TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting

Dear Reviewer QYcP,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Note that we only report MSE metrics in this rebuttal due to its limited space.

---

[Q1-1, W3] How does your architecture differ from T5?

[Response] TimePerceiver is fundamentally different from Chronos-T5 in the perspectives about how the model operates based on the context of time. Firstly, TimePerceiver consists of a decoder with query tokens containing time and channel information about target segments. Therefore, the decoder flexibly extracts information based on the time point we want to predict from the representation captured by the encoder through the latent bottleneck, which captures temporal-channel dependencies. However, as stated in the Chronos paper, Chronos-T5 completely omits time and frequency information and treats time series as a plain sequence of tokens. The decoder works in an auto-regressive manner, looking only at the tokens generated so far and producing the probability distribution for the next single token. As a result, the model relies purely on sequential token patterns to predict the next value, without any position-specific information that is unique to time-series data.

To sum up, TimePerceiver can respond flexibly even if the input and output intervals vary each time. On the other hand, Chronos-T5 remains stuck in a token-next-token approach, making it difficult to structurally solve generalized formulation other than predicting specific future intervals.

---

[Q1-2] Please discuss how your generalized formulation can be projected to other architectures (or is this impossible?).

[Response] We believe that our generalized formulation can be easily projected to other patch-based architectures by replacing their prediction layers with our query-based decoder. PatchTST [1] is a clear example: replacing its linear decoder by our query-based decoder and adopting it with our generalized formulation yields a TimePerceiver variant that uses PatchTST’s encoder design. Not only that, but the more naturally a structure can handle temporal positional information, the more easily our idea can be projected.

[1] Nie et al., A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. ICLR 2023

---

[Q2-1] In Table 2, could the authors provide all combinations?

[Response] Below are the results for the combinations that were not shown for the standard formulation in Experiment 4.2 and Appendix D.

---

[Q2-2] It is unclear what impact positional embedding sharing has on the standard formulation.

[Response] In the standard formulation, PE sharing means that channel positional embedding (CPE) is shared across the encoder and the decoder in our architecture. In contrast, our generalized formulation, PE sharing means that both CPE and temporal positional embedding (TPE) are shared. Therefore, PE sharing has less impact in the standard formulation, as we observed a slight performance difference in its ablation study. We will clarify the ablation experiments in the final manuscript.

---

[Q3, W1] While the focus is on transformer-based methods, a global view should include other architectures. Please provide results for modern RNN methods (xLSTM-Mixer), MLPs (CycleNet, TimeMixer, TimeMixer++), and convolutional models (ModernTCN).

[Response] We provide comparison results for the methods you mentioned below. All evaluations were conducted with a lookback window L=96 for consistency. Since public results and scripts for xLSTM-Mixer under this setting are not available, we instead include SegRNN for comparison.

Experimental results show that TimePerceiver consistently outperforms the baselines. While TimeMixer++ is reported to perform slightly better than our method, it is worth noting that the reported standard deviation in the original paper is relatively high (around 0.2 across most datasets), and the code has not been publicly released. Moreover, when we run experiments using an unofficially released version of the code, we were unable to reproduce the reported results. For these reasons, we are cautious about making a direct comparison with TimeMixer++.

---

[Q4] In your ablation, you report that ETTh1 works best for contiguous sampling. Could the authors provide a rationale? Simply stating that this highlights the importance of sampling ratio is insufficient, especially since one of the main claims is that previous pretraining methods do not model this behavior correctly.

[Response] Based on our visualization and analysis of the datasets, we observed that (i) fully contiguous sampling is effective when capturing global patterns is important for the data within a single instance, whereas (ii) fully disjoint sampling performs better when the data contains frequently repeated local patterns. First, fully contiguous sampling proved effective not only for ETTh1 but also for the Weather dataset. Both datasets exhibit strong global patterns rather than repetitive local behaviors. In such cases, removing long contiguous segments from the input forces the model to infer high-level temporal dependencies, which helps in learning global trends more effectively. Second, datasets such as Solar, ECL, and Traffic showed more consistent global patterns but contained local variations, often with repetitive short-term patterns. For these datasets, fully disjoint sampling led to better performance, as it encourages the model to solve many small-scale reconstruction tasks and thus better capture local dynamics. Therefore, when prior knowledge about the characteristics of the data is available, we recommend selecting a sampling strategy accordingly to best exploit that knowledge. In the absence of such prior information, using the mixed version serves as a robust default choice. We will incorporate this clarification into the revision.

