acceptodds
Under review as a conference paper at ICLR 2027

History-Guided Slice Self-Attention for Mitigating Forecast Smoothing in Irregular Time Series

Abstract

Irregular multivariate time series (IMTS) often provide sparse, asynchronous observations, leaving fine-scale future dynamics weakly constrained. Although diffusion models can generate fine-scale temporal structure, mixing across distinct dynamics in existing IMTS forecasting architectures can suppress temporal variation, leading to forecast smoothing with high-frequency attenuation. Grouping based on noisy query states can amplify such dynamics mixing, motivating history-guided grouping that maintains distinct slice representations and preserves structured variation. To address this, we propose Diffusion Coordinate Transformer with history-guided Slice self-attention (DiCTS) for joint probabilistic IMTS forecasting. Its denoiser employs History-guided Slice Self-Attention (HiSSA), which derives query-to-slice assignments from encoded history and query coordinates rather than noisy query states, while aggregating sample-specific features from the current denoising states. This history-guided assignment prevents injected diffusion noise from influencing the grouping and distinguishes queries with distinct dynamics, preventing attenuation of structured variation. We demonstrate DiCTS with established IMTS benchmarks, indicating competitive probabilistic forecasting performance and point-prediction accuracy. Qualitative and spectral analyses further show that DiCTS mitigates forecast smoothing and more faithfully follows future trajectories.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series, Probabilistic Forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aE3uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses a meaningful problem and achieves promising results. However, its methodological novelty is incremental, and the experiments do not sufficiently validate the central claim that history-guided routing mitigates forecast smoothing. More targeted ablations and experimental analyses are needed.

Critical strengths

The proposed history-guided slice assignment provides a reasonable approach to reducing noise-driven feature mixing during diffusion denoising. The method achieves competitive forecasting results across five benchmark datasets. The additional spectral analyses provide useful evidence regarding temporal variation preservation.

Critical weaknesses
  1. Limited methodological novelty. HiSSA largely follows existing slice-based attention mechanisms. Its main contribution is history-guided slice assignment. The authors should clarify its differences and advantages over Transolver and Transolver++.
  2. Insufficient validation of the core motivation. The paper claims that noisy-state-based grouping causes dynamics mixing and forecast smoothing, but lacks direct evidence. A controlled comparison between history-guided and noisy-state-guided slice assignments is necessary.
Questions for authors
  1. Insufficient ablation and sensitivity analyses. The current ablations do not directly demonstrate HiSSA's contribution to frequency preservation. Additional frequency-domain comparisons, routing strategy ablations, and sensitivity analyses of the number of slices are recommended.
  2. Limited representativeness of visualization results. Figure 1 shows that DiCTS captures a sharp change at the final prediction timestamp, which appears potentially anomalous, while other methods fail to capture it. Although the appendix provides additional visualizations, more representative examples and clarification of the selection criteria would strengthen the conclusions.
  3. Missing computational efficiency analysis. The paper lacks comparisons of training time, inference time, and computational overhead, making it difficult to assess the practical efficiency of DiCTS.
  4. Inconsistencies in experimental results. The VS values for USHCN and MIMIC-III differ between Tables 1 and 3.
  5. Presentation issues. The abstract should more clearly emphasize forecast smoothing as the primary motivation. The Introduction should better explain Figure 1. In Figure 2, the output following the Head module is unclear, and the distinction between DiCT and DiCTS should be clarified. The conclusion would also benefit from a brief discussion of limitations and future work.
7 comments
AuthorsOpenReview ↗

Author Response 1: Summary

We thank the reviewer for the careful and constructive assessment. We provide additional experiments and clarifications addressing the following concerns:

  1. Novelty over Transolver and Transolver++: Our contribution lies in the routing source rather than in slice-based aggregation. HiSSA derives assignments from query coordinates and observed history instead of the noisy states, which makes the grouping independent of the injected noise and reusable across denoising steps and samples.
  2. Validation of the core hypothesis: A controlled routing ablation shows that noise-dependent routing is worse than standard self-attention in CRPS on all five datasets, whereas history-guided routing lowers CRPS on all five (by 5.9–26.5% relative to noise-dependent routing). A counterfactual noise intervention shows that resampling the noise alone regroups queries under noise-dependent routing (10.4–67.9% of a permutation reference), while history-guided assignments are unchanged.
  3. Frequency preservation and sensitivity: Frequency-domain comparisons of the ablation variants show closer agreement with the observed band amplitudes under history-guided routing (lower energy score than noise-dependent routing in 12 of 15 dataset–band cells; lower absolute log-amplitude error than standard self-attention in all 15), with the exceptions reported explicitly. The routing comparison holds for every tested number of slices (G = 8, 16, 32, 64).
  4. Visualization and efficiency: We state how the displayed channels and trajectories were selected, add a further example, and base representativeness on aggregate statistics over the full test set. We also report training time, sampling latency, FLOPs, and parameter counts.

We have also corrected the VS entries in Table 1 and revised the abstract, figure descriptions, notation, and limitations. Detailed results and discussions follow below.

AuthorsOpenReview ↗

Author Response 2: Detailed Response

W1. Limited Methodological Novelty

We thank the reviewer for the opportunity to clarify our contribution. Our contribution is noise-decoupled routing: identifying that the routing source of slice attention becomes a failure point inside a diffusion denoiser, and separating the organization of query interactions from the stochastic states being denoised.

Why Transolver-style routing is problematic in a denoiser. Transolver [1] and Transolver++ [2] compute slice weights from the same point features that are aggregated. In PDE solving these features are clean, and feature similarity reflects similar physical states. In a denoiser they are noisy query states: similarity is partly determined by the injected noise, so queries with distinct temporal dynamics can be grouped together, and because the same assignments govern aggregation and redistribution, mismatched groupings propagate in both directions.

Differences. HiSSA derives assignments only from query coordinate embeddings EqE^q and the history context CqC^q (Eqs. 4–5), obtained by a dedicated routing cross-attention to the history memory. History thus determines which queries exchange information, and the denoising states determine what is exchanged. Transolver++ adds a per-point adaptive temperature and a Gumbel-softmax reparameterization, which sharpen the assignments but do not change their source, so the noise dependence remains.

