acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Lightweight Long-Term Time Series Forecasting with Period-Aligned Summarization

Abstract

Long-term time series forecasting (LTSF) requires models that capture temporal dependencies while remaining computationally efficient. We propose PRISM, a lightweight forecasting architecture that exploits periodic structure as a computational prior. PRISM reorganizes the input sequence into a period-aligned representation with cross-period and intra-period axes, and assigns asymmetric computational roles to them. Sequence summarization compresses the cross-period dimension before forecasting, while lightweight temporal mixing refines the intra-period structure. This allows the forecasting head to operate on a compact representation rather than the full historical sequence. Experiments on seven LTSF benchmarks across four forecasting horizons show that PRISM maintains competitive forecasting accuracy with a small computational footprint. Efficiency measurements and structural ablations further confirm that the period-aligned design substantially reduces computational cost while preserving predictive performance. Overall, PRISM provides a simple and effective approach to balancing forecasting accuracy and computational efficiency in LTSF.

Primary area:
learning on time series and dynamical systems
Keywords:
long-term time series forecasting, lightweight forecasting, periodic structure, period-aligned representation, sequence summarization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fqdVOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a lightweight method for long-term time series forecasting. PRISM reshapes a one-dimensional time series into a two-dimensional, period-aligned representation based on a predefined period length. It separately models cross-period patterns and intra-period temporal dynamics using lightweight convolutional modules and a shared forecasting head.

Critical strengths

s1: Balancing forecasting accuracy and computational efficiency is an important objective for long-term time series forecasting.

s2: The period-aligned representation provides an intuitive way to separate cross-period dependencies from intra-period dynamics.

Critical weaknesses

w1: The central idea of PRISM is to reorganize a one-dimensional time series into a period-aligned representation and use lightweight convolutional modules to model patterns along the two dimensions. While intuitive, this design appears to be closer to a combination or structural adaptation of existing ideas in periodic modeling, temporal decomposition, and lightweight forecasting than a clearly differentiated methodological contribution. The paper does not sufficiently establish what fundamentally new modeling capability PRISM introduces.

w2: The paper evaluates PRISM on seven benchmark datasets and compares it with several baselines. However, the authors should provide a more detailed analysis across forecasting horizons, datasets, and different temporal characteristics, and explain the underlying reasons for performance variations.

w3: The period length and cross-period compression mechanism may jointly determine which temporal information is retained. Without sufficiently targeted ablations and analyses, it is difficult to determine whether the reported performance gains arise primarily from period alignment, the convolutional architecture, or other implementation choices.

Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer

Thank you for the constructive review. We appreciate your recognition of the importance of accuracy--efficiency trade-offs and the intuitive separation of temporal structures. We revised the manuscript to clarify the methodological contribution, provide a more detailed interpretation of forecasting results, and explicitly discuss the limitations of our ablation analyses.

W1. Methodological novelty and distinction from prior work. We agree that period-aligned representations, convolution, and temporal compression are established techniques. PRISM's methodological contribution lies in the specific structured mapping that combines sparse, weight-shared cross-period compression with dimension-preserving intra-period mixing. This design explicitly reduces the forecasting-head dimension while preserving the intra-period axis, rather than introducing an entirely new primitive operation.

PRISM applies a shared dilated convolution to compress the cross-period dimension from NN to N′N' before forecasting, while a shared circular convolution mixes intra-period positions without reducing their dimension. This constrains the forecasting head to operate on a compact representation with N′MN'M parameters. Unlike SparseTSF's shared cross-period forecasting, TimeBase's learned temporal bases, and MixLinear's time/frequency pathways, PRISM combines structured cross-period compression with dimension-preserving intra-period mixing.

We revised Sections 1 and 2.2 to distinguish PRISM from these approaches and additional period-aware methods, including TimesNet, PDF, and PhaseFormer. Sections 3.4--3.5 further characterize PRISM as a structured linear predictor with sparse, weight-shared cross-period mapping and circulant intra-period mixing. We therefore position the contribution as a lightweight computational design rather than claiming novelty for period alignment or its individual operations.