---

[W2] Masking and reconstruction tasks are common pretraining strategies but not the only ones; prior time series pretraining works include TS-TCC, TFC, TS2Vec, and others that utilize multiple self-supervised losses.

[Response] Our contribution lies not only in the generalized forecasting formulation but also in the encoder-decoder architecture that is closely aligned with this formulation. Furthermore, we believe that our framework is compatible with various self-supervised learning objectives mentioned (e.g., contrastive learning) and can be extended to incorporate them effectively.

---

[W4] While the unified view is appealing, it is confusing how this relates to general time series model analysis; if a unified view is employed, methods from imputation should also be included in the evaluation or the related work section.

[Response] Following your suggestion, we conduct imputation experiments with three datasets (ETTh1, Weather, Electricity) and three baselines (TimesNet, iTransformer, PatchTST). We here set a more challenging task of imputation via patch-wise masking with a patch length of 24, as opposed to the easier setup of masking at individual timesteps. Due to the space limit, please refer to [Reviewer DDKH, W3 Response] for the imputation results. As shown in the results, our TimePerceiver consistently outperforms the baselines across all datasets. We will add these results with more baselines into the final manuscript.

---

[W5] The experiments over different input lengths are appreciated, but results without multiple seeds are less meaningful.

[Response] In Appendix B, we clarify that all reported results are the average of 5 runs using different random seeds. Also, across all 32 experimental settings with lookback window L=384, 25 cases exhibit a standard deviation less than or equal to 0.0005, indicating that TimePerceiver remains highly stable. For details, refer to [Reviewer DDKH, W1 Response].

---

[W6] The number of latent tokens is not sufficiently explored; reporting that 8, 16, and 32 tokens work best is not enough. What happens when decreasing or increasing this number? From a practical perspective, it is unclear how to choose this hyperparameter.

[Response] Thank you for mentioning the important point. In our experiments, we simply chose the latent size based on validation. Following your suggestion, we further explore its sensitivity as reported below. We observe that the presence of a latent bottleneck clearly improves performance over having none, but the specific number of latent tokens plays only a minor role in determining the final performance, as they are all relatively small compared to the number of input tokens. This indicates that our model is not overly sensitive to this parameter, and even with a small latent size, we observe substantial performance gains.

---

Dear Reviewer QYcP,

Thank you very much for your thoughtful comments and for taking the time to review our manuscript. We are glad that our additional clarifications and experiments have addressed your concerns. We will incorporate all the findings and the discussions from the rebuttal into the final manuscript.

If you have any further suggestions, we would be happy to hear them.

Sincerely,
Authors

I thank the authors again for their engagement and elaboration. For the camera-ready/revision, the authors can think about plotting those (dim-reduced) latent tokens to give a qualitative intuition of what is happening for different configurations. Otherwise, I have no remaining concerns.

I raised my score accordingly.

[W6] I would assume a latent size of 1 yields the same outcome as 0. When does the latent-size improvement start to take effect? Is 8 the smallest value that shows a benefit?
As shown below, interestingly, even a single latent is sufficient to yield significant improvements over the no-latent case, suggesting that the presence of the latent itself plays a key role in forecasting performance. Moreover, the performance is not highly sensitive to the number of latents. This indicates that our latent bottleneck structure effectively aggregates temporal and channel dependencies, even with minimal capacity. We appreciate the suggestion to conduct this ablation study, and we believe that the resulting observations further strengthen the contribution of our latent bottleneck. We will include these results in the final manuscript.

---

Please do not hesitate to let us know if there are any remaining concerns or clarifications you would like us to address. We are more than willing to make the most of the remaining discussion time to further improve our paper.

Thank you very much,
Authors

[Q3/W1] More comparisons with recent baselines
Following your request, we conduct additional experiments with xLSTM-Mixer and TimeMixer++ on ETTh1, ETTh2, ETTm1, and ETTm2 with various lookback window sizes (L=96, 384) and forecasting horizons (H=96,192,336,720). As shown below, our TimePerceiver consistently outperforms both xLSTM-Mixer and TimeMixer++ across these diverse settings.