Transolver [1] / Transolver++ [2] in a denoiserHiSSA (Ours)
Routing inputFeatures being aggregated (noisy query states)Query coordinates and observed history
Access to history when groupingOnly indirectly, through the noisy statesDirect, via routing cross-attention (Eq. 4)
Dependence on injected noise and stepYesNone (by construction)
Assignments during samplingRecomputed for every step and sampleComputed once, reused across steps and samples

Advantages (same backbone and content path; only the routing differs).

  • Forecasting accuracy. History-guided routing reduces mean CRPS on all five datasets (by 5.9–26.5%) and mean VS on four of five (W2 & Q1 (1)).
  • Slice attention alone is not sufficient. Noise-dependent routing has worse mean CRPS than standard self-attention on all five datasets, so directly transplanting Transolver-style slicing into the denoiser degrades performance; the gain of HiSSA comes from the routing source rather than from slice-based aggregation itself.
  • Stable grouping under noise. Resampling the noise alone changes the grouping of noise-dependent routing by 10.4–67.9% of a random-permutation reference, whereas history-guided assignments are unchanged (W2 & Q1 (2)).
  • Frequency preservation. Relative to noise-dependent routing, history-guided routing lowers the band-wise energy score in 12 of 15 dataset–band cells (by 0.7–5.4%), including the high band on four of five datasets, and the absolute log-amplitude error in 12 of 15. Noise-dependent routing raises band amplitudes relative to standard self-attention (14 of 15 cells) without a consistent gain in agreement with the observed amplitudes (7 of 15), whereas history-guided routing improves agreement in all 15 (W2 & Q1 (3)).
  • Lower sampling cost than noise-dependent routing. Assignments are computed once per block and reused across denoising steps and forecast samples, which reduces counted sampling FLOPs by 7.00–10.75%, with 0.25–0.33M additional router parameters. The cost relative to standard self-attention is reported in Q3.

We have revised the manuscript to state these differences explicitly.

[1] Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for PDEs on general geometries. In ICML, volume 235, pp. 53681–53705, 2024. [2] Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Transolver++: An accurate neural solver for PDEs on million-scale geometries. In ICML, volume 267, pp. 41432–41449, 2025.

AuthorsOpenReview ↗

Author Response 3: Detailed Response

W2 & Q1. Insufficient Validation of the Core Hypothesis and Ablation/Sensitivity Analyses

We thank the reviewer for pointing out that our ablation lacked the noise-dependent routing control needed to test the core hypothesis directly, since we compared HiSSA only with standard self-attention. We have added (1) a controlled comparison of routing strategies, (2) a counterfactual noise intervention, (3) a frequency-domain comparison of the ablation variants, and (4) a sensitivity analysis of the number of slices.

1. Controlled comparison of routing strategies. We add a noise-dependent routing variant, which keeps the slice–attend–deslice block (Eqs. 6–8) but derives assignments from the noisy denoising states rather than from history and coordinates. Aggregation, slice attention, deslicing, and all training settings are identical; history-guided routing additionally contains the routing cross-attention of Eq. 4 (0.25–0.33M parameters), which acts only through the assignments. The variants are named as in Table 3 of the revised manuscript:

  • DiCTS (Standard Self-Attn): (formerly "Ours w/o HiSSA"): standard query self-attention, without slice-based routing.
  • DiCTS (Noise-dependent routing): slice-based attention with assignments derived from the noisy denoising states.
  • DiCTS (Ours): slice-based attention with assignments derived exclusively from observed history and target coordinates.

Table. Ablation on five IMTS benchmarks (mean ± sample standard deviation over five seeds; lower is better).

ModelMetricHumanActivityPhysioNet'12USHCNMIMIC-IIIMIMIC-IV
DiTCRPS0.1133 ± 0.00150.2778 ± 0.00140.1657 ± 0.00540.2972 ± 0.00220.2399 ± 0.0013
VS0.0358 ± 0.00050.0910 ± 0.00030.0999 ± 0.00980.1232 ± 0.00080.1193 ± 0.0015
DiCTS (Standard Self-Attn)CRPS0.0981 ± 0.00250.2632 ± 0.00250.1617 ± 0.00090.2767 ± 0.00340.1939 ± 0.0010
VS0.0325 ± 0.00060.0867 ± 0.00020.1031 ± 0.00950.1153 ± 0.00200.0972 ± 0.0030
DiCTS (Noise-dependent routing)CRPS0.1128 ± 0.00230.2784 ± 0.00140.1678 ± 0.00480.2993 ± 0.00150.2444 ± 0.0018
VS0.0357 ± 0.00060.0921 ± 0.00020.0938 ± 0.01070.1320 ± 0.00170.1352 ± 0.0021
DiCTS (Ours)CRPS0.0829 ± 0.00060.2487 ± 0.00040.1579 ± 0.00230.2565 ± 0.00100.1934 ± 0.0010
VS0.0321 ± 0.00030.0880 ± 0.00140.0988 ± 0.00260.1132 ± 0.00110.0959 ± 0.0038
  • Noise-dependent routing has worse mean CRPS than standard self-attention on all five datasets, so slice-based aggregation alone does not explain the improvement of DiCTS.
  • With the content path unchanged and only the routing replaced, history-guided routing improves mean CRPS on all five datasets and mean VS on four of five. On USHCN, noise-dependent routing has the lower mean VS (by 0.0050), which is within its seed standard deviation (0.0107).
  • Against standard self-attention, the VS gains are not uniform: HiSSA improves VS on four of five datasets (not PhysioNet'12).

2. Counterfactual noise intervention. To test directly whether the grouping under noise-dependent routing is driven by the injected noise, we fix the observed history, query coordinates, ground-truth future, and diffusion noise level for each test instance, and independently resample the Gaussian noise eight times at each τ∈{0.1,0.3,0.5,0.7,0.9}\tau \in \lbrace 0.1, 0.3, 0.5, 0.7, 0.9 \rbrace. We measure the assignment-induced mixing matrix WR=Sdiag⁡(1/m)S⊤,W^R = S \operatorname{diag}(1/m) S^\top, where SS denotes the query-to-slice assignments and mm the slice masses, and report Dnoise(WR)/Dperm(WR)D_{\mathrm{noise}}(W^R)/D_{\mathrm{perm}}(W^R): the noise-induced change relative to the change under random query permutations (0 = noise-invariant; 1 = permutation reference).

Table. Noise-induced change in query grouping on five IMTS benchmarks (0 = noise-invariant; 1 = permutation reference).