W2. Performance variations across datasets and forecasting horizons. We expanded the performance analysis in Section 4.2 using Table 1. PRISM achieves the lowest MSE across all four forecasting horizons on ETTh1 and lower MSE than MixLinear in 22 of 28 dataset--horizon settings. However, TimeBase and FITS consistently outperform PRISM on Weather, while PatchTST performs better on Traffic. PRISM's MSE also increases with the forecasting horizon across all seven datasets.

Section 4.4 provides additional evidence of dataset-dependent behavior. At H=720H=720, removing period reorganization increases ETTh1 MSE from 0.4170.417 to 0.5140.514, whereas the reported MSE remains unchanged on Electricity and Traffic. This suggests that the predictive benefit of period-aligned modeling varies across datasets.

One possible explanation is that the effectiveness of PRISM's structured compression and linear mapping depends on dataset-specific temporal characteristics. Longer forecasting horizons also require predicting more distant observations from the same historical window. However, the current experiments do not isolate the underlying causes of these performance variations, and these interpretations remain hypotheses.

W3. Interaction between period alignment and cross-period compression. We agree that period length, compression strength, and architectural choices jointly influence the information retained for forecasting. We therefore expanded the sensitivity analysis in Appendix D.3 and added Table 8 to report the representation dimensions and parameter counts for each evaluated period length.

For Electricity and Traffic, changing PP from 2424 to 3030 reduces the retained cross-period dimension N′N' from 99 to 33 and the parameter count from 279279 to 8181. Consequently, the performance degradation at P=30P=30 cannot be attributed to period misalignment alone.

Appendix D.1 further examines the accuracy--efficiency trade-off through dilation. On Electricity, increasing dilation from d=1d=1 to d=7d=7 reduces N′N' from 2727 to 99 and MACs from 8.118.11M to 2.702.70M while maintaining relatively stable MSE. More aggressive compression at d=9d=9 increases forecasting error. Table 3 also shows that removing sequence summarization on Traffic increases MACs from 7.267.26M to 21.7221.72M while slightly reducing MSE from 0.4510.451 to 0.4470.447.

These results clarify the trade-off between retained information and computational efficiency. However, changing PP or dilation also changes representation dimensions or receptive fields, and component removal changes model complexity. We explicitly acknowledge in Sections 4.4--4.5 and Appendix D that these analyses do not independently isolate the causal contributions of period alignment, convolution, and compression.

We hope these clarifications address the main concerns while making the scope and limitations of our contribution more explicit.

Thank you for your detailed response and clarifications. While I appreciate the more explicit positioning of PRISM as a lightweight computational design, my main concern regarding its methodological novelty remains unresolved. Furthermore, the additional analyses do not sufficiently explain the performance variations across datasets or disentangle the contributions of period alignment, convolution, and compression, as the authors also acknowledge. Therefore, I will maintain my original rating.

Reviewer_QYtXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PRISM, a lightweight architecture for long‑term time series forecasting using period‑aligned summarization. It reorganizes input sequences into a 2D period‑aligned representation and applies an asymmetric dual‑axis computation paradigm: shared dilated convolutions compress the cross‑period dimension to reduce historical redundancy, while circular‑padded weight‑shared convolutions capture intra‑period patterns without modifying period length. A lightweight linear head produces forecasts. PRISM uses channel‑independent processing, so its parameter count does not scale with input variables.

The work leverages periodicity as a computational prior to cut complexity for long‑horizon forecasting. Evaluated on seven LTSF benchmarks against lightweight and deep baselines, PRISM is assessed by both forecasting accuracy and practical efficiency metrics (MACs, memory, training/inference latency). Experiments show competitive prediction performance with low computational cost, supported by ablations for each core component.

Overall, this is a simple, promising efficient forecasting architecture. However, the novelty of its asymmetric dual‑axis design versus existing period‑aware lightweight methods is not sufficiently established. Its dependence on a predefined period length and associated divisibility constraints also raise concerns about general robustness and applicability. More systematic comparisons and broader empirical validation are needed to quantify PRISM’s practical advantages.

Critical strengths
  1. The period‑aligned 2D representation cleanly separates cross‑period dependencies and intra‑period temporal patterns. Critically, the model assigns distinct computational roles to the two axes: the cross‑period dimension is explicitly compressed for redundancy reduction, while intra‑period patterns are refined without dimension reduction. This decomposition offers an intuitive inductive bias for long‑term forecasting.

  2. PRISM leverages weight‑shared convolutions, cross‑period summarization, and a channel‑independent scheme, where parameters do not scale with variable numbers. Beyond parameter count and MACs, the paper reports comprehensive hardware‑oriented metrics, including peak memory, training time, and GPU/CPU inference latency. It demonstrates a small computational footprint, showing promise for resource‑constrained forecasting scenarios.

  3. Experiments are conducted on seven LTSF datasets with multiple forecasting horizons. Quantitatively, PRISM obtains lower MSE than MixLinear in 22 out of 28 dataset‑horizon settings and outperforms TimeBase in 19 settings (with additional ties). It achieves strong results without resorting to over‑parameterized backbones.

  4. Ablations validate the necessity of each core component. Removing period reorganization drastically increases computation cost, verifying its central role for efficiency. Results also identify cross‑period summarization as the primary source of computational savings. Supplementary sensitivity analyses for period length, dilation, and kernel size further enrich understanding of model behavior.

  5. The overall architecture is conceptually simple and avoids complex pipelines such as frequency‑domain transforms or basis learning. It provides an interesting alternative design for lightweight long‑term time‑series forecasting.

Critical weaknesses
  1. Exploiting periodicity for redundancy reduction is a known direction, and individual components (period reshaping, shared convolution, temporal compression) are not new. The paper fails to clearly articulate mechanistic differences compared with SparseTSF, TimeBase and MixLinear. Direct mechanism‑level comparisons are needed to establish what makes PRISM fundamentally distinct.

  2. PRISM requires a hand‑set period PP and enforces that input length LL and forecast horizon HH must be divisible by PP. Real‑world data can have multiple time‑varying absent periodicities. Current sensitivity tests cover only a small number of datasets, and it remains unclear how performance degrades under misspecified PP, with no evaluation on weakly periodic time series. Furthermore, circular padding for intra‑period mixing implicitly assumes continuity between period boundaries, which may introduce spurious dependencies for signals with sharp boundary discontinuities, yet this risk lacks systematic analysis.

  3. PRISM uses fully channel‑independent processing and ignores cross‑channel interactions. While this improves parameter scalability, the paper does not analyze the performance cost from discarding cross‑variable predictive signals. It is unclear under which real‑world scenarios this assumption severely degrades forecasting quality.

  4. Full hardware‑oriented efficiency benchmarks are only reported for one single configuration (Electricity, L=H=720L=H=720), so consistent efficiency gains across datasets input windows hardware are not proven. In addition, empirical claims would be strengthened by more thorough variance analysis or significance tests to validate whether observed accuracy gaps are statistically meaningful.

  5. Ablations confirm cross‑period summarization yields major computational savings. Nevertheless, the paper does not systematically investigate how compression strength balances computational cost and forecasting error. Such analysis is essential to understand model behavior and characterize suitable application scenarios for PRISM.

Questions for authors
  1. What is the key methodological difference between PRISM and existing period‑aware lightweight forecasting methods, particularly SparseTSF, TimeBase, and MixLinear? Can the authors identify a specific modeling capability or mechanism that cannot be readily obtained from these methods?
  2. How sensitive is PRISM to the choice of period length? Can the authors provide experiments with incorrect period lengths, multiple seasonalities, or nonstationary periodic patterns? Is automatic period selection feasible within the proposed framework?
  3. How does PRISM perform when cross‑channel interactions are essential? Would incorporating a lightweight cross‑variable interaction mechanism improve forecasting accuracy without substantially compromising efficiency?
  4. Can the authors provide parameter counts, FLOPs or MACs, memory usage, training time, and inference latency across additional datasets and input/output lengths? Can the efficiency claims be validated under matched hardware and evaluation settings?
  5. Are the reported improvements consistent across multiple random seeds? Could the authors provide standard deviations or confidence intervals, especially for settings where the differences between competing methods are small?
  6. How do the compression ratio and the number of retained cross‑period representations affect forecasting accuracy and computational cost? Such an analysis could help explain the role of cross‑period summarization beyond the aggregate ablation results.
2 comments
AuthorsOpenReview ↗

Author Response to Reviewer

We revised the manuscript to clarify PRISM's distinction from prior work, its limitations, and the empirical evidence.

W1. Methodological novelty. We agree that period reshaping, convolution, and temporal compression are not individually new. Sections 1 and 2.2 clarify PRISM's asymmetric combination: a shared dilated convolution compresses the cross-period axis before forecasting, while circular mixing preserves the intra-period length. Unlike SparseTSF's shared cross-period prediction, TimeBase's learned temporal bases, or MixLinear's time/frequency pathways, PRISM uses this specific structured compression-and-mixing arrangement. We claim novelty in the architecture, not its components.

W2. Period assumptions and circular padding. Section 4.5 and Appendix D.3 extend period sensitivity to all seven datasets. Electricity and Traffic degrade at P=30P=30 relative to P=24P=24, but N′N' also drops from 99 to 33; period mismatch and compression are therefore confounded. The current implementation requires PP to divide both LL and HH. We have not tested weak, multiple, or time-varying periodicities or sharp boundary discontinuities. Circular padding connects positions at opposite ends of a period and may be unsuitable for such discontinuities.

W3. Channel independence. Section 3.2 now explains that each target uses only its own history, with parameters shared across variables. This yields channel-independent parameter counts but cannot exploit cross-variable predictive signals. We have not quantified the cost on strongly coupled data or evaluated a cross-channel variant.

W4. Efficiency and variability. Appendix D.4 adds parameter counts, MACs, peak GPU memory, epoch time, and GPU/CPU latency for the other six datasets at L=H=720L=H=720, under the same experimental environment. Other sequence lengths and hardware platforms were not evaluated. Appendix C already reports three-run means and standard deviations for PRISM and all baselines; we do not claim formal statistical significance.

W5. Compression versus accuracy. We expanded the interpretation of the existing dilation study in Appendix D.1 by explicitly relating the retained cross-period dimension N′N' to MACs and forecasting error. On Electricity at H=720H=720, increasing dilation from d=1d=1 to d=7d=7 reduces N′N' from 2727 to 99 and MACs from 8.118.11M to 2.702.70M, while MSE remains relatively stable. More aggressive compression at d=9d=9 reduces N′N' to 33 but increases forecasting error. Since dilation also changes the receptive field, these results characterize the observed accuracy--efficiency trade-off without isolating compression strength independently.

Q1. Mechanistic distinction. PRISM constrains computation before prediction through sparse, shared cross-period summarization and separately mixes intra-period positions without shortening that axis. This differs from the mechanisms described for SparseTSF, TimeBase, and MixLinear. We do not claim an exclusive capability that could not be implemented in other frameworks.

Q2. Misspecified periods and selection. Figures 2 and 5 evaluate alternative PP values; some settings, notably Electricity and Traffic at P=30P=30, perform substantially worse. Table 8 shows that PP also changes representation size and parameters. Validation-based selection of PP is technically possible but entails extra tuning; the evaluated design intentionally uses a fixed dataset-level prior. Multiple seasonalities and changing periods remain untested.

Q3. Cross-variable interactions. PRISM cannot condition a target forecast on other channels (Section 3.2). A lightweight cross-channel module could be considered, but we have neither evaluated its accuracy benefit nor measured its overhead.

Q4. Scope of hardware evaluation. Table 2 and Appendix D.4 cover all seven datasets with matched input/output lengths and a common evaluation environment, reporting all requested efficiency metrics. Appendix A.5 specifies the timing and memory protocol. Results across other lengths or hardware remain unavailable.

Q5. Random seeds and uncertainty. Tables 6--7 in Appendix C report the mean and standard deviation from three independent runs for PRISM and all baselines. PRISM's reported MSE standard deviations are at most 0.0080.008. For example, on ETTm1 at H=720H=720, PRISM obtains 0.416±0.0010.416\pm0.001 MSE, compared with 0.415±0.0020.415\pm0.002 for TimeBase. We do not interpret this small difference as evidence of a meaningful performance advantage. The three-run results characterize observed variability but provide limited statistical power, and we do not claim formal statistical significance.

Q6. Number of retained representations. For Electricity, N=30N=30, k=4k=4, and s=1s=1 give N′=30−3dN'=30-3d. Figure 3 shows N′=27,9,3N'=27,9,3 for d=1,7,9d=1,7,9, respectively. Fewer retained features reduce MACs, but excessive compression increases error. This result also reflects changing dilation/receptive field, as acknowledged in Appendix D.1.

