Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models

We would like to sincerely thank Reviewer C47o for providing a detailed review with insightful questions! We will revise our paper accordingly.

W1: Explanation of terminology definition.

Thank you for your advice on the rigor of Definition 2. Although some works [1] use "modeling" as a substitute for "representation learning", this indeed lacks precision. To ensure clarity and rigor, we will revise the term "Temporal Text-attributed Graph Modeling" to "Temporal Text-attributed Graph Representation Learning" in our revised manuscripts.

As for Definition 3, we explicitly distinguish between LLMs and LMs to prevent potential misunderstandings, as some literature [2, 3] considers LLMs to be a specific subclass of LMs. However, in our work, we enhance the LLM with dynamic reasoning capabilities to generate evolving textual summaries, which are subsequently encoded into d-dimensional vectors by an LM. This separation clarifies their respective roles and helps prevent conceptual confusion.

[1] Towards Adaptive Neighborhood for Advancing Temporal Interaction Graph Modeling, KDD 2024.
[2] A Comprehensive Overview of Large Language Models, ACM TIST, 2023.
[3] GLBench: A Comprehensive Benchmark for Graph with Large Language Models, NeurIPS 2024.

W2-1: Comparison with temporal knowledge graph reasoning (TKGR).

Technical Comparison. In the domain of NLP, several LLM-based TKGR methods [1, 2, 3, 4, 5] have been proposed to enhance the graph reasoning capabilities of LLMs within temporal knowledge graphs (TKGs). We compare these methods with CROSS as follows.

• Similarity. Essentially, both works can predict future knowledge based on historical knowledge in dynamic graph-structured data.
• Key Differences. CROSS enhances LLMs for representation learning on TTAGs while LLM-based TKGR methods aim to improve LLMs for graph reasoning over TKGs. Moreover, CROSS extracts structural information via GNN-based aggregation, while LLM-based TKGR methods typically rely on explainable rule-based quadruples [1].

LLM-based TKGR methods cannot be fully applied to TTAGs, as TTAGs are not necessarily TKGs. For example, the Googlemap_CT dataset records user–restaurant review interactions rather than structured knowledge triples. Additionally, although LLM-based TKGR methods primarily focus on graph reasoning tasks, CROSS is designed to generate node representations to support various downstream applications. Therefore, these methods, while valuable, may fall outside the scope of our work. We will discuss, analyze, and appropriately cite these works in our revised paper.

Key differences are summarized as follows.

Empirical Comparison. We conduct empirical comparison with a recent LLM-based TKGR method, LLM-DA, using TiRGN as its graph-based reasoning module [1]. Other details are available in its original publication. Experiments are conducted on the ICEWS1819 dataset, as it can be structured as a temporal knowledge graph. The results are presented in Tab. 1.

From the results, we observe that CROSS consistently outperforms LLM-DA, highlighting its effectiveness once again.

Table 1: MRR result (%) for temporal link prediction on ICEWS1819.

[1] Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning, NeurIPS 2024.
[2] Back to the Future: Towards Explainable Temporal Reasoning with Large Language Models, WWW 2024.
[3] TimeR4: Time-aware Retrieval-Augmented Large Language Models for Temporal Knowledge Graph Question Answering, EMNLP 2024.
[4] Chain of History: Learning and Forecasting with LLMs for Temporal Knowledge Graph Completion, ACL 2023.
[5] ChatRule: Mining Logical Rules with Large Language Models for Knowledge Graph Reasoning, PAKDD 2025.

W2-2: Baseline clarification.

We would like to clarify that our baselines already cover both types of models: (1) those that explicitly incorporate textual information (e.g., LKD4DyTAG), and (2) those that do not (e.g., FreeDyG, DyGFormer). Under this comprehensive comparison, CROSS consistently achieves the best performance, demonstrating its effectiveness.

We will highlight this more prominently in our revised version.

Q1: Sampling of LLM reasoning timestamp.