DatasetNoise-dependent routing (range across τ\tau)History-guided routing (by construction)
HumanActivity0.114–0.2740.000
PhysioNet'120.104–0.1690.000
USHCN0.505–0.6790.000
MIMIC-III0.133–0.2520.000
MIMIC-IV0.296–0.5650.000
  • Under noise-dependent routing, resampling the noise alone regroups queries on every dataset and at every noise level, by 10.4–67.9% of the permutation reference, although the history and the underlying future dynamics are unchanged.
  • This confirms the premise of Section 3.3: under noise-dependent routing, which queries exchange information is determined in part by the injected noise rather than by the temporal evidence. Removing this dependence improves CRPS on all five datasets (1) and the band-wise energy score in 12 of 15 dataset–band cells (3).
Reviewer_6CTROpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes DiCTS, a coordinate-based conditional diffusion model for irregular multivariate time series forecasting, whose denoiser uses History-guided Slice Self-Attention (HiSSA): queries are softly assigned to learned slices using an assignment derived from query coordinates and history cross-attention rather than from the noisy denoising state. The paper reports CRPS/VS and MAE/MSE on five benchmarks, a sample-level spectral diagnostic for forecast smoothing, and an ablation.

Critical strengths
  • The coordinate-based formulation avoids re-gridding or interpolation and is a clean fit for irregular time series.
  • The HiSSA routing is genuinely noise-free and τ\tau-independent, and the claim that fixed assignments do not imply fixed interactions is correct given eqs. (6)-(8).
Critical weaknesses
  1. HiSSA (eqs. 5-8) maps one-to-one onto Transolver's Physics-Attention [1]: softmax slice weights, slice summaries normalised by slice mass, self-attention across slices, deslice. The only conceptual change is the routing input, MLP([Eq;Cq])\mathrm{MLP}([E^q; C^q]) instead of the input features, plus a learnable temperature and an ε\varepsilon-floored slice mass. Routing from clean rather than noise-corrupted features is already published for diffusion models in [3], with the same stated diagnosis and the same remedy principle. The problem statement is also not new: [4] already attributes high-frequency suppression in diffusion forecasting to the forward process and proposes a remedy, and [5] already targets over-smoothing in time-series forecasting. Transolver++ [2] is cited but never discussed or compared.

  2. The core hypothesis is never tested. HiSSA's claim is that history-guided grouping beats noise-guided grouping. The ablation (Table 3, Section 4.5) compares HiSSA only against standard query self-attention, i.e. no grouping at all. The control that would test the claim — the same block with routing computed from the modulated noisy state mods(Xτ,τ)\mathrm{mod}_s(X_\tau, \tau) instead of [Eq;Cq][E^q; C^q] — is absent, so the paper's central mechanism claim is asserted rather than demonstrated. The other ablation row is also confounded: "DiT" versus "Ours w/o HiSSA" changes the conditioning pathway, the history retrieval pathway and the attention mechanism simultaneously, and Section 3.2 pre-declares the DiT-style conditioning inadequate; no parameter counts or FLOPs are reported for any row.

  3. The smoothing claim is unquantified, and where it can be checked it fails on the paper's own dataset. The claim rests on visual inspection of histograms: no summary statistic, no confidence interval and no test is reported anywhere, and the Low/Middle/High band edges are never defined. Recovering the bin masses from the vector figures shows that on PhysioNet'12 — the dataset used for the main-text demonstration — DiCTS still attenuates (mean log⁡2r=−0.228\log_2 r = -0.228, about 0.85×0.85\times amplitude; P(r<1)=0.57P(r<1)=0.57), and MOSES is better on both natural summaries (∣mean∣=0.091|\text{mean}| = 0.091 vs 0.2280.228; P(∣log⁡2r∣<1)=0.681P(|\log_2 r|<1) = 0.681 vs 0.6760.676). DiCTS is best on three of the four datasets analysed, but not on the one chosen to demonstrate the effect. The rank histograms show DiCTS edge mass of 0.2940.294-0.2990.299 against a uniform 0.2000.200, the same pattern the paper interprets as insufficient ensemble spread for LLapDiff, so the diagnostic is applied with a double standard.

  4. Table 1's DiCTS VS entries for USHCN (0.1031±0.00950.1031 \pm 0.0095) and MIMIC-III (0.1153±0.00200.1153 \pm 0.0020) are character-for-character Table 3's "Ours w/o HiSSA" values; Table 3 gives 0.0988±0.00260.0988 \pm 0.0026 and 0.1132±0.00110.1132 \pm 0.0011 for the same model, metric and dataset. Two of the paper's ten DiCTS cells therefore report the ablated model. Section 4.5 then states that "Ours w/o HiSSA" achieves lower mean CRPS and MSE than DiT on all five datasets — Table 3 has no MSE column, and on VS the variant is worse than DiT on USHCN (0.10310.1031 vs 0.09990.0999) — and that adding HiSSA further improves both metrics across all five datasets, which is false for VS on PhysioNet'12 (0.0867→0.08800.0867 \to 0.0880; the table's own bold marks 0.08670.0867 as better). The MIMIC-IV CRPS change it relies on is 0.00050.0005 against a standard deviation of 0.00100.0010.

Minor

  • On USHCN, VS is 44%44\% worse than MOSES (0.10310.1031 vs 0.07160.0716, p=0.0005p=0.0005). "Competitive" in the abstract understates a significant loss on a reported metric.
  • Fairness is not assessable. Section 4.1 gives one shared training recipe for all models, contradicting Appendix B's statement that baselines follow their original papers' settings; the number of forecast samples behind CRPS/VS is never stated; the rule converting DiCTS samples into the point prediction used for MAE/MSE is never stated; and DiCTS's across-seed standard deviations are 2.42.4-3×3\times tighter than every baseline, suggesting an undocumented multi-sample averaging. Table 2's GraFITi and HyperIMTS cells differ from those models' published tables by more than the published standard deviations (e.g. GraFITi MAE on USHCN 0.25990.2599 vs 0.30290.3029), so "reproduced" is unverifiable.
Questions for authors
  1. Why do Table 1 and Table 3 disagree on DiCTS's VS for USHCN and MIMIC-III, and which values are correct?

  2. Section 4.5's two claims are contradicted by Table 3 and refer to an MSE column that does not exist.

  3. How is the point prediction for MAE/MSE obtained?

  4. How does HiSSA differ in substance from Transolver's Physics-Attention [1] and from noise-free routing in diffusion MoE [3]?

9 comments
AuthorsOpenReview ↗

Author Response 1: Summary

We thank the reviewer for the careful and detailed assessment. We provide additional experiments, statistics, and corrections addressing the following concerns:

  1. Novelty over Transolver, Transolver++, and SharpMoE: Our contribution is the routing source, not the slice-attention operator. HiSSA routes on query coordinates and observed history rather than noisy states or model predictions, so its grouping is invariant to the injected noise and reusable across steps and samples.
  2. Validation of the core hypothesis: The requested noise-dependent control is worse than standard self-attention in CRPS on all five datasets, whereas history-guided routing is the best on all five. Resampling the noise alone regroups queries under the control (10.4–67.9% of a permutation reference), and history-guided routing has the lower band-wise energy score in 12 of 15 dataset–band cells (paired 95% intervals exclude zero in all 12).
  3. Quantification of forecast smoothing: We report summary statistics with 95% confidence intervals and define the frequency bands. On PhysioNet'12, DiCTS has the lowest absolute log-amplitude error and rank-histogram edge mass among the seven models; we acknowledge its residual attenuation and apply the same reading of the rank histograms to every model.
  4. Evaluation protocol: We have corrected Table 1 and Section 4.5, documented the evaluation protocol, and shown that the conclusions of Table 2 hold with the published GraFITi and HyperIMTS values.

Detailed results and discussions follow below.

AuthorsOpenReview ↗

Author Response 2: Detailed Response

W1 & Q4. Lack of Novelty Relative to Existing Studies

We thank the reviewer for the opportunity to clarify and further highlight the novelty of our work. Our contribution lies not in introducing a new slice-attention operator, but in noise-decoupled routing for organizing interactions among irregular future queries during diffusion-based IMTS forecasting. As the review does not include bibliographic details for references [1]–[5], we have interpreted these citations as referring to the works listed below.

Existing diffusion models may attenuate fine-scale temporal variations by mixing queries with heterogeneous dynamics. While slice-based routing provides a mechanism for organizing query interactions, noise-dependent routing can induce noise-driven dynamics mixing, because stochastic similarities may group queries with distinct temporal dynamics. Since routing assignments govern both aggregation and redistribution, these mismatched associations can propagate through both directions of information exchange, potentially suppressing fine-scale variations. This noise dependence is examined through the counterfactual noise intervention (Response to W2-1).

Relation to Transolver-based architectures. DiCTS addresses this problem by explicitly decoupling the organization of query interactions from the stochastic representations being denoised. In Transolver [1] and Transolver++ [2], slice weights are computed from the same point features that are subsequently aggregated; placed inside a diffusion denoiser, this design makes the routing a function of the noisy query states. Transolver++ additionally introduces a per-point adaptive temperature and a Gumbel-softmax reparameterization of the slice weights, which sharpen the assignments but do not change their source. The DiCTS (Noise-dependent routing) variant in W2-1 uses this Transolver-style routing source within the same denoiser and serves as the corresponding empirical comparison. In contrast, DiCTS derives assignments exclusively from observed history and target time–channel coordinates, without relying on noisy states or estimated clean future representations. While assignments remain fixed across denoising steps, slice representations and their interactions remain sample-dependent. Thus, historical evidence determines which queries exchange information, while denoising states determine what information is exchanged. This separation is designed to mitigate noise-induced dynamics mixing and preserve fine-scale temporal variations.

Relation to SharpMoE [3]. The two methods differ in what is routed, in the routing signal, and in the problem addressed. First, [3] assigns tokens to experts in a diffusion MoE so that more computation is allocated to salient tokens, whereas HiSSA determines which queries exchange information. Second, [3] routes on the model's clean prediction, whereas HiSSA routes only on the observed history and the query coordinates. These inputs are observed rather than estimated, so the assignments are exactly invariant to the injected noise and to the denoising step, and they can be computed once and reused across denoising steps and forecast samples. Third, [3] is a post-training framework for pretrained diffusion MoE models in image generation, whereas HiSSA is trained end-to-end for IMTS forecasting.

Regarding the problem statement. DiCTS addresses forecast smoothing in IMTS, where sparse and asynchronous observations leave fine-scale future dynamics weakly constrained and mixing across queries with distinct dynamics can suppress temporal variation. In comparison, [4] attributes the loss of structured temporal features to the forward diffusion process and modifies that process, and [5] attributes over-smoothing to latent dynamical mode compression. None of [3]–[5] directly addresses noise-induced dynamics mixing among irregular future time–channel queries and its potential role in attenuating fine-scale temporal variations during denoising.

[1] Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for PDEs on general geometries. In ICML, volume 235, pp. 53681–53705, 2024. [2] Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Transolver++: An accurate neural solver for PDEs on million-scale geometries. In ICML, volume 267, pp. 41432–41449, 2025. [3] Haoyou Deng, Keyu Yan, Chaojie Mao, Xiang Wang, Yu Liu, Changxin Gao, and Nong Sang. Focusing on what matters: Saliency-harnessing accurate routing for diffusion MoE. In ECCV, 2026. [4] Francisco Caldas, Sahil Kumar, and Cláudia Soares. A decomposable forward process in diffusion models for time-series forecasting. arXiv preprint arXiv:2601.21812, 2026. [5] Xingyu Zhang, Jingyao Wang, Xin Yu, Zeen Song, Jianqi Zhang, Changwen Zheng, and Wenwen Qiang. Dirichlet-guided group forecasting for alleviating over-smoothing in time series forecasting. arXiv preprint arXiv:2606.10592, 2026.

AuthorsOpenReview ↗

Author Response 3: Detailed Response

W2-1. Validation of the Core Hypothesis

We thank the reviewer for pointing out that our ablation lacked the noise-dependent routing control needed to test the core hypothesis directly.

1. Routing ablation. We add DiCTS (Noise-dependent routing): slice attention with assignments derived from the noisy denoising states. DiCTS (Standard Self-Attn) is the former "Ours w/o HiSSA" (no slice-based routing), and DiCTS (Ours) routes only on observed history and target coordinates. Section 4.5 is revised accordingly.

Table R1. Ablation on five IMTS benchmarks (mean ± sample standard deviation over five seeds; lower is better; best in bold).