Based on these results, our TimePerceiver outperforms all the baselines you mentioned (CycleNet, TimeMixer, TimeMixer++, xLSTM-Mixer, ModernTCN). This demonstrates the superiority of TimePerceiver over recent state-of-the-art methods. We will include these results in the final manuscript.

Note: We are currently reproducing the results of TimeMixer++ and xLSTM-Mixer on the remaining datasets (Weather, Solar, ECL, Traffic), and we will share them as soon as they become available.

---

[Q3/W1] On which datasets or settings did your method outperform other baselines?
As shown in Appendix C, TimePerceiver demonstrates consistently strong performance across a wide range of settings. It achieves competitive or superior results in various horizons, which highlights the robustness of our framework across diverse scenarios.

Also, when combined with the discussion in [Q3, W1], our method shows remarkable stability in datasets with relatively long sequences such as ETTm1, ETTm2, Weather and ECL. The standard deviation in these settings is notably low, indicating stable results. We attribute this to our generalized formulation, which enables the model to make more robust predictions in diverse scenarios, especially when the datasets contain large numbers of instances.

Dear Reviewer QYcP,

Thank you for your comments and for taking the time to review our manuscript. We are happy that our rebuttal and experiments addressed most of your concerns.

We here provide an additional response to [Q1-2], [Q3/W1] and [W6] for addressing your remaining concerns.

---

[Q1-2] Do you by any chance have a few preliminary results for this projection?
Yes. In the early stage of our research project, we conducted a few preliminary results showing that our generalized formulation can improve other encoders. We here provide the results obtained using PatchTST as an example, under ETTh1 with a lookback window L=384. As shown below, we observed that training with the generalized formulation consistently outperforms the standard formulation across various lookback window settings. While the TimePerceiver achieves the highest performance overall, these results suggest that the benefits of the generalized formulation are not limited to a specific model and can be effectively applied to enhance various encoder architectures.

Dear Reviewer QYcP,

Thank you for your follow-up question! We appreciate the opportunity to clarify further.

We think that the latent may tend to learn redundant information as you mentioned. This could explain why performance does not vary significantly, especially when the latent dimension is large (e.g., dim=128 in the previous experiments). To examine the hypothesis, we conduct additional configurations with different latent dimensions (dim=16, 64) under the Traffic dataset. As shown below, we observe a clear performance drop as the dimension decreases.

Furthermore, we compare two different latent configurations with the same latent capacity (i.e., number of latent x latent dim): 4x16 vs 1x64. As shown below, interestingly, we observe that using more latent tokens tends to perform better. This suggests that the latent structure plays an important role in our framework.

Although finding the optimal latent configuration remains an open problem, we would like to emphasize that our latent bottleneck achieves strong performance over the no-latent case even without carefully tuned configurations. We again appreciate your comment, which helps further strengthen our paper’s contribution. We will include this new ablation study in the final manuscript.

Thank you very much,
Authors

Dear Reviewer pmxb,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Note that we only report MSE metrics in this rebuttal due to its limited space.

---

[W1] The novelty is weak, since the latent bottleneck representations of encoder is introduced in time series forecasting, “Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting” (ICLR2023). Also, there is no theoretical proofs for the bottleneck representation, please add. Therefore, the novelty is the learnable query-based decoder.

[Response] The use of latent bottleneck in TimePerceiver is fundamentally different from Crossformer [1] in the perspectives about (i) which information is aggregated into the latent and (ii) how to utilize the latent. First, our latent aggregates temporal and channel dependencies simultaneously via our latent bottleneck structure. This design explicitly captures interactions across both dimensions in a single operation. In contrast, Crossformer implicitly aggregates temporal-channel dependencies as interactions between temporal and channel information are considered separately. Here, Crossformer's bottleneck structure is used only in the cross-channel stage. Second, our latent is updated through multiple self-attention blocks. Such approach enables the latent to progressively refined through multiple self-attention blocks. The enhanced latent is then projected back to the input tokens, enabling it to incorporate more comprehensive and informative context. However, Crossformer's bottleneck structure is designed only for reducing computational cost, which is the reason why our TimePerceiver can outperform Crossformer.

[1] Zhang et al., Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. ICLR 2023

---

[W2] As a generalized forecasting model, some generalized baselines should compare, "One Fits All: Power General Time Series Analysis by Pretrained LM", "TOTEM: TOkenized Time Series EMbeddings for General Time Series Analysis" and "UniTS: A Unified Multi-Task Time Series Model".