We would like to clarify that CROSS samples LLM reasoning timestamps based on fixed interaction intervals, rather than fixed time intervals.

Motivation Analysis. As noted in Line 138, for each node, we first derive its interaction timestamps and then sample timestamps for LLM reasoning at fixed interaction intervals. As discussed in Line 135, this strategy ensures that each LLM reasoning step accesses a balanced subset of interactions, thereby mitigating the issue of uneven interaction density across different time periods.

Empirical Evaluation. To empirically validate the above analysis, we conduct an ablation study with two additional variants:

• w/ FTI (Fixed Time Interval): For each node, we compute its interaction time span and uniformly sample $m$ timestamps to enable the LLM to summarize at fixed time intervals.
• w/ RN (Recent Neighbors): For each node at each timestamp, we directly instruct the LLM to summarize the $b$ most recent neighbors.

The corresponding results are reported in Tab. 2. From the results, we find that:

• Summarizing at fixed time intervals leads to inferior performance. This confirms the presence of uneven interaction density across different time periods. CROSS addresses this issue by sampling at fixed interaction intervals, ensuring balanced input interactions at each LLM reasoning step.
• Summarizing recent interactions also yields suboptimal results. This is likely due to its narrow focus on short-term contexts, failing to capture the long-range semantic dynamics crucial for TTAG modeling.

Table 2: MRR results (%) for temporal link prediction on Enron under $b=20$.

Q2: Motivation of Co-encoder.

As stated in Line 189 and illustrated in Fig. 2 (Line 132), we clarify that the structural layer already incorporates historical structural information via TGNN-based aggregation.

Motivation Analysis. In CROSS, the semantic layer effectively captures semantic information, while the structural layer fully models structural contexts. As mentioned in the Introduction, the fundamental motivation of our Co-encoder is the assumption that semantics and structures can reinforce each other. Therefore, unification should occur throughout the entire encoding process, rather than shallow fusion solely at input or output. To achieve this, we propose to iteratively and hierarchically unify these two modalities during encoding, enabling their mutual reinforcement and producing cohesive representations for effective TTAG modeling (Lines 181–185, 206–209, and 223–224).

Empirical Evaluation. Our ablation study (Tab. 2, Line 275) provides empirical support for the encoder design. Specifically, we conduct experiments to evaluate different design choices of the encoder, including semantic-only encoding, structural-only encoding, and the proposed co-encoding architecture. Under such extensive comparison, CROSS consistently outperforms all variants, demonstrating the effectiveness and necessity of the Co-encoder design.

Once again, thank you for your valuable review. You truly have helped us to improve our work!

Thank you very much for your positive assessment! Your insightful comments have been truly valuable to our work, and we will revise the paper accordingly.

Once again, we sincerely appreciate your thoughtful feedback and support.

We would like to sincerely thank Reviewer db5y for providing a detailed review with insightful questions.

W1: Comparison with existing works.

Unlike existing works in both LLM4Graph and TGNN domains, we propose a concise yet effective framework for LLM4TTAG, which fundamentally opens new avenues for future research in TTAG modeling.

To improve readability, we first summarize the key differences between existing works and CROSS, with detailed explanations in the subsequent discussion.

Problem Novelty.

We address a largely underexplored yet practically important problem of TTAG modeling, with limited methods in this area.

Methodological Novelty.

We emphasize the novelty of CROSS by addressing the fundamental limitations inherent to both the LLM4Graph and TGNN domains:

• LLM4Graph domain: LLMs in existing LLM4Graph methods lack dynamic reasoning capabilities and typically adopt shallow output modality fusion after encoding.
  • Most LLMs are pretrained on static corpora without temporal consideration, which inherently limits their capacity for dynamic reasoning. Existing LLM4Graph methods rely on static graphs and still integrate these LLMs through static, one-off reasoning for each node, rendering them ill-suited to capture the temporal dynamics within TTAGs (Line 61).
  • Furthermore, these methods integrate structural and semantic representations only at the output stage, resulting in shallow modality fusion after encoding (Line 782). This limits the model’s ability to capture the intricate interplay between structures and semantics, ultimately leading to suboptimal representations for downstream tasks.
• TGNN domain: Existing TGNNs cannot effectively capture the semantic dynamics of TTAGs and often employ shallow input modality fusion before encoding.
  • Existing TGNNs rely heavily on their structural encoders to capture the dynamics of graph structures, overlooking the temporal evolution of text semantics within TTAGs (Line 39). As a result, their learned representations tend to be biased and overly reliant on structural information.
  • Additionally, they typically pre-embed textual information as input semantic features and incorporate them only at the input stage, leading to shallow modality fusion before encoding (Line 45). This early fusion fails to enable mutual reinforcement between structures and semantics, thereby neglecting their interplay for TTAG modeling.

To tackle the above limitations, in this paper, we propose CROSS, which (1) enhances LLMs with dynamic reasoning capabilities to capture semantic dynamics; and (2) facilitates hierarchical modality fusion throughout the entire encoding process. Specifically:

• We introduce a temporal-aware prompting paradigm to empower LLMs with dynamic reasoning capabilities for extracting semantic dynamics. Unlike statically employing LLMs, as discussed in Line 125, we propose a novel temporal reasoning chain paradigm to advance LLMs with dynamic semantic reasoning capabilities for TTAG modeling. It adopts a recurrent prompting paradigm to dynamically summarize the evolving neighborhoods of nodes, effectively extracting semantic dynamics within TTAGs.
• We propose a cohesive architecture to facilitate hierarchical modality fusion throughout the entire encoding process. Unlike shallow fusion solely at input or output, as stated in Line 158, we design a novel co-encoding architecture that encourages information propagation at the layer-encoding stage, enabling hierarchical modality fusion during encoding. It facilitates deep cross-modal unification and allows both modalities to mutually reinforce each other, leading to cohesive representations for TTAG modeling.

W2, L1: Zero-shot and extension to edge classification.

Our focus is on TTAG representation learning. Therefore, we argue that text generation falls outside the scope of this work. Nevertheless, CROSS can be easily extended to edge classification. For completeness, we evaluate two edge classification settings:

• Supervised setting: Models are trained directly using edge labels.
• Zero-shot setting: Models are first trained on temporal link prediction, and a 2-layer MLP is subsequently fine-tuned on top of the frozen encoder for edge classification.

Both implementation and model configurations follow DTGB, with DyGFormer adopted as the backbone. The results on Enron are reported in Tabs. 1 & 2. From these results, we observe that:

• CROSS still achieves the best performance in both supervised and zero-shot edge classification, proving its strong generalization and transferability across tasks.
• All models exhibit a performance drop in the zero-shot setting. This highlights the presence of a transfer learning challenge in TTAG modeling, which warrants future research.

Table 1: Precision, Recall, F1 (%) results for supervised edge classification.

Table 2: Precision, Recall, F1 (%) results for zero-shot edge classification.

W3: Costs of LLM calls and model training.

We analyze the cost of LLM calls in Sec. G (Line 806) and provide an empirical efficiency study of model training in Sec. H (Line 835).

• Despite involving LLM calls, CROSS remains applicable for large-scale TTAGs. We report the token cost and time consumption of LLM calls in Tab. 7 (Line 817). From these results, we find that CROSS maintains acceptable LLM costs even on our industrial dataset, demonstrating its scalability for large-scale TTAGs.
• During training, CROSS exhibits no significant extra overhead compared to its backbone, as they share the same number of encoding layers. We report the training times of models in Fig. 13 (Line 841). The cost of CROSS does not increase significantly compared to its backbone. This is because we strictly follow the hyper-parameters of DyGLib during training (Line 750), ensuring the same number of layers between CROSS and its backbone.
• CROSS can adjust the number of encoding layers to fit different computational resources while consistently achieving the best results. As shown in Fig. 6 (Line 314), CROSS demonstrates remarkable robustness to the number of layers. This indicates that CROSS can flexibly adapt its encoding layers as needed while achieving the best performance.