ModelMetricHumanActivityPhysioNet'12USHCNMIMIC-IIIMIMIC-IV
DiTCRPS0.1133 ± 0.00150.2778 ± 0.00140.1657 ± 0.00540.2972 ± 0.00220.2399 ± 0.0013
VS0.0358 ± 0.00050.0910 ± 0.00030.0999 ± 0.00980.1232 ± 0.00080.1193 ± 0.0015
MSE0.0611 ± 0.00260.3337 ± 0.00210.2083 ± 0.03100.4736 ± 0.00460.2787 ± 0.0025
DiCTS (Standard Self-Attn)CRPS0.0981 ± 0.00250.2632 ± 0.00250.1617 ± 0.00090.2767 ± 0.00340.1939 ± 0.0010
VS0.0325 ± 0.00060.0867 ± 0.00020.1031 ± 0.00950.1153 ± 0.00200.0972 ± 0.0030
MSE0.0504 ± 0.00410.3103 ± 0.00320.1911 ± 0.00220.4309 ± 0.00630.2133 ± 0.0008
DiCTS (Noise-dependent routing)CRPS0.1128 ± 0.00230.2784 ± 0.00140.1678 ± 0.00480.2993 ± 0.00150.2444 ± 0.0018
VS0.0357 ± 0.00060.0921 ± 0.00020.0938 ± 0.01070.1320 ± 0.00170.1352 ± 0.0021
MSE0.0595 ± 0.00320.3354 ± 0.00140.2205 ± 0.02520.4786 ± 0.00360.2866 ± 0.0041
DiCTS (Ours)CRPS0.0829 ± 0.00060.2487 ± 0.00040.1579 ± 0.00230.2565 ± 0.00100.1934 ± 0.0010
VS0.0321 ± 0.00030.0880 ± 0.00140.0988 ± 0.00260.1132 ± 0.00110.0959 ± 0.0038
MSE0.0418 ± 0.00090.2857 ± 0.00110.1734 ± 0.01390.3875 ± 0.00540.2127 ± 0.0029
  • Noise-dependent routing has worse CRPS than Standard Self-Attn on all five datasets, so slice-based aggregation alone does not explain the improvement of DiCTS.
  • DiCTS (Ours) has the lowest mean CRPS on all five datasets and, against noise-dependent routing, lower mean VS on four of five. This supports our core hypothesis that the routing source, rather than slice-based aggregation alone, is critical for information exchange during denoising.

2. Counterfactual noise intervention. For each test instance and noise level τ∈{0.1,0.3,0.5,0.7,0.9}\tau \in \lbrace 0.1, 0.3, 0.5, 0.7, 0.9 \rbrace, we fix the history, query coordinates, and ground-truth future, and resample the Gaussian noise eight times. With WR=Sdiag⁡(1/m)S⊤W^R=S \operatorname{diag}(1/m) S^\top (SS: query-to-slice assignments; mm: slice masses), we report Dnoise(WR)/Dperm(WR)D_{\mathrm{noise}}(W^R)/D_{\mathrm{perm}}(W^R): the noise-induced change relative to the change under random query permutations.

Table R2. Counterfactual noise intervention: Dnoise(WR)/Dperm(WR)D_{\mathrm{noise}}(W^R)/D_{\mathrm{perm}}(W^R) for noise-dependent and history-guided routing on five IMTS benchmarks (lower is more noise-invariant; best in bold).

DatasetNoise-dependent routing (range across τ\tau)History-guided routing
HumanActivity0.114–0.2740.000
PhysioNet'120.104–0.1690.000
USHCN0.505–0.6790.000
MIMIC-III0.133–0.2520.000
MIMIC-IV0.296–0.5650.000
  • Noise alone regroups queries under noise-dependent routing on every dataset and noise level, reaching 67.9% (USHCN) and 56.5% (MIMIC-IV) of the permutation reference, although the history and the future dynamics are unchanged.

3. Frequency-domain comparison. Both variants use the same test instances and 100 joint forecast samples per instance. Intervals are paired 95% percentile bootstrap intervals.

Table R3. Paired difference in the band-wise energy score of the complex NUDFT coefficients (a proper scoring rule that penalizes attenuated, over-amplified, and misplaced fluctuations alike), noise-dependent minus history-guided [95% interval; 1,000 cluster-level resamples]. A positive value favors history-guided routing.

DatasetLowMiddleHigh
HumanActivity+0.00123 [+0.00064, +0.00195]+0.00229 [+0.00158, +0.00362]+0.00222 [+0.00149, +0.00343]
PhysioNet'12+0.00202 [+0.00075, +0.00321]+0.00117 [+0.00004, +0.00231]+0.00126 [+0.00002, +0.00237]
USHCN-0.00281 [-0.00664, +0.00049]-0.00369 [-0.00645, -0.00107]-0.00218 [-0.00512, +0.00019]
MIMIC-III+0.00821 [+0.00556, +0.01097]+0.00493 [+0.00304, +0.00696]+0.00399 [+0.00226, +0.00582]
MIMIC-IV+0.00624 [+0.00437, +0.00808]+0.00311 [+0.00219, +0.00413]+0.00355 [+0.00251, +0.00461]
  • History-guided routing has the lower score in 12 of 15 dataset–band cells (by 0.7–5.4%; all 12 intervals exclude zero), including the high band on four of five datasets (by 0.8–5.1%) and all three bands of PhysioNet'12.
  • The gain is not limited to slow trends: in the high band it holds on four of five datasets (by 0.8–5.1%)
Reviewer_vj8mOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the problem of forecast smoothing in diffusion-based probabilistic forecasting for irregular multivariate time series (IMTS). The authors attribute this phenomenon to dynamics mixing caused by grouping queries based on noisy intermediate diffusion states, which erroneously clusters queries with similar random noise patterns rather than shared underlying physical dynamics.

To address this issue, the paper proposes the Diffusion Coordinate Transformer with history-guided Slice self-attention (DiCTS). DiCTS incorporates History-guided Slice Self-Attention (HiSSA), which decouples the assignment of query tokens to slices from the diffusion noise.

Evaluations across five established IMTS benchmarks (spanning clinical health records, climate monitoring, and human motion capture) demonstrate that DiCTS achieves competitive marginal probabilistic accuracy (CRPS), improves intra-trajectory temporal structure preservation (Variogram Score), and yields strong deterministic point-prediction accuracy (MAE/MSE).

