansion1.004 ± 0.1860.982 ± 0.2471.403 ± 0.160
DiSCo (Ours)0.754 ± 0.0740.975 ± 0.1611.366 ± 0.158
Table: Empirical gradient-error proxy comparison. Lower is better.

W3/Q3. Computational and Memory overhead of sampling. We would like to clarify that DiSCo constructs compact heterogeneous replay contexts through representative target selection and bounded relation-aware neighbor expansion, avoiding the storage and computation associated with full historical neighborhoods. Specifically, Appendix C provides empirical evidence of HERO's computational and memory efficiency. As shown in Table 4, HERO achieves the lowest average and peak GPU memory usage among the compared replay-based methods. Furthermore, Table 6 shows that HERO is trained faster than FTF-ER and MetaCLGraph across all four datasets. These results demonstrate that DiSCo (our sampling method) can be achieved with practical computational costs and memory efficiency.

W4/Q4. Scope and Diversity of Evaluation. We would like to emphasize that HERO has been extensively evaluated across diverse heterogeneous graph benchmarks. As detailed in Section 4 and Appendix B.1, the four datasets contain different application domains and exhibit substantial diversity, ranging from 21K to 180K nodes, 6 to 36 relation types, and up to 32.5M edges. Furthermore, we evaluate HERO using multiple HGNN backbones, including both metapath-based and relation-based architectures. To further strengthen the evaluation, we additionally conducted experiments using RGCN as a non-attention backbone and under Class-IL setting (without task boundaries at inference), where HERO achieves the highest AP across all evaluated datasets. These results demonstrate HERO's effectiveness across diverse graph structures, model architectures, and CL settings.

W5/Q5. Comparison with recent method. We would like to clarify that methods designed for heterogeneous continual graph learning (HCGL) remain relatively scarce. Therefore, we evaluate HERO against a broad range of representative and recent continual graph learning methods, including FTF-ER (2024) and DMSG (2025).

To further strengthen the comparison, we additionally evaluated UFO (NeurIPS 2026)[1], a recent strong continual graph learning method, under our HCGL evaluation protocol. As shown below, HERO outperforms UFO by up to 19 AP points.

MethodDBLP AP (%) ↑IMDB AP (%) ↑Freebase AP (%) ↑Yelp AP (%) ↑
UFO89.2 ± 1.075.8 ± 0.956.0 ± 0.731.9 ± 12.7
HERO93.6 ± 0.878.4 ± 1.760.9 ± 1.555.9 ± 0.5
Table: Performance comparison with UFO.

Regarding generalizability, HERO has been evaluated across multiple HGNN backbones, achieving the best or tied-best AP across all configurations. Our additional experiments with RGCN and Class-IL setting further demonstrate its effectiveness. We will add these results to the revised manuscript.

[1] UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback. We have carefully addressed your comments in our point-by-point rebuttal and provided additional experimental results to further support our contributions.

We would greatly appreciate it if you could take a look at our responses. We would be happy to provide further clarification or address any remaining questions.

Thanks again for your time.

Thank you for the rebuttal, which addressed all of my concerns. I remain positive about the paper and will keep my original rating.

Reviewer_uqDwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses continual node classification on growing heterogeneous graphs. HERO combines DiSCo, a replay strategy that selects target nodes and reconstructs relation-aware contexts; heterogeneity-aware knowledge distillation (HKD), which preserves output and semantic responses; and a lightweight look-ahead adaptation step. A gradient-approximation analysis motivates retaining both representative targets and their typed neighborhoods. Experiments on DBLP, IMDB, Freebase, and Yelp evaluate three heterogeneous GNN backbones. The results support competitive average performance and forgetting, with additional ablations and task-order studies.

Critical strengths

S1. The paper targets a meaningful issue specific to heterogeneous continual learning: replaying labeled target nodes alone may fail to reproduce the typed message-passing contexts that generated historical predictions. The target-selection/context-reconstruction decomposition is a useful conceptual framing (Sec. 3.1; Propositions 1–2). S2. Combining relation-aware replay with prediction-level and semantic-response distillation is reasonable for heterogeneous GNNs. The experiments with HAN, MAGNN, and HGT suggest that the general framework is not tied to a single backbone (Secs. 3–4; Appendix D.1). S3. The paper reports meaningful main results and controlled ablations. With HAN, HERO obtains 78.4% average performance on IMDB and 55.9% on Yelp; removing HKD or DiSCo degrades performance substantially in several settings (Tables 1–2).

Critical weaknesses

W1. The theoretical argument does not yet establish a direct connection between the proposed heuristic and reduced gradient error. The decomposition follows from the triangle inequality, while Proposition 2 is an existence result. In particular, the rule selecting nodes with the smallest number of nearby different-class samples is not shown to minimize target-selection error. An empirical gradient-error proxy would help validate this connection (Sec. 3.1; Eq. (1)). W2. Replay fairness and cumulative storage deserve clearer accounting. DiSCo retains contextual nodes and edges beyond selected targets, and HKD may retain teacher models. Although the appendix studies memory and teacher windows, comparison under matched total bytes and model-storage budgets would better isolate the algorithmic gain (Appendix C). W3. The setting assumes a fixed heterogeneous schema and known task context at evaluation. Such assumptions are legitimate for the studied benchmark but limit the evidence for open-ended streams with new relation types and longer task sequences (Sec. 2; Sec. 4). W4. Equation (7) appears to index the denominator of the distillation softmax over nodes rather than classes, whereas the surrounding discussion treats the result as a predictive label distribution. Please clarify whether this is notation only or matches the implementation (Sec. 3.2).