We acknowledge that computing MRR is relatively more time-consuming than AUC. However, this overhead is not introduced by CROSS itself. We adopt MRR because it is a discriminative and increasingly popular metric in temporal graph learning, as detailed in W5 to Reviewer B85N. We also report other metrics with AUC and Hits@10 in Tab. 6 (Line 797).

Q1: Prompt adaptation for different datasets.

As detailed in Lines 801–805, we have carefully considered prompt adaptation across datasets. Only minor keyword adjustments are required for the prompts when applied to different datasets. For instance, in the Enron dataset, the terms “item” and “review” are replaced with “user” and “email”, respectively. Further details of prompt design can be found in Sec. F (Line 798).

Q2: Motivation of Mixer.

We discuss the motivation of our Mixer in Lines 213-219, and all design choices are empirically validated in ablation study (Line 322).

Motivation Analysis. As mentioned in the Introduction, the key motivation of our Mixer is the assumption that semantics and structures can reinforce each other. Consequently, unification should occur throughout the entire encoding process rather than solely after encoding (Lines 181–185, 206–209, and 223-224). Moreover, as discussed in Lines 216–219, the Mixer integrates only the last semantic representation based on three considerations: efficiency, temporal awareness, and empirical evidence where mixing all semantic representations leads to suboptimal performance.

Empirical Evaluation. Our ablation study (Line 322) provides empirical support for the Mixer's design:

• Mixing at output leads to suboptimal performance. The w/o CM variant (mix at output) exhibits performance degradation. This suggests that fusing bi-modal information only after encoding fails to exploit their mutual reinforcement, resulting in suboptimal performance.
• Mixing all semantic representations also leads to suboptimal performance. The w/ CM$_{\text{all}}$ variant (mix all semantics) also exhibits performance degradation. This is because aggregating all semantic representations into a single vector will lead to uniform attention scores in the next Transformer layers (since this vector closely resembles other semantic representations), making representations indistinguishable and ultimately degrading their quality.

Once again, thank you for your invaluable suggestions. Your questions were truly helpful, and we hope our responses have addressed your concerns satisfactorily.

We are sincerely grateful for your positive assessment and for raising your score! We truly appreciate your thoughtful recognition of our rebuttal, particularly regarding the novelty and scalability of CROSS. We will revise our paper based on your reviews.

Once again, thank you for your valuable insights. Your detailed comments have been immensely helpful to us!

Many thanks to Reviewer Fqmi for providing thorough and insightful comments.

W1: Comparison with existing works.

Unlike existing works in both LLM4Graph and TGNN domains, we propose a concise yet effective framework for LLM4TTAG, which fundamentally opens new avenues for future research in TTAG modeling.

To improve readability, we first summarize the key differences between existing works and CROSS, with detailed explanations in the subsequent discussion.

Problem Novelty.

We address a largely underexplored yet practically important problem of TTAG modeling, with limited methods in this area.

Methodological Novelty.

We emphasize the novelty of CROSS by addressing the fundamental limitations inherent to both the LLM4Graph and TGNN domains:

• LLM4Graph domain: LLMs in existing LLM4Graph methods lack dynamic reasoning capabilities and typically adopt shallow output modality fusion after encoding.
  • Most LLMs are pretrained on static corpora without temporal consideration, which inherently limits their capacity for dynamic reasoning. Existing LLM4Graph methods rely on static graphs and still integrate these LLMs through static, one-off reasoning for each node, rendering them ill-suited to capture the temporal dynamics within TTAGs (Line 61).
  • Furthermore, these methods integrate structural and semantic representations only at the output stage, resulting in shallow modality fusion after encoding (Line 782). This limits the model’s ability to capture the intricate interplay between structures and semantics, ultimately leading to suboptimal representations for downstream tasks.
• TGNN domain: Existing TGNNs cannot effectively capture the semantic dynamics of TTAGs and often employ shallow input modality fusion before encoding.
  • Existing TGNNs rely heavily on their structural encoders to capture the dynamics of graph structures, overlooking the temporal evolution of text semantics within TTAGs (Line 39). As a result, their learned representations tend to be biased and overly reliant on structural information.
  • Additionally, they typically pre-embed textual information as input semantic features and incorporate them only at the input stage, leading to shallow modality fusion before encoding (Line 45). This early fusion fails to enable mutual reinforcement between structures and semantics, thereby neglecting their interplay for TTAG modeling.