Critical strengths
  1. The design of HiSSA decouples the grouping topology from the sample-specific content. By computing slice routing strictly from time-channel coordinates and historical context, the model prevents diffusion noise from perturbing query interactions. Moreover, because the soft assignment matrix SS is invariant to the diffusion timestep τ\tau, it can be computed once and reused across all sampling steps.

  2. Rather than relying solely on marginal error metrics, the authors employ diagnostic tools directly tailored to evaluating trajectory smoothing. Specifically, the use of Lomb-Scargle power spectrum analysis appropriately accommodates irregular observation intervals, providing rigorous quantitative evidence that the model avoids band-amplitude attenuation without inducing excessive noise amplification. The inclusion of the Variogram Score further corroborates the preservation of temporal dependencies.

  3. The empirical validation encompasses five datasets with distinct dimensionalities, sparsity levels, and sampling behaviors. DiCTS is compared against six recent probabilistic IMTS architectures (including diffusion models, neural SDEs, and continuous normalizing flows) and seven deterministic baselines, demonstrating consistent gains in both probabilistic calibration and point forecasts.

Critical weaknesses
  1. The absence of a direct control group, specifically one involving "slice allocation based on noisy states" (Noisy-State Slicing), makes it impossible to definitively rule out the possibility that the performance gains stem solely from the low-rank bottleneck effect inherent to slice attention itself.

  2. Although the reusability of precomputed data is highlighted, no quantitative evaluation data (such as end-to-end sampling latency, peak VRAM usage, or FLOPs) is provided.

  3. Slice routing relies heavily on historical sequences; consequently, when faced with extreme sparsity or abrupt, dynamic shifts, it may be subject to biases stemming from historical priors.

Questions for authors
  1. How would the CRPS, VS, and frequency-band amplitude ratio change if the slice allocation were modified to be dynamically calculated from the noisy state at each denoising step?

  2. Could authors provide empirical comparisons of DiCTS against full self-attention and other diffusion baselines, specifically in terms of sampling latency and VRAM usage?

  3. When dealing with datasets that differ significantly in the number of variables (ranging from 5 to 100 channels), does fixing the number of slices at G=32G=32 lead to overly coarse clustering or a collapse due to overfitting?

6 comments
AuthorsOpenReview ↗

Author Response 1: Summary

We thank the reviewer for the careful and constructive assessment. To address the concerns about the missing control, computational cost, and the design choices of history-guided routing, we provide additional evidence on four aspects:

  1. Validation of the core hypothesis (W1 & Q1): We add a noisy-state slicing control that keeps the same slice attention. Resampling only the diffusion noise changes its grouping by 10.4–67.9% of a permutation reference, and it is worse than standard self-attention in CRPS on all five datasets, so slice attention alone does not explain the gains. In the frequency-band diagnostic, history-guided routing has the lowest absolute amplitude error of the three variants in 12 of 15 dataset–band cells.
  2. Computational cost (W2 & Q2): We report sampling FLOPs, latency, and peak VRAM on MIMIC-IV. DiCTS samples 1.5–23.3× faster than the diffusion baselines, and reusing the history-guided assignments reduces FLOPs by 7.00% and latency by 17.43% relative to noise-dependent routing, while its latency remains higher than that of standard self-attention.
  3. Reliance on history (W3): We clarify that the history determines only which queries exchange information, not what they exchange, and we state the resulting limitation in the revised discussion.
  4. Fixed number of slices (Q3): The median number of queries per instance is below G=32G=32 on all five datasets, and a sensitivity study over G∈{8,16,32,64}G \in \lbrace 8, 16, 32, 64 \rbrace shows that the CRPS of the default is within 1.5% of the best setting on every dataset.

We provide the experimental details and quantitative results below.

AuthorsOpenReview ↗

Author Response 2: Detailed Response

W1 & Q1. Validation of the Core Hypothesis

We thank the reviewer for pointing out the missing control. We added it and report the quantities requested in Q1.

Control. DiCTS (Noise-dependent routing) keeps the slice attention of Eqs. (6)–(8) but recomputes the assignments at every denoising step from the noisy query states (same GG, temperature, and training recipe). DiCTS (Standard Self-Attn) is "Ours w/o HiSSA"; its attention weights also depend on the noisy query states.

1. Noise destabilizes noise-dependent routing. With the history, query coordinates, future, and noise level τ∈{0.1,0.3,0.5,0.7,0.9}\tau \in \lbrace 0.1, 0.3, 0.5, 0.7, 0.9 \rbrace fixed, we resample only the Gaussian noise and report the change in the mixing matrix WR=Sdiag⁡(1/m)S⊤W^R = S \operatorname{diag}(1/m) S^\top relative to random query permutations (0 = unchanged; 1 = permutation reference).

DatasetNoise-dependent routing (range across τ\tau)History-guided routing
HumanActivity0.114–0.2740.000
PhysioNet'120.104–0.1690.000
USHCN0.505–0.6790.000
MIMIC-III0.133–0.2520.000
MIMIC-IV0.296–0.5650.000

Noise alone regroups queries by 10.4–67.9% of the permutation reference. History-guided assignments take no noisy input, so they are unchanged by construction.

2. CRPS and VS (Q1). Mean ± sample std over five seeds; lower is better.

ModelMetricHumanActivityPhysioNet'12USHCNMIMIC-IIIMIMIC-IV
DiCTS (Standard Self-Attn)CRPS0.0981 ± 0.00250.2632 ± 0.00250.1617 ± 0.00090.2767 ± 0.00340.1939 ± 0.0010
VS0.0325 ± 0.00060.0867 ± 0.00020.1031 ± 0.00950.1153 ± 0.00200.0972 ± 0.0030
DiCTS (Noise-dependent routing)CRPS0.1128 ± 0.00230.2784 ± 0.00140.1678 ± 0.00480.2993 ± 0.00150.2444 ± 0.0018
VS0.0357 ± 0.00060.0921 ± 0.00020.0938 ± 0.01070.1320 ± 0.00170.1352 ± 0.0021
DiCTS (Ours)CRPS0.0829 ± 0.00060.2487 ± 0.00040.1579 ± 0.00230.2565 ± 0.00100.1934 ± 0.0010
VS0.0321 ± 0.00030.0880 ± 0.00140.0988 ± 0.00260.1132 ± 0.00110.0959 ± 0.0038
  • Slice attention alone does not explain the gains. With the same slice attention, noise-dependent routing is worse than Standard Self-Attn in CRPS on all five datasets (3.8–26.0%) and in VS on four. Moreover, the median number of queries per instance is below G=32G=32 on all five datasets (Q3), so the slices rarely impose a low-rank constraint.
  • History guidance matters. Relative to DiCTS, noise-dependent routing increases CRPS on all five datasets (6.3–36.1%) and VS on four (4.7–41.0%); on USHCN its mean VS is lower, within its seed-level standard deviation.