Questions for authors

Q1. Can the authors compare DiSCo with target-only and random replay using an empirical proxy for target-selection and context-reconstruction gradient errors? Q2. Do the main improvements persist when the total stored node/edge data and teacher-model storage are matched across methods? Q3. Could the authors clarify the class-wise normalization in Eq. (7) and discuss behavior on longer or schema-evolving task streams?

2 comments

We thank the reviewer for these constructive suggestions.

W1/Q1. Empirical validation of gradient approximation errors To directly check the connection between DiSCo and the theoretical error decomposition in Section 3.1, we conducted additional experiments measuring empirical proxies for target-selection error (ϵsel\epsilon_{\mathrm{sel}}), context-reconstruction error (ϵctx\epsilon_{\mathrm{ctx}}), and overall replay-gradient error (ϵB\epsilon_B). Specifically, we compare full DiSCo against target-only replay with Coverage Maximization (CM) and random target selection with DiSCo's context expansion.

Replay StrategyRel. ϵsel\epsilon_{\mathrm{sel}} ↓Rel. ϵctx\epsilon_{\mathrm{ctx}} ↓Rel. ϵB\epsilon_B ↓
CM + Target-only0.776 ± 0.0401.027 ± 0.2441.455 ± 0.246
Random + DiSCo Expansion1.004 ± 0.1860.982 ± 0.2471.403 ± 0.160
DiSCo (Ours)0.754 ± 0.0740.975 ± 0.1611.366 ± 0.158
Table: Empirical gradient-error proxy comparison. Lower is better.

The results provide direct empirical support for DiSCo's two-stage design. Stage 1 reduces the target-selection error by 24.9% compared with random selection, while Stage 2 reduces the context-reconstruction error by 5.1% compared with target-only replay. Full DiSCo achieves the lowest overall replay-gradient error.

W2/Q2. Memory-Matched Performance Comparison. We would like to clarify that HERO achieves its improvements with competitive memory efficiency. As described in Appendix B.4, replay-based methods follow the same per-type node buffer budgets. Furthermore, Appendix C (Table 4) shows that HERO achieves the lowest average and peak GPU memory usage among the compared replay-based methods. Table 5 further shows that a fixed teacher window bounds teacher storage to O(k)O(k) while maintaining strong performance.

To further validate the improvements under matched storage constraints, we conducted an additional experiment with the same total storage budget of 11 MB, including replay data, teacher models, and other persistent model states.

MethodAP (%) ↑AF (%) ↓
ER-GNN71.67 ± 2.1020.47 ± 2.68
MetaCLGraph75.94 ± 2.7612.20 ± 4.86
HERO78.15 ± 1.637.07 ± 3.24
Table: Performance comparison under the same total storage budget on IMDB.

Under the same total budget, HERO outperforms MetaCLGraph and ER-GNN by 2.22 and 6.48 AP points, respectively, while reducing AF by 5.14 and 13.41 points. These results further demonstrate HERO's effectiveness under matched storage budget.

W3/Q3. Evolving heterogeneous graphs and longer task streams. We would like to clarify that a fixed heterogeneous schema does not imply static graph structures or semantic contexts. As defined in Section 2, while the node and relation type sets remain fixed, graph snapshots continuously evolve with newly introduced nodes and edges across tasks. Since tasks involve different classes of target nodes, their typed neighborhoods, relation instances, and semantic interactions also evolve. Preserving these evolving heterogeneous contexts is precisely the central motivation of HERO. The introduction of entirely new relation types represents a different schema-expansion setting.

For longer sequences, Appendix C shows that retaining three teachers achieves AP within 0.38 points of retaining all teachers on the 8-task Yelp setting, while bounding teacher storage to O(k)O(k).

W4/Q3. Clarification on normalization equation. We thank the reviewer for pointing out the need for clarification. Eq(7) follows commonly used temperature-scaled softmax used in knowledge distillation, where normalization is performed over the KK class logits. Formally, for node (ziz_i), teacher (ll), and class (cc), the normalized probability is pT,l,c(zi,t)=exp⁡(fT,l,c(zi)/t)∑j=1Kexp⁡(fT,l,j(zi)/t).p_{T,l,c}(z_i,t)=\frac{\exp(f_{T,l,c}(z_i)/t)}{\sum_{j=1}^{K}\exp(f_{T,l,j}(z_i)/t)}. This is consistent with our implementation, which applies softmax along the class dimension. We will make the class indices explicit in Eq.(7) in the revised manuscript.

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback. We have carefully addressed your comments in our point-by-point rebuttal and provided additional experimental results to further support our contributions.

We would greatly appreciate it if you could take a look at our responses. We would be happy to provide further clarification or address any remaining questions.

Thanks again for your time.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your constructive feedback to help us improve our work. We have uploaded a revised version and provided detailed point-by-point answers in the rebuttal. We would greatly appreciate it if you could take a look at it.

We would welcome any additional feedback or suggestions and would be happy to discuss them to further improve the paper.

Thank you again for your time and feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.