[Response] Thank you for the suggestion. We would like to clarify that our generalized formulation differs from the use of the term generalized in foundation models in that we redefine time-series forecasting itself to include extrapolation, interpolation, and imputation in a unified framework. In contrast, the three papers you mentioned (One Fits All, TOTEM, and UniTS) treat generalized as building a single foundation model that supports a wide variety of tasks, including classification, anomaly detection, and forecasting.

Nevertheless, following your suggestion, we compare long-term forecasting results with the mentioned baselines with lookback window L=96. Note that “One Fits All” did not provide the results of L=96, we here omit the baseline. As shown below, our TimePerceiver outperforms TOTEM and is comparable to UniTS.

---

[W3] In the learnable query of decoder, did the author consider adding covariates corresponding to the prediction target to enhance the expressiveness of the query?

[Response] We constructed the query not merely as learnable parameters, but as a composition of covariates corresponding to the prediction target including temporal information (referred to TPE) and channel-specific information (referred to CPE). As shown below, experiments on the ECL dataset demonstrate that removing CPE, which encodes channel information, leads to a performance drop. This demonstrates that adding covariates corresponding to the prediction target enhances the expressiveness of the query, as you expected. We also experimented with adding an embedding for patch size (referred to PSE). However, since our trained model uses a fixed patch size, this addition had minimal impact.

---

[W4] Since the author states that TimePerceiver can encompasses extrapolation, interpolation, and imputation tasks along the temporal axis in lines 54-56, could the author do some comparisons within these tasks, "Optimal Transport for Time Series Imputation", "Csdi: Conditional score-based diffusion models for probabilistic time series imputation".

[Response] Following your suggestion, we compare our framework with PSW-I [1] under the standard imputation setup [2,3] using 6:2:2 training/validation/test splits. As shown below, our TimePerceiver outperforms the baseline, PSW-I, under the ETTh1 dataset. We will add more comparisons (with other baselines, e.g., CSDI, and other datasets, e.g., Weather) into the final manuscript.

[1] Wang et al., Optimal Transport for Time Series Imputation, ICLR 2025
[2] Wu et al., TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis, ICLR 2023
[3] Goswami et al., MOMENT: A Family of Open Time-series Foundation Models, ICML 2024

Dear Reviewer rg1y,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Note that we only report MSE metrics in this rebuttal due to its limited space.

---

[Q1] What are the computational costs (training time, memory footprint) of TIMEPERCEIVER compared to the baselines, especially given the dual attention mechanisms and bottleneck structure? Please include a runtime and memory usage comparison across different sequence lengths.

[Response] As shown below, TimePerceiver demonstrates efficiency in both training time and memory footprint compared to baselines, regardless of the sequence length. This efficiency stems from the latent bottleneck that enables compression of core information efficiently. Additional analyses, such as the impact of longer prediction lengths, are provided in Appendix B.

---

[Q2] Under what conditions does TIMEPERCEIVER struggle? Are there failure cases or performance drop-offs in certain data regimes (e.g., small datasets, irregular sampling)? Discuss known limitations and include a sensitivity analysis or case examples showing where the model might be less effective.

[Response] TimePerceiver is designed to be effective even in small datasets, as our generalized formulation enables the model to learn from a richer context beyond standard future prediction, especially by leveraging extrapolation, interpolation and imputation. However, in cases where the available time-series data is extremely short, the benefits of the generalized formulation may become less pronounced compared to the standard formulation. However, even in such cases, we observe that performance remains comparable to that of standard formulation.

---

[Q3] How does each architectural component (e.g., latent bottlenecks, learnable queries, self-attention in bottlenecks) individually contribute to performance? Please clearly discuss each innovation's role. For example, what is the performance delta when removing learnable queries alone?

[Response] We would like to note that the analysis of each component is discussed in detail in Section 4.2 component ablation studies and Appendix D. Specifically, in the analysis, we conducted experiments by replacing each architectural component of TimePerceiver with a different structure or modifying the model formulation to verify the effectiveness of each component. We observe that the latent bottleneck contributed to performance improvement by efficiently aggregating temporal-channel dependencies. Additionally, the combination of learnable queries and a generalized formulation led to performance improvements due to its ability to handle diverse temporal contexts. We here further provide additional experiments with varying the number of latent tokens. As shown below, the results show a clear performance drop without latent bottleneck, highlighting its role in effectively modeling temporal and cross-channel dependencies.

