Comment for Reviewer CXWD:

Thank you very much for your professional review and valuable suggestions. We have carefully considered and responded to the questions you raised.

$\color{blue}{W.1:}$ The writing could be improved. Some figures should be polished.

$\color{blue}{Re:}$ We really appreciate your comments. We will thoroughly revise the manuscript to improve the overall writing quality. Additionally, we'll review and refine all figures to ensure clarity and visual appeal.

$\color{blue}{W.2:}$ LLM-DA leverages GPT 3.5 as the LLM model, while the other baselines use only 7B LLM models. The authors should also conduct experiments to study the impact of different LLM models.

$\color{blue}{Re:}$ Thank you very much for your professional comments. We first want to point out that there are some larger LLMs (e.g., GPT-NeoX: 20B, and Mixtral-8x7B: 56B) selected as baselines for fair comparison. To further demonstrate the flexibility of our method, we conduct experiments about using different LLMs (Qwen 1.5 chat (7B) and Qwen 1.5 chat (72B)) in our proposed method. The experimental results on both datasets are shown as follows:

For your convenience, we list the experimental results of baselines and the original results (using ChatGPT 3.5 as LLM) as follows:

From the above tables, it is evident that both the new LLMs, Qwen 1.5 chat (7B) and Qwen 1.5 chat (72B), perform better than the baselines, and achieve comparable performance to our proposed method (ChatGPT-3.5). This demonstrates that the proposed dynamic adaptation mechanism effectively captures temporal evolution patterns, allowing the generated rules to generalize better to future data. This validation across multiple model scales and architectures underscores the robustness of our approach, confirming its capability to leverage different LLMs effectively.

In the final version, we will include additional experiments and provide a more comprehensive analysis to examine the impact of different LLMs.

Comment for Reviewer thzY:

Weaknesses:

$\color{blue}{W.1:}$ There is limited analysis of the constrained Markovian random walks.

$\color{blue}{Re:}$ Thanks for your suggestions. In the paper, we have provided theoretical analysis of the constrained Markovian random walks in Appendix A from the aspects of "Impact of Time Weighting" and "Traversal Properties". Specically, the time weighting enhances the model's sensitivity to recent data, improving its ability to capture short-term changes, while traversal properties ensure that the exploration of paths respects temporal continuity, leading to more accurate insights.

$\color{blue}{W.2:}$ While Figure 1 intuitively reflects the paper's motivation, it could be improved. Moreover, the dynamic adaptation strategy, which should be the core focus, occupies a relatively small proportion in Figure 2.

$\color{blue}{Re:}$ We appreciate your kind suggestion. We will refine the Figure 1 and Figure 2 in the final manuscript.

$\color{blue}{W.3:}$ Several minor grammatical and expression issues need attention.

$\color{blue}{Re:}$ Thank you for pointing out the minor grammatical and expression issues in our manuscript. We will thoroughly review and correct these issues in the final manuscript to ensure clarity and readability.

Questions:

$\color{blue}{Q.1:}$ Do the other modules in the paper, such as the contextual relation selector and the graph-based reasoning function, also require pre-training in addition to the LLMs?

$\color{blue}{Re:}$ Thank you for your insightful question. Both the contextual relation selector and the graph-based reasoning function are freezed without any tuning in our framework.

$\color{blue}{Q.2:}$ The current experimental analysis is too superficial. To better validate the proposed method, a more in-depth analysis of the experiments is needed.

$\color{blue}{Re:}$ Thank you for your suggestions. To further evaluate the proposed method, we explore the generalizability of our framework with different LLMs. Specifically, we replace the closed-source ChatGPT with two open-source LLMs of different sizes. (e.g., Qwen-1.5-chat-7B and Qwen-1.5-chat-72B). The detailed results can be found in our response to reviewer CXWD. W2. For your convenience, we compile the results showing the impact of different LLMs on experimental performance as follows:

Experiment results show that the proposed method is flexible with the used LLMs showing the generalizability of our framework. Meanwhile, this provides deeper insights into the selection of LLMs and further validate the significant advantages of our proposed method in handling temporal data.

In the final version, we will add more detailed experimental analysis to thoroughly demonstrate the robustness and effectiveness of our proposed method.

Comment for Reviewer 9wEQ:

Thank you very much for your professional review and valuable suggestions. We have carefully considered and responded to the questions you raised.

$\color{blue}{W.1:}$ Long-horizon forecasting concerns.

$\color{blue}{Re:}$ Thank you for your insightful comments regarding the long-horizon forecasting. We want to clarify that we have conducted the long-horizon forecasting experiment in the paper, which can be found in Appendix C.5 (Figures 6 and 7) and Sec 5.3 RQ2 (Figure 4).

-Long-horizon Forecasting (time point): To verify that the proposed method can generalize to diverse further distributions, we conduct the long horizontal link prediction in Appendix C.5. In this experimental setup, the model is adapted using only data available up to time point $t$, and then employed for predictions at subsequent time points $t+k\triangle T$ without any further adaptation. Here, $k$ controls how far we predict and $\triangle T$ denotes the interval between time points. In experiments, we vary $k$ to evaluate the performance over different horizons. Results depicted in Figures 6 and 7 demonstrate that our proposed LLM-DA maintains strong performance across various time points, indicating its robustness in dynamically adapting to new data patterns over long horizons.

-Long-horizon Forecasting (time interval): We further evaluate the long-horizon forecasting results from the perspective of time intervals. Specifically, we divided the timeline into several intervals in chronological order, and these findings are thoroughly detailed in Section 5.3, RQ2. As shown in Figure 4, our method consistently outperforms existing methods across different time intervals, which validates the effectiveness of using LLMs to capture the evolving patterns of temporal knowledge.

In the final version, we would add more details about the experiment settings and clarify the confusion.

$\color{blue}{W.2:}$ Limited evaluation of generated rules.

$\color{blue}{Re:}$ We appreciate your point about evaluating the generated rules. Given the large number of generated rules, a comprehensive manual evaluation is impractical. Instead, we use the quantitative metric Support to evaluate the quality of rule $\rho$, which represents the number of facts in KGs that satisfy the rules. Support is important for assessing rules, as it shows how often the facts can be inferred by the rules within KGs.

Specifically, the rule $(\rho)$ defines the relation between two entities $e_s$ and $e_o$ at timestamp $t_l$,

$\rho: = r(e_s, e_o, t_l) \leftarrow \wedge_{i=1}^{l-1}r^*(e_s,\ e_o,\ t_i),$

where the left-hand side denotes the rule head with relation $r$ that can be induced by ($\leftarrow$) the right-hand rule body. The rule body is represented by the conjunction ($\wedge$) of a series of body relations $(r^* \in$ {$r_1,...,r_{l-1}$}).

Support $s_\rho$ denotes the number of facts in Temporal Knowledge Graphs (TKGs) that satisfy the rule $\rho$, which is formally defined as:

$s_\rho := (e_s, e_o, t_l) : \exists\wedge_{i=1}^{l-1}r^*(e_s,\ e_o,\ t_i) : \text{rule}$_$\text{body}(\rho) \wedge (e_s, r, e_o, t_l) \in \mathcal{G},$

where $\wedge_{i=1}^{l-1}r^*(e_s,\ e_o,\ t_i)$ denotes a series of facts in KGs that satisfy the rule body \text{rule}$_$\text{body}(\rho) and $e_s, r, e_o, t_l$ denotes the fact satisfying the rule head $r$.

Following the Support metric, we can evaluate the quality of the rule $\rho$, and the results are shown as follows:

In this Table, $s_\rho$ is the generated rules of LLMs, $s_\rho$ (filtered) denotes the filtered rules in Section 4.4, Eq. (8), which is the final rules used for rule application part; and $s_\rho(ALL)$ indicates the generated rules of LLMs are applied in the whole dataset.

From the results, we can see that rules generated by LLM achieve a high support, indicating the good quality of them. In the final version, we will add experiments for rule evaluation in the appendix and conduct a comprehensive analysis.