3. Frequency-band amplitude ratio (Q1). Diagnostic of Section 4.3 (100 joint samples per instance; one seed per model). Signed = E[log⁡2r]\mathbb{E}[\log_2 r] (below zero: attenuation); Abs. = E[∣log⁡2r∣]\mathbb{E}[\lvert\log_2 r\rvert] (absolute log-amplitude error; lower is better). Cells list Low / Middle / High.

ModelMetricHumanActivityPhysioNet'12USHCNMIMIC-IIIMIMIC-IV
DiCTS (Standard Self-Attn)Signed-0.182 / -0.073 / -0.030-0.415 / -0.404 / -0.352-1.620 / -1.596 / -1.612-0.181 / -0.247 / -0.255-0.423 / -0.188 / 0.031
Abs.1.222 / 1.200 / 1.2451.033 / 1.011 / 1.0012.290 / 2.283 / 2.2821.104 / 1.155 / 1.1622.080 / 2.221 / 2.253
DiCTS (Noise-dependent routing)Signed-0.013 / 0.269 / 0.325-0.335 / -0.312 / -0.283-1.331 / -1.324 / -1.352-0.252 / -0.219 / -0.169-0.291 / 0.196 / 0.450
Abs.1.251 / 1.233 / 1.2880.969 / 0.961 / 0.9662.087 / 2.093 / 2.1041.165 / 1.204 / 1.2002.090 / 2.202 / 2.272
DiCTS (Ours)Signed-0.072 / -0.078 / -0.105-0.374 / -0.372 / -0.321-1.299 / -1.305 / -1.326-0.193 / -0.196 / -0.149-0.282 / 0.157 / 0.412
Abs.1.191 / 1.176 / 1.2280.992 / 0.975 / 0.9672.076 / 2.090 / 2.0971.082 / 1.120 / 1.1202.039 / 2.179 / 2.251
  • Closest amplitude agreement. DiCTS has the lowest absolute error in 12 of 15 dataset–band cells: lower than Standard Self-Attn in all 15 (0.1–9.3%) and lower than noise-dependent routing on four datasets (0.1–7.1%); on PhysioNet'12 it is higher by 0.1–2.4%. The 95% bootstrap intervals of the two routing variants overlap in every cell.
  • Less attenuation than noisy-state self-attention. In the high band, DiCTS reduces the attenuation of Standard Self-Attn on PhysioNet'12 (-0.352 → -0.321), USHCN (-1.612 → -1.326), and MIMIC-III (-0.255 → -0.149); its signed mean is closer to zero in 11 of 15 cells.

Summary. These results support the claim that grouping from noisy states can mix queries with distinct dynamics and that history-guided routing mitigates the resulting forecast smoothing.

AuthorsOpenReview ↗

Author Response 3: Detailed Response

W2 & Q2. Computational cost: sampling latency, memory, and FLOPs

We thank the reviewer for requesting quantitative cost measurements and report them below. Sampling costs on MIMIC-IV (RTX A6000, FP32, batch size 32, 100 joint forecasts per series; DiCTS variants use 32 sampling steps, baselines the configurations in Appendix B; FLOPs averaged over three batches):

ModelCounted sampling FLOPs (G/batch)Sampling latency (ms/batch)Peak VRAM (GiB)
TSDiff990.9847,927.20.297
TDSTF779,301.13751,016.71.130
LLapDiff1,268,721.349124,147.76.591
DiCTS (Standard Self-Attn)9,979.9873,223.52.963
DiCTS (Noise-dependent routing)10,636.8856,462.22.963
DiCTS (Ours)9,892.0445,335.62.982
  • Diffusion baselines. TSDiff, TDSTF, and LLapDiff take 1.5×, 9.6×, and 23.3× the sampling latency of DiCTS; the peak VRAM of DiCTS is 10.0×, 2.6×, and 0.45× theirs.
  • Standard self-attention. DiCTS has comparable FLOPs (0.9% lower) and peak VRAM (0.019 GiB higher) but 1.66× the latency, so we do not claim a cost advantage over it.
  • Noise-dependent routing. History-guided assignments are reused across forecast samples and denoising steps, which reduces FLOPs by 7.00% and latency by 17.43%, with 0.019 GiB additional peak VRAM.

W3. Reliance on History under Extreme Sparsity or Abrupt Shifts

We thank the reviewer for raising this point and agree that this is a real trade-off of history-guided routing.

  • What the history determines. Within HiSSA, the history sets only which queries exchange information (Eqs. 4–5); what they exchange follows the current denoising states (Eqs. 6–7). The routing feature also contains the query coordinates, which remain available when the history is sparse.
  • Evidence. Noise-dependent routing, which can regroup queries at every step, has a worse mean CRPS than history-guided routing on all five datasets. These averages do not isolate extremely sparse or abruptly shifting instances.
  • Limitation. Assignments are shared across forecast samples and denoising steps, so a future that departs from the history can be expressed through slice contents and attention weights but not by regrouping queries. We have added this limitation in Section 5 and leave adaptive routing to future work.

Q3. Fixed Number of Slices across Datasets

We thank the reviewer for this question and address it with the number of queries per instance and a sensitivity study over GG.

Queries per instance. Slices group future time–channel queries, not channels. The median (maximum) number of queries per test instance is 27 (33), 19 (42), 3 (5), 7 (60), and 4 (47) for HumanActivity, PhysioNet'12, USHCN, MIMIC-III, and MIMIC-IV. The median is below G=32G=32 on all five datasets and even the largest instance has fewer than two queries per slice, so a fixed G=32G=32 does not force a coarse clustering.

Sensitivity to GG. We vary G∈{8,16,32,64}G \in \lbrace 8, 16, 32, 64 \rbrace with all other settings fixed (five seeds; bold marks the lowest mean).

