acceptodds
Under review as a conference paper at ICLR 2027

ResoPath: Disentangling Multi-Scale Continuous Paths with Missingness-Aware Neural CDEs for Irregular Time Series Forecasting

Abstract

Irregularly sampled multivariate time series play a pivotal role across science and engineering, yet non-uniform observation intervals and severe missingness make continuous dynamic modeling challenging. Existing neural controlled differential equation based methods collapse sparse observations into single interpolated paths, conflating multi-scale patterns with synthetic smoothing artifacts and ignoring observation reliability. In this paper, we introduce ResoPath, an end-to-end framework that constructs multi-resolution continuous control paths by disentangling coarse temporal trends from fine-scale fluctuations. ResoPath couples wavelet decomposition with frequency-conditioned denoising to construct continuous control paths that drive multiscale latent dynamics through vector fields conditioned on missingness-aware information. Extensive comparisons with diverse baselines across five real-world datasets demonstrate that ResoPath achieves state-of-the-art accuracy and strong robustness in the irregular multivariate time series forecasting task.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series Forecasting, multi-scale forecasting, neural controlled differential equation

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_bSdVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ResoPath, combining wavelet denoising, multiscale NCDEs, and observation-gap-conditioned FiLM modulation for irregular multivariate time series forecasting, and reports overall improvements over eighteen ODE/CDE baselines across five datasets.

Critical strengths
  1. Using separate frequency components to drive continuous-time dynamics provides a valuable modeling approach for handling trends, local fluctuations, and interpolation artifacts.
  2. Comparisons across five datasets show good overall performance, with particularly noticeable gains on Human Activity supporting the method’s practical potential.
Critical weaknesses
  1. The relationship between the future ODE in Section 3.4 and the linear temporal projection head in Table 3 is unexplained, leaving the forecasting architecture underlying the reported results and its continuous-time prediction capabilities unclear.
  2. Experiments compare only ODE/CDE models and omit cited, directly relevant methods such as APN, Hi-Patch, and TFMixer, so the evidence does not establish the claimed state-of-the-art performance for irregular time series forecasting overall.
Questions for authors
  1. Under Equation (15), when all channels are observed synchronously at every knot, both endpoints have zero gaps and unit recency, so how does MaDF capture declining reliability within a long intervening gap after interpolation?
  2. Is the Haar decomposition performed over observation indices or a uniform physical-time grid, and how does it avoid interpreting irregular sampling and zero-filled missing entries as genuine high-frequency changes?
  3. Could the authors provide total NFEs and end-to-end latency with units for Figure 3, clarifying whether these include wavelet processing, all NCDE branches, and future prediction, to establish the practical accuracy–cost tradeoff?
9 comments

We sincerely thank you for the constructive feedback of our formulation and experimental evaluation. Below we address each weakness and question in detail.

W1: Forecast-Head Specification & Reconciliation of Section 3.4 with Table 3

We clarify that all reported results in Tables 1–2 were produced using the autonomous latent ODE rollout followed by a point-wise output readout, as described in Section 3.4. The phrase "Shared linear temporal projection from the input length to the forecasting horizon" in Table 3 (Line 959) was an editing remnant from an early discrