- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2(+1)
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes SPLICE, a modular framework for long-horizon time series imputation. It combines a JEPA-based predictive encoder, a conditional latent bridge (supporting deterministic, noise-perturbed, or generative variants), an hourly-conditioned decoder, and adaptive conformal inference (ACI) for distribution-free uncertainty calibration. Experiments across multiple datasets demonstrate that SPLICE achieves lower errors.
Critical strengths
- The integration of JEPA as a predictive latent backbone for long-gap time-series inpainting appears novel to me.
- Overall, the experiments is solid, including those presented in the appendix.
- The code is provided in the Supplementary Material, which facilitates reproducibility.
Critical weaknesses
- The manuscript is difficult to read and understand from a human reader’s perspective due to its writing and organization. First, neither the Introduction nor the overall narrative of the paper clearly explains why JEPA is suitable for the problem at hand or what it can address that previous generative machine learning approaches cannot. The same concern applies to the other components of the proposed framework. Some extracted tables/figures in the text appear corrupted; while likely an artifact of PDF parsing, it occasionally interrupts readability. Is Figure 1 AI-generated? The font is so small that it is illegible. Even worse, the figure contains excessive whitespace around it, making the layout visually unappealing. Is it appropriate to devote so much space in the main text to describing the setup while relegating many important experiments to the appendix? This seems to be a misplaced prioritization. Furthermore, the experiments, particularly those in the appendix, lack sufficient discussion of key findings and analysis of observed phenomena. ... I do not believe the manuscript, in its current form, meets the acceptance threshold for ICLR, at least for its presentation.
- The proposed method also lacks novelty and reads more like a simple stacking of existing machine learning techniques than a genuinely new contribution. As a point of comparison, I am curious why the related work only covers studies published before 2025, as this makes it difficult to assess the method's distinctiveness relative to existing approaches.
- While the proprietary datasets are operationally meaningful, heavy reliance on them makes independent verification difficult; only four public datasets are included in the main table. Statistical tests indicate differences vs strong baselines (SAITS/BRITS) are not significant at p<0.05; more seeds or broader public benchmarks could strengthen claims of superiority.
- The imputation comparison omits some recent strong generative imputers, especially for algorithms after 2025.
- How maps into updated band quantiles from ensemble samples is not fully specified; more algorithmic clarity would help reproducibility and theoretical alignment. The coverage guarantee is marginal/long-run; no conditional or within-window guarantees are offered, which may limit operational interpretations for fixed gaps. The inference time claim (“6.5–9.0 ms for a 91-day gap”) appears implausibly low given sampling and decoding; this may be a units typo (likely seconds) and warrants correction.
Questions for authors
See Critical Weaknesses.
4 comments
Response to Reviewer C931 (2/2)
W3. Proprietary data, four public datasets, p > 0.05
Section 6.2 and Appendix P state that SPLICE is not significantly better than SAITS, BRITS or PatchTST. A one-sided Wilcoxon test over 12 datasets gives p = 0.102, 0.102 and 0.117, respectively. SPLICE is significantly better than ImputeFormer (p = 0.021) and than CSDI, TimesNet and TimeMixer++ (p < 0.001).
Public data. Four of the 13 datasets in Table 2 (Table 1 in the submitted version) are public: three UCI Electricity feeds and ETTh1. Section 6.5 also replicates the three controlled swaps on twelve catchments from the public CAMELS-SPAT dataset (Knoben et al., 2025). The target is hourly discharge, the covariates are ERA5 weather variables, and the pipeline is unchanged. The same orderings hold (Table 3):
- JEPA beats TS2Vec on 9 of 12 catchments and the VAE latent on 10 of 12;
- 5-step flow matching beats 50-step DDIM on 8 of 12.
Both encoder comparisons keep their ordering under hydrological criteria: six of ten criteria, including KGE and NSE, favor JEPA (Appendix F). The catchments were chosen before any training, on data completeness and hydrological diversity only (Table 7). CAMELS-SPAT is now listed alongside the other public datasets in the Data Availability appendix (Appendix H) and in the reproducibility statement.
W4. No imputers after 2025
We added TimeMixer++ (Wang et al., ICLR 2025), a multiscale and multiresolution architecture. It is trained through PyPOTS with the same splits, gap masks, three seeds and training protocol as the revised TimesNet column.
- Mean result: its mean Load-only MSE over the twelve datasets is 0.298, against 0.056 for SPLICE.
- Per dataset: SPLICE is lower on 11 of 12 datasets (one-sided Wilcoxon, p < 0.001).
- Exception: RIInd1014, where TimeMixer++ reaches 0.149 against 0.202 for SPLICE; Seasonal Naïve is best there.
Overall win counts do not change: SPLICE is best on 9 of 12 datasets. TimeMixer++ is now in Table 2, the MAPE table (Table 15), Appendix P and the baseline figure (Figure 7).
W5. From α_t to the bands; marginal guarantee; 6.5–9.0 ms
At step t, the conformal quantile q̂_t is the empirical quantile at level ⌈(n+1)(1−α_t)⌉/n of the n pooled pointwise calibration scores. The band is [q_0.025 − q̂_t, q_0.975 + q̂_t], where q_0.025 and q_0.975 are the quantiles of the generated ensemble. Section 4.3 now states this. As noted in the limitations (Section 8), the guarantee is marginal and holds in the long run.
We now measure timing per 91-day gap using a script included in the supplementary material (RTX A4000, median over the nine datasets):
- Deterministic pipeline: 2.8 ms.
- Single trajectory: 32 ms for FM-A (5 steps), 0.36 s for DDIM (50 steps).
- Ensembles for the intervals: 20 DDIM samples drawn in one batch take 1.1 s per gap, and the 50 calibration samples take 2.5 s. With FM-A, the same two figures are 0.11 s and 0.26 s.
Table 13 (Table 12 in the submitted version) now reports these values, listing single trajectory and ensemble separately.
We hope these answers address your concerns, and we are happy to clarify any remaining point during the discussion.
Thank you for the detailed response. I have updated my score to reflect the clarifications. However, I believe some problems, particularly regarding the writing, remain unresolved, and I therefore still lean negative.
Response to the follow-up
Thank you for considering our response and for updating your score. Could you please point us to the specific sections or passages that you still find difficult to follow? In the revised PDF, we have added the motivation for the predictive latent space in Section 4.1, redrawn Figure 1, and moved the encoder comparison to Section 6.1. We would be happy to address any specific passages you identify in a new revision of the PDF before the end of the discussion period, time permitting, and otherwise in the final version.
Response to Reviewer C931 (1/2)
Thank you for your review, and for finding the use of JEPA as a predictive latent backbone for long-gap inpainting novel, the experiments solid, including those in the appendix, and the provided code helpful for reproducibility. We answer each point below; all changes to the PDF are listed in our general comment.
W1. Readability, why JEPA, and Figure 1
Why JEPA. Filling a 91-day gap means inferring unobserved days from the days around them, and JEPA is trained on this exact task. A predictor infers the embeddings of masked days from the observed context, with the loss computed directly in latent space (Section 4.1, Eq. 1). The latent therefore retains what is predictable from context: load regime, daily and weekly periodicity, and response to weather. Hourly detail is restored from covariates by the hourly-conditioned decoder (Section 4.4). Masked reconstruction (PatchTST) optimizes input fidelity, and contrastive learning (TS2Vec) optimizes invariance between views. Neither objective targets the prediction of missing segments.
We test this choice against both alternatives, with the same bridge architecture:
Section 4.1 now opens with this rationale, and Section 8 states the resulting design principle.
JEPA also gives the pipeline the structure of a world model (Section 1). The bridge rolls out predicted latent states instead of sampling observations directly. This is where operational or physical constraints could be imposed, which we leave for future work.
Presentation. Figure 1 is now redrawn as a vector graphic at its printed size, which increases its text from about 3 pt on the page to 7–8 pt. The empty areas are removed, and the figure shows the 64-dimensional interface shared by the modules. The three-seed encoder swap has moved from the appendix into Section 6.1 (new Table 1).
W2. Novelty, and related work before 2025
We isolate, through controlled module swaps, which component drives reconstruction accuracy in long-gap inpainting, and build an operational pipeline on that finding. We swap each module individually while keeping the rest of the pipeline fixed:
These results give the design rule of Section 8: learn the predictive state space first, then use the cheapest sampler that gives the required diversity. Our setting is contiguous 91-day outages, not randomly missing values. The pipeline provides online conformal intervals whose coverage holds under the deployable protocol: 0.949 on gaps never used before scoring (see our reply to 6FkT, W3). To our knowledge, no previous imputer combines a deep generative model with ACI (Table 10).
Recent work. Section 2 now covers work from 2025–2026:
TS-JEPA (Ennadir et al., 2025) was already discussed in the submitted Section 2.