The author has responded to all questions. I consider the explanations on experimental transparency and limitations to be well addressed. However, concerns regarding the methodological innovation remain insufficiently resolved. I will therefore keep my original score.

Reviewer_EKJCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PRISM, a lightweight channel-independent model for long-term time series forecasting. The input window is mean-centered and reshaped into an N×P matrix using a fixed period length P. A shared convolution compresses the cross-period axis. A circular convolution then mixes positions within each period. A linear head forecasts the future periods from this compact grid. The model has only k + t + N′M parameters, independent of the channel count. Experiments use public benchmark datasets and compare PRISM with many recent forecasting models. The paper also reports efficiency, ablation and sensitivity analyses. And I downloaded the anonymous code provided by authors and checked them.

Critical strengths

S1. The design goal is clear and the model is indeed lightweight. The paper aims at an ultra-lightweight forecaster, and the resulting parameter count is very low. It does not grow with the number of channels. The design is systematic: period reorganization, cross-period summarization, intra-period mixing and a shared linear head each have a clear role. One section specifies every shape, and the exact parameter and MAC counts are given in the paper.

S2. The experiments support the lightweight claim. The efficiency study goes beyond parameter counts. It reports MACs, peak memory, training time and inference latency, and PRISM compares favorably with recent lightweight baselines on these measures. The ablations show where the savings come from. Accuracy stays competitive while the model remains very small.

Critical weaknesses

W1. The novelty is limited, and important related work is not cited. Each component of PRISM has a close precedent in prior work. Period reorganization, mean-centering and the shared period-level head follow SparseTSF (ICML 2024). Compressing the cross-period axis before the head is the core idea of TimeBase (ICML 2025). Period-aligned 2D modeling also appears in TimesNet (ICLR 2023), PDF (ICLR 2024) and PhaseFormer (ICLR 2026).

These closely related works are not cited or discussed: PhaseFormer(ICLR 2026)., CycleNet (NeurIPS 2024) and PDF(ICLR 2024) . To be honest, I am quite surprised by how few works the paper cites. The authors should spend more time searching and discussing related work, and cite enough of related literatures correctly.

W2. The linear design is not explained from a representation-learning perspective. Every operator in PRISM is linear, so the whole model is a structured linear regressor. The paper does not say what this design teaches the community about representation learning. It also does not explain why such a linear model keeps enough information for long-term forecasting. At least a more detailed analysis is needed, for example of the learned weights or of the low-rank structure the model imposes. This gap also shows in the experiments and presentation, which have much room for improvement. Accuracy gains are concentrated on ETTh1, efficiency is measured in a single configuration, and the ablations and sensitivity analyses are confounded. I downloaded the anonymous code provided by authors and checked them, so I think maybe there still have room to improve the design of exps and representing, also with anonymous codes.

Questions for authors
  1. Both PRISM and TimeBase (ICML 2025) compress the cross-period axis before the forecasting head. It would be helpful if the authors could clarify the functional difference between the two, perhaps with a comparison against TimeBase at a matched rank (R = N′).

  2. I would appreciate a discussion of how PRISM relates to PhaseFormer (ICLR 2026), CycleNet (NeurIPS 2024), LightTS (arXiv 2022) and PDF (ICLR 2024). Adding the most relevant ones as baselines would further strengthen the paper.

  3. PRISM is a linear model. It would be valuable to explain why this design keeps enough information for long-term forecasting, for example through an analysis of the learned weights or of the low-rank structure the model imposes.

  4. I would also be interested in the authors' view on what this linear design suggests to the community from a representation-learning perspective, beyond parameter savings.

  5. Since the model is linear, a comparison with closed-form least squares and with a reduced-rank linear model on the same period grid would be informative.

  6. It would help readers if parameters, MACs, memory and latency were reported for every dataset, including ETTm1 and ETTm2. Releasing the baseline and ablation code would also make the comparisons easier to reproduce.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer

Thank you for the careful review. We revised the manuscript to sharpen the distinction from prior work, explain the linear architecture and its limitations, and expand the efficiency evaluation. We address each point below.