To tackle the above limitations, in this paper, we propose CROSS, which (1) enhances LLMs with dynamic reasoning capabilities to capture semantic dynamics; and (2) facilitates hierarchical modality fusion throughout the entire encoding process. Specifically:

• We introduce a temporal-aware prompting paradigm to empower LLMs with dynamic reasoning capabilities for extracting semantic dynamics. Unlike statically employing LLMs, as discussed in Line 125, we propose a novel temporal reasoning chain paradigm to advance LLMs with dynamic semantic reasoning capabilities for TTAG modeling. It adopts a recurrent prompting paradigm to dynamically summarize the evolving neighborhoods of nodes, effectively extracting semantic dynamics within TTAGs.
• We propose a cohesive architecture to facilitate hierarchical modality fusion throughout the entire encoding process. Unlike shallow fusion solely at input or output, as stated in Line 158, we design a novel co-encoding architecture that encourages information propagation at the layer-encoding stage, enabling hierarchical modality fusion during encoding. It facilitates deep cross-modal unification and allows both modalities to reinforce each other, leading to cohesive representations for TTAG modeling.

W2: Scalability and ablation.

Scalability. We provide the scalability clarification in Sec. G (Line 806) and present an empirical efficiency study in Sec. H (Line 835).

• From the cost results of LLM calls, we find that CROSS remains applicable for large-scale TTAGs, validating its scalability. We report the token cost and time consumption of LLM calls in Tab. 7 (Line 817). From these results, we find that CROSS maintains acceptable LLM costs even on our industrial dataset, proving its scalability for large-scale TTAGs.
• From the training time results, we find that CROSS exhibits no significant extra overhead compared to its backbone, further validating its scalability. We also report the training time of CROSS in Fig. 13 (Line 841). From these results, we find that the training cost of CROSS does not increase significantly compared to its backbone, proving its scalability once again.

Ablation. We perform an ablation study across different LLMs in Fig. 8 (Line 354) and a parameter study for $m$ in Fig. 14 (Line 852). Lowering LLM capacity or reducing $m$ will lead to slight performance drops. Nevertheless, we emphasize that all these variants still outperform baselines, reaffirming the effectiveness of CROSS.

We will revise our manuscript to highlight the above aspects more prominently.

W3: Baseline and hyper-parameter.

Directly using existing TGNNs to encode dynamic semantic features brings marginal gains or even slight degradation. In the "Structural Encoding" column of Tab. 2 (Line 275), we report the performance of existing TGNNs when fed with raw static texts or dynamic LLM-generated texts. Under this setting, the performance of LLM-generated texts yields marginal gains or even slight degradation. We attribute this to the excessive irrelevant temporal semantics from high-order relations, which pollutes the representations.

CROSS and its backbone share the same set of well-tuned hyper-parameters, ensuring a fair comparison. As noted in Line 750, we strictly follow the hyper-parameters from DyGLib for CROSS and baselines, which have been carefully selected via an exhaustive grid search.

Q1: Motivation of Mixer.

Motivation Analysis. As discussed in Lines 216–219, the Mixer integrates only the last semantic representation based on three considerations: efficiency, temporal awareness, and empirical evidence where mixing all semantics leads to suboptimal results.

Empirical Evaluation. Our ablation study (Line 322) provides empirical support for the Mixer design. From the results of the w/ CM$_{\text{all}}$ variant (mix all semantics), we find that mixing all semantic representations leads to suboptimal performance. This is because aggregating all semantic representations into a single vector will lead to uniform attention scores in the next Transformer layer (since this vector closely resembles other semantic representations), making representations indistinguishable and ultimately degrading their quality.

Q2: Comparison with LLM4DyG.

Technical Comparison. We first provide a technical comparison between CROSS and LLM4DyG:

• Similarity. Both works explore the spatiotemporal reasoning capabilities of LLMs on dynamic graph-structured data.
• Key Differences. CROSS is an LLM-based framework that enhances LLMs' dynamic reasoning capabilities for TTAG representation learning, while LLM4DyG is a benchmark that promotes LLMs' spatiotemporal reasoning capabilities via its designed prompting schemes. Moreover, LLMs in CROSS act as Enhancers for providing dynamic semantics, while in LLM4DyG, they serve directly as Predictors. Additionally, CROSS adopts the time-ordered text stream to describe the graph, whereas LLM4DyG employs triplets with node IDs.

The key differences are outlined below.

Empirical Comparison. We also conduct an empirical comparison between CROSS and LLM4DyG. For LLM4DyG, we adopt its prompting strategy (illustrated in Tab. 7 of its publication) to instruct DeepSeek-v2 to perform temporal link prediction. The results are reported in Tab. 1.

From the results, we observe that CROSS significantly outperforms LLM4DyG. This demonstrates that LLM-as-Enhancer can better exploit the spatiotemporal modeling capabilities of LLMs for TTAG modeling.

Table 1: F1 result (%) for temporal link prediction.

Q3: Alternative theoretical analysis.

Here, we provide an alternative theoretical analysis to prove the effectiveness of CROSS, showing that unifying text semantics and graph structures instead of solely considering graph structures is bounded.

Theorem 1 (Information Bound). Given the target task $Y$, there exists a constant $\beta\in(0,1]$ such that:

I(Z_G,Z_T;Y)\geq I(Z_G;Y)+\beta\min\\{H(Y|Z_G),H(Z_T|Z_G)\\}, \tag{1}

where $Z_G$ denotes the information derived from graph structures, $Z_T$ represents the information captured from text semantics, $I(\cdot|\cdot)$ denotes mutual information, and $H(\cdot|\cdot)$ depicts the condition entropy.

Due to character limitations, we could provide the corresponding proofs at the discussion phase (if permitted). Please feel free to request.

Q4: Clarification of Tab. 2.

As stated in Line 275, the "Semantic Encoding" column of Tab. 2 refers to the setting where only the semantic encoder is used, without any structural encodings from TGNNs. Consequently, the results under this setting are independent of TGNNs, and thus identical values are reported across them.

Thanks again for your invaluable suggestions! We hope we have addressed your concerns.

Many thanks to Reviewer B85N for the thorough and insightful comments, and we will revise our paper accordingly.

W1: LLM cost.

Clarification for LLM cost. We would like to clarify the misunderstanding of the costs for LLMs:

• For each node, timestamps are sampled over its full interaction timestamps, resulting in only $m$ LLM calls per node instead of $n \times m$. As stated in Eq. 1 (Line 138), we sample $m$ timestamps from the complete $n$ interaction timestamps for each node. This reduces the number of LLM calls per node from $n$ to $m (m \ll n)$.
• LLMs are not repeatedly invoked at every timestamp. As illustrated in the pseudo-code (Line 635), CROSS pre-generates $m$ summaries for each node in advance. During co-encoder training, for node $u$ at time $t$, we only need to derive $u$'s pre-generated summaries that occur before $t$ to construct its representation, thus avoiding redundant LLM calls.
• Sampling ensures that the number of LLM calls remains controllable, effectively balancing temporal granularity and cost efficiency. As shown in Line 133, by restricting the number of LLM calls to $m (m \ll n)$, we avoid the impractical costs from invoking LLMs at every timestamp. We acknowledge that this sampling may unavoidably introduce information loss, but it indeed helps CROSS to achieve the trade-off between temporal granularity and cost efficiency. We also report the results across different sampling intervals in Fig. 14 (Line 851).

Evaluation for LLM cost. We also present the empirical cost statistics of LLM calls in Tab. 7 (Line 817). From the results, we observe that CROSS maintains acceptable LLM costs even on our industrial dataset, demonstrating its scalability for large-scale TTAGs.

W2: Lengthy prompt.

As illustrated in Fig. 12 (Line 805), historical interactions within the prompt are organized in reverse chronological order. For each prompt, we impose a token budget of 4K tokens. Once the prompt exceeds this limit, earlier interactions are truncated.

W3: Rationality of model design.

CROSS consists of two decoupled components with different responsibilities for TTAG modeling:

• From the data perspective, the LLM is dynamically enhanced to generate evolving summaries that capture semantic dynamics, which is responsible for enriching the textual modality for the original TTAG data.
• From the model perspective, the Co-encoder hierarchically unifies semantic and structural information to produce representations, which is responsible for specific spatiotemporal modeling and task adaptation.

Based on the above clarification, we argue that the design of CROSS (including LLM usage) is reasonable. This is because:

• In CROSS, the LLM is designed as an Enhancer for data augmentation, rather than a Predictor that requires fine-tuning for task adaptation. CROSS dynamically harnesses the LLM to capture the semantic dynamics over time, generating reusable dynamic summaries for semantic data augmentation. This LLM is not designed for label generation or task-specific prediction, making fine-tuning unnecessary.
• In CROSS, task adaptation is achieved through the Co-encoder, not the LLM. Recent studies [1, 2] reveal that directly using LLMs for task prediction exhibits limited effectiveness in spatiotemporal problems. Empirical results from Q2 to Reviewer Fqmi and W2 to Reviewer C47o further confirm these limitations in the context of TTAG modeling. These observations highlight the important role of a Co-encoder in bridging the gap between LLMs and TTAG modeling, enabling the learning of expressive representations for downstream predictions.
• Off-the-shelf LLM already yields highly effective summaries that enable CROSS to achieve state-of-the-art performance. As presented in Tab. 1 (Line 242), off-the-shelf LLMs can already generate sufficient summaries to support the best performance in TTAG modeling. This empirical evidence further validates the rationality of model design.

[1] Are Language Models Actually Useful for Time Series Forecasting?, NeurIPS 2024.
[2] Can LLMs Understand Time Series Anomalies?, ICLR 2025.

W4: Text-driven dynamics.

Our prompting paradigm guides LLMs to capture the semantic evolution over time, effectively facilitating text-driven dynamics for TTAG modeling. This is evident by:

• Ablation study for text-driven dynamics. In our ablation study, we evaluate the w/o TRC variant, where reasoning timestamps and their order are shuffled. This prevents the LLM-generated summaries from reflecting the accurate temporal dynamics within TTAGs. As shown in Fig. 7 (Line 330), the w/o TRC leads to a substantial performance drop, confirming the presence and significance of text-driven dynamics.
• Ablation study for time encoding. We further evaluate the contribution of time encoding. We introduce an additional variant, w/o TE, where time encoding is directly removed. From the results in Tab. 1, we find that the w/o TE exhibits only a marginal performance drop. This indicates that time encoding alone captures limited temporal dynamics for TTAG modeling.

Table 1: MRR results (%) for temporal link prediction using TGN backbone.

W5: MRR metric.

We report model performance using AUC and Hits@10 in Tab. 6 (Line 797), and we choose MRR as the main metric because:

• Similar trends are observed across all metrics, where CROSS performs best across all datasets and metrics.
• MRR has become an increasingly popular metric in temporal graph learning [1, 2].
• CROSS boosts many backbones to achieve high scores (> 98%) on metrics like AUC/AP, resulting in similar values across variants. This makes it hard to compare CROSS's performance across different backbones.

[1] TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction, ICML 2025.
[2] TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs, NeurIPS 2024.