GGMetricHumanActivityPhysioNet'12USHCNMIMIC-IIIMIMIC-IV
8CRPS0.0885 ± 0.00020.2480 ± 0.00050.1645 ± 0.00400.2527 ± 0.00050.2014 ± 0.0010
VS0.0326 ± 0.00120.0883 ± 0.00100.1018 ± 0.00400.1248 ± 0.00100.0979 ± 0.0013
16CRPS0.0828 ± 0.00050.2490 ± 0.00030.1620 ± 0.00490.2572 ± 0.00040.1941 ± 0.0004
VS0.0318 ± 0.00100.0858 ± 0.00110.0985 ± 0.01090.1234 ± 0.00110.0939 ± 0.0048
32 (default)CRPS0.0829 ± 0.00060.2487 ± 0.00040.1579 ± 0.00230.2565 ± 0.00100.1934 ± 0.0010
VS0.0321 ± 0.00030.0880 ± 0.00140.0988 ± 0.00260.1132 ± 0.00110.0959 ± 0.0038
64CRPS0.0833 ± 0.00020.2512 ± 0.00060.1612 ± 0.00200.2604 ± 0.00180.1974 ± 0.0011
VS0.0319 ± 0.00100.0892 ± 0.00110.0890 ± 0.01790.1231 ± 0.00110.1033 ± 0.0018
  • The default is close to the best setting. The CRPS of G=32G=32 is within 1.5% of the best GG on all five datasets (best on USHCN and MIMIC-IV), and its VS is within 2.6% on four; on USHCN, G=64G=64 has the lower mean VS, within its seed-level standard deviation.
  • Too few slices can hurt. G=8G=8 raises CRPS by 6.8% on HumanActivity, 4.2% on USHCN, and 4.1% on MIMIC-IV, although it is the best on PhysioNet'12 and MIMIC-III (0.3% and 1.5%).
  • If G=32G=32 were too coarse, a finer partition would help, but G=64G=64 has a higher CRPS on all five datasets (0.5–2.1%). If it overfit where queries are few, a smaller GG would generalize better, but on USHCN and MIMIC-IV the default has the lowest test CRPS.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Final Remarks

Dear AC and Reviewers,

We sincerely thank you for your valuable time, constructive comments, and detailed feedback on our work. We greatly appreciate the reviewers' efforts in carefully evaluating both the technical soundness and the empirical evidence of our paper.


[Acknowledgement of Strengths]

We appreciate the reviewers' recognition:

  • History-guided routing. Computing slice routing from time–channel coordinates and historical context prevents diffusion noise from perturbing query interactions (vj8m). The routing is noise-free and independent of the denoising step, and fixed assignments do not imply fixed interactions (6CTR). It is a reasonable approach to reducing noise-driven feature mixing during denoising (aE3u).
  • Reusable assignments. The assignments can be computed once and reused across all sampling steps (vj8m).
  • Coordinate-based formulation. It avoids re-gridding or interpolation and is a clean fit for irregular time series (6CTR).
  • Smoothing diagnostics. The Lomb–Scargle analysis is tailored to evaluating trajectory smoothing under irregular observation intervals, and the Variogram Score further corroborates the preservation of temporal dependencies (vj8m). The spectral analyses provide useful evidence on the preservation of temporal variation (aE3u).
  • Empirical validation. The evaluation covers five datasets with six probabilistic and seven deterministic baselines (vj8m), with competitive results across the five benchmarks (aE3u).

[Summary of Rebuttal]

We summarize our responses and additional experiments below:

  • Validation of the core hypothesis (vj8m, 6CTR, aE3u). We added a noise-dependent routing variant that keeps the same slice attention but computes assignments from noisy states. It is worse than standard self-attention in CRPS on all five datasets, whereas history-guided routing has the lowest CRPS on all five. Resampling only the noise changes the query grouping under the noise-dependent variant on every dataset and at every noise level, while history-guided assignments are unchanged.
  • Novelty over prior work (6CTR, aE3u). Our contribution is the routing source, not the slice-attention operator. Unlike Transolver, Transolver++, and SharpMoE, HiSSA routes on query coordinates and observed history, so its grouping is invariant to the injected noise. Grouping by noisy states can mix queries with distinct temporal dynamics through stochastic similarities and suppress fine-scale variations; by grouping queries according to observed evidence instead, HiSSA mitigates this noise-driven mixing and the resulting forecast smoothing.
  • Quantification of forecast smoothing (6CTR, aE3u). We report summary statistics with 95% confidence intervals and define the frequency bands. On PhysioNet'12, DiCTS has the lowest absolute amplitude error and rank-histogram edge mass among the seven models, and history-guided routing has a lower band-wise energy score than noise-dependent routing in 12 of 15 dataset–band cells.
  • Computational cost (vj8m, 6CTR, aE3u). We report parameter counts, FLOPs, sampling latency, peak VRAM, and training time. On MIMIC-IV, DiCTS samples 1.5–23.3× faster than the diffusion baselines, and reusing the assignments reduces FLOPs by 7.00% and latency by 17.43% relative to noise-dependent routing.
  • Number of slices (vj8m, aE3u). A sensitivity study over G ∈ {8, 16, 32, 64} shows that the CRPS of the default G = 32 is within about 1.5% of the best setting on every dataset.
  • Reliance on history (vj8m). The history determines only which queries exchange information, not what they exchange. We acknowledge the resulting limitation under very sparse histories or abrupt shifts and have added it to the Discussion section.
  • Reporting and evaluation protocol (6CTR, aE3u). We clarified the inconsistency in the VS entries between Tables 1 and 3 and in the ablation, and explained the evaluation protocol, including how the point prediction is obtained. We also compared our reproduced GraFITi and HyperIMTS results with the published ones: with the published values, DiCTS still has the lowest MAE on all five datasets and the lowest MSE on four.
  • Visualization (aE3u). We explained how the displayed examples were selected and that our conclusions rest on aggregate statistics over the full test set rather than on individual examples.

After the rebuttal, Reviewer vj8m confirmed that their concerns had been addressed, Reviewer 6CTR stated that the new experiments addressed their main concern and raised their rating, and Reviewer aE3u confirmed that the rebuttal clarified several concerns.


We hope these clarifications and additional experiments address the concerns and support a fair final assessment of our work.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.