We would like to clarify that our generalized formulation cannot be implemented without our query-based decoder, so we did not provide the ablation study for our query-based decoder.

---

[Q4] Can the authors elaborate on specific real-world scenarios where the simultaneous handling of extrapolation, interpolation, and imputation is essential? Add at least one case study or detailed scenario (e.g., in health, finance, or sensor networks) where this generalized forecasting capability provides a clear advantage over traditional models.

[Response] A concrete example is climate and environmental monitoring using sensor networks that measure variables such as temperature, humidity, and air quality. In such systems, our generalized forecasting capability is essential. First, imputation or interpolation is needed when sensor failures or communication issues lead to missing data. Second, extrapolation enables the estimation of past values from periods before sensor deployment. Lastly, forecasting is required to predict future conditions for effective planning and decision-making.

We believe our unified framework, built upon our generalized formulation, can flexibly support all of these objectives, making it particularly well-suited for real-world applications like environmental monitoring. We will incorporate this discussion into the final manuscript.

---

[Limitations] Thank you for the insightful comments. We provide additional responses regarding the mentioned limitations below. We also would like to note that we discussed limitations in Appendix E of the supplementary material.

[a] Please refer to [Q2].
[b] Please refer to [Q1].
[c] We believe our framework is well-suited to address these biases, thanks to the generalized formulation that enables diverse contextual understanding and the latent bottleneck that effectively captures various temporal-channel dependencies.
[d] Thanks to the the efficiency of our framework as shown in [Q1], we believe that it is executable in resource-limited edge devices without network access (e.g., uploading/downloading private data). This data isolation could mitigate such privacy risks.

Dear Reviewer DDKH,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Note that we only report MSE metrics in this rebuttal due to its limited space.

---

[W1] Since the results were averaged over 5 runs, it could have been interesting to show the standard deviation of the results to show the statistical significance of the results (though it is not standard in the field).

[Response] Following your suggestion, we evaluate the stability of TimePerceiver by reporting mean and standard deviation over five runs with lookback window L=384. As shown below, across all 32 experimental settings, 25 cases exhibit a standard deviation less than or equal to 0.0005, indicating that TimePerceiver remains highly stable.

---

[W2] Multiple latent sizes are tested, I suppose the choice is based on performance, but the procedure is not detailed. Studying the sensitivity of the model to this parameter should be included in the study.

[Response] Thank you for mentioning the important point. In our experiments, we simply chose the latent size based on validation. Following your suggestion, we further explore its sensitivity as reported below. We observe that the presence of a latent bottleneck clearly improves performance over having none, but the specific number of latent tokens plays only a minor role in determining the final performance, as they are all relatively small compared to the number of input tokens. This indicates that our model is not overly sensitive to this parameter, and even with a small latent size, we observe substantial performance gains.

---

[W3] Since the method was trained on an extended task able to handle imputation and interpolation, it could have been interesting to evaluate the performance on such tasks.

[Response] Following your suggestion, we conduct imputation experiments with three datasets (ETTh1, Weather, Electricity) and three baselines (TimesNet [1], iTransformer [2], PatchTST [3]). We here set a more challenging task of imputation via patch-wise masking with a patch length of 24, as opposed to the easier setup of masking at individual timesteps. As shown below, our TimePerceiver consistently outperforms the baselines across all datasets. We will add these results with more baselines into the final manuscript.

[1] Wu et al., TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis, ICLR 2023
[2] Liu et al., iTransformer: Inverted Transformers Are Effective for Time Series Forecasting, ICLR 2024
[3] Nie et al., A Time Series is Worth 64 Words: Long-term Forecasting with Transformers, ICLR 2023

---

[W2] Furthermore, the reprojection in the original size is not motivated either.
[Q1] The decoder operates by projecting contextualised query tokens back into full patch outputs. It remains unclear whether bypassing re-expansion to full resolution (i.e., using latent tokens directly) might yield a more efficient architecture. Have the authors tried directly using the encoded latent token as input for the decoder? If so, how much was the performance degrading?