W6: Clarification of robustness study.

Our robustness study is both well-motivated and reasonable:

• Motivation. We clearly state the motivation of our robustness study in Lines 274–276: From the results in Tab. 1 (Line 242), we find that CROSS produces similar results across different structural encoding backbones. We hypothesize this stability is due to the robustness of text semantics, and thus design a robustness study to validate this assumption.
• Rationality. We conduct a robustness study under different graph structural environments, including structural perturbations (e.g., injected noise) and structural expansions (e.g., increased encoding layers). This allows us to evaluate model performance in more degraded or complex structural scenarios. The observed robustness of CROSS emphasizes the significance of text semantics in maintaining effectiveness for TTAG modeling.

W7: Comparison with MoMent.

We would like to clarify that we do not claim to be the first to explore TTAG modeling. In fact, we explicitly mention a recent TTAG method, LKD4DyTAG (AAAI 2025), and include it in our empirical comparison (Tab. 1, Line 242). Our work presents the first attempt to unify text semantics and graph structures with LLMs for TTAG modeling.

Technical Comparison. We first compare CROSS with MoMent [1]:

• Similarity. Both works adopt a multi-modal perspective for TTAG modeling.
• Key Differences. CROSS emphasizes modality fusion, facilitating deeper unification of semantic and structural signals. In contrast, MoMent focuses on modality alignment via a pre-existing alignment loss. Additionally, CROSS advances LLMs to explicitly capture temporal dynamics within the semantic modality, while MoMent still relies on static node texts to derive semantic features.

The key differences are summarized below.

Empirical Comparison. We notice that MoMent was not accepted by KDD 2025, which remains an arXiv preprint without official code. Therefore, we re-implement it based on the details in its manuscript. We use TGN as the backbone.

From the results in Tab. 2, we find that CROSS consistently outperforms MoMent, further validating the effectiveness of CROSS.

Table 2: MRR results (%) for temporal link prediction.

[1] Unlocking Multi-Modal Potentials for Dynamic Text-Attributed Graph Representation, Arxiv 2025.

Q1: Summarizing recent neighbors.

We conduct additional experiments by summarizing the $b$ most recent neighbors for each node at each timestamp, leading to the variant w/ RN. Notably, this significantly increases the number of LLM calls, resulting in substantial costs. As shown in Tab. 3, simply summarizing recent neighbors results in inferior performance. This is likely due to the narrow focus on short-term contexts, which fails to capture the long-range semantic dynamics for TTAG modeling.

Table 3: MRR results (%) for temporal link prediction on Enron under $b=20$.

Q2: Clarification of $k$.

During training, for node $u$ at time $t$, we incorporate all $k$ summaries that occur before $t$ to derive its semantic representation. This avoids information leakage from future texts.

Thank you again for your valuable comments. Your feedback truly helps a lot!

Thanks for your feedback. We provide further clarification below.

In our preliminary experiments, we evaluate model performance using five metrics: AP, AUC, MRR, NDCG@3, and Hits@10. We select MRR as the main metric because it is the most discriminative and increasingly popular in temporal graph learning, as discussed in W5.

We report AUC and Hits@10 results in Tab. 6, omitting some datasets for brevity. This is because the same trend holds across all datasets, with CROSS achieving the best performance.

To further confirm this, we retrieve our training logs and compile the complete results for all datasets. From the results, we find that:

• CROSS consistently outperforms all baselines across all datasets and metrics.
• Our reported baseline scores are generally higher than those of DTGB. This is due to the difference in text embedder: DTGB employs early BERT for text embedding, while CROSS adopts the advanced MiniLM (Line 258), which provides greater efficiency and effectiveness.
• AUC and AP scores tend to be less distinguishable compared to MRR, reinforcing the benefits of MRR.

Once again, thanks for your suggestions. We will include these results in our revised paper.

Table 1: AP results (%) for temporal link prediction.

Table 2: AUC results (%) for temporal link prediction.