W1. Novelty and positioning. We agree that period alignment, convolution, and linear forecasting are not new individually. PRISM's contribution is their asymmetric computational organization: a shared dilated convolution compresses the cross-period axis before prediction, while a shared circular convolution mixes positions within each period without shortening that axis. We revised Sections 1 and 2.2 to clarify this distinction rather than claim novelty for the individual components.

W2. Dataset dependence and confounded analyses. We agree that PRISM is a structured linear predictor. To clarify its representation-learning implications, Section 3.4 now derives the mapping Y~2D=WSX2DT\widetilde{Y}_{2D} = WSX_{2D}T with rank⁡(WS)≤min⁡(M,N′)\operatorname{rank}(WS) \leq \min(M,N'). The sparse, weight-shared operator SS and circulant operator TT impose structural constraints that reduce complexity but also limit expressiveness. Compression may discard information, and we do not claim guaranteed information preservation. We also expanded the dataset- and horizon-dependent performance analysis in Section 4.2 and efficiency measurements to all seven datasets in Appendix D.4. Section 4.4 acknowledges that ablations change model dimensions and parameter counts. Furthermore, Appendix D.3 and Table 8 show that changing PP also changes compression strength; for Electricity and Traffic, P=24P=24 to P=30P=30 reduces N′N' from 99 to 33 and parameters from 279279 to 8181. Thus, these experiments characterize observed trade-offs without independently isolating their causal contributions.

Q1. Difference from TimeBase. Both models reduce historical information before forecasting, but their reduction mechanisms differ. TimeBase extracts learned temporal bases, whereas PRISM uses a sparse, weight-shared dilated convolution to compress the cross-period dimension, followed by a separate circular convolution for intra-period mixing (Section 2.2). We have not conducted a matched-rank comparison at R=N′R=N', so we claim a structural distinction rather than superiority under matched conditions.

Q2. Coverage of relevant prior work. We expanded Section 2.2 to discuss LightTS, CycleNet, PDF, and PhaseFormer, in addition to SparseTSF, TimeBase, and MixLinear. The revision acknowledges that period-aware representations and lightweight forecasting precede PRISM and specifies how our arrangement of compression and mixing differs. We do not claim empirical superiority over methods without direct comparisons.

Q3. Information preservation in linear compression. Section 3.4 now expresses the centered forecast as Y~2D=WSX2DT\widetilde{Y}_{2D}=WSX_{2D}T, where SS represents cross-period convolution and TT circular mixing, with rank⁡(WS)≤min⁡(M,N′)\operatorname{rank}(WS) \leq \min(M,N'). Compression may discard input information. End-to-end training encourages the compressed features to retain information useful for forecasting, but does not guarantee information preservation or reconstruction.

Q4. Representation-learning interpretation. We now describe PRISM as a constrained linear predictor rather than a nonlinear feature learner. Weight sharing and sparse connectivity in SS, together with the circulant structure of TT, provide structural inductive biases with limited expressiveness. When the temporal kernel size is one, the operation is only a shared scalar and does not mix intra-period positions. These limitations are now explicit.

Q5. Comparison with OLS/reduced-rank regression. Section 3.5 now contrasts PRISM with unrestricted least-squares mappings and reduced-rank regression. PRISM additionally constrains the cross-period factor by convolutional weight sharing and the intra-period operator by circular structure. Its parameter count is k+t+N′Mk+t+N'M, independent of the number of channels. We have not added OLS or reduced-rank baseline experiments and do not claim predictive superiority over those models.

Q6. Efficiency and reproducibility. Beyond Electricity in Section 4.3, Appendix D.4 reports parameters, MACs, peak GPU memory, epoch time, and GPU/CPU inference latency on the remaining six datasets at L=H=720L=H=720. On Electricity, PRISM uses 2.702.70M MACs and 487.48487.48 MB of peak memory versus TimeBase's 3.053.05M MACs and 982.51982.51 MB. The extended results also show that PRISM is not best on every metric or dataset. Appendix A.5 describes the measurement protocol, and the Reproducibility Statement provides an anonymized implementation and experiment-script link.

We hope these revisions clarify both the intended contribution and the limits of the current evidence.

I appreciate the authors' thorough response. While these additions help clarify certain aspects of the paper, they do not fully alleviate my fundamental concern regarding the marginal technical novelty. Considering the overall significance and contribution to the community, I feel the paper is not yet ready for publication at ICLR. Hence, I will maintain my initial score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.