[Response] Thank you for the thoughtful comment. We have conducted experiments with the variant (referred to as DirectLT) you suggested in which the decoder uses latent tokens directly without re-expansion to full resolution. As shown below, our decoder design is significantly better than directly using the encoded latent token as input for the decoder.

We believe the superiority of our design stems from the following insight. In our framework, the latent bottleneck is designed to capture global temporal-channel patterns, while the original input sequence retains fine-grained local signals. Relying solely on the latent makes it difficult to reconstruct these detailed signals, ultimately leading to performance degradation. In contrast, our framework enriches input patch representations by leveraging the latent, enabling each to incorporate both global contexts and local details.

---

[Q2] I am not sure I understand the difference between interpolation and imputation in the description of the project.

[Response] Your understanding is generally correct. Interpolation can indeed be considered a subset of imputation. However, we made the distinction because interpolation refers to filling in values within continuous gaps, while imputation includes both such cases and those where individual points are missing. As you insightfully pointed out, our model does not explicitly separate these cases, but rather handles them within a generalized formulation.

---

[Q3-1] I am not fully convinced that this framework would be able to directly handle arbitrary time segments as is, as this would blur the captured seasonality.

[Response] We gracefully disagree with the concern that our framework may hinder the model’s ability to capture seasonality. On the contrary, we believe that the ability to handle arbitrary time segments enables the model to learn a broader range of recurring seasonal patterns.

Our generalized formulation allows the model to learn from a wider range of configurations such as predicting the present given both past and future, or inferring the past from the future alone. By exposing the model to such diverse temporal contexts, we enhance rather than blur its ability to capture seasonal patterns. Crucially, latent tokens aggregate relationships among patches across every temporal‑channel tokens, creating a compressed global summary that captures seasonal patterns. This perspective is supported by our empirical results, where the generalized formulation consistently outperforms standard formulation across multiple settings.

---

[Q3-2] I think additional information on the size of the patch should be integrated into the representation, perhaps through the same system as temporal and covariate index information. Have the authors tried anything in that regard?

[Response] Regarding your suggestion to provide additional information on the size of the patch, since the current TimePerceiver implementation is trained with a fixed patch length, we did not include patch size embeddings in our framework. As shown below with our experiments on ETTh1, we inject a patch length embedding into both inputs and queries that produced no measurable gains in our ablation study. Therefore such a patch-size embedding has not proved necessary for the current setting.

Instead, based on your idea, we further validate a new multi-resolution design with patch-size embeddings: (i) we patchify the input time-series with various patch sizes 12, 24, 48, and then (ii) use all patches as input tokens in our framework. As shown below, it has not yet delivered a clear performance lift.

However, your insightful suggestion will be strongly considered in the future when developing time-series foundation models, especially when combined with our generalized formulation, since we believe that enabling the model to process time-series at multiple resolutions is critical for scaling it to handle large-scale data.

Dear Reviewer QYcP,

Thank you for your comments and for taking the time to review our manuscript.

For W3, we apologize for the confusion. Our intention was to emphasize that TimePerceiver consistently achieves strong performance across all datasets, not necessarily that it outperforms all baselines on every single one.

In addition, it is worth noting that, even though we did not specifically tailor our framework to the imputation task, we were still able to achieve competitive results as reported in the table.

Please do not hesitate to let us know if there are any remaining concerns or clarifications you would like us to address. We are more than willing to make the most of the remaining discussion time to further improve our paper.

Thank you very much,
Authors

Thanks for the active engagement in the discussion with the authors and filling in the final justification.

Please remember to also submit the mandatory acknowledgement as well.

Thanks,

-Area Chair

I want to thank the authors for the time and effort invested in conducting additional experiments to address the raised concerns. The new results help clarify the strengths and design choices of the proposed method.

While the imputation performance does not surpass all baselines, it remains competitive—particularly given the advantages in training efficiency and memory usage compared to existing methods (as noted in the response to Reviewer rg1y).

I also found the analysis of latent size (extended in the answer to reviewer QYcP) particularly insightful. Though surprising, the observation that performance saturates with as few as 8 latent tokens supports the authors’ intuition in [W2-Q1], namely that the latent tokens primarily capture global channel-level patterns, while the reprojection helps recover local details.

All of my questions and weaknesses have been addressed; provided the additional experiments are included in the appendix for the final version, this paper would make a strong addition to the field of time series forecasting. I raised my score and would recommend this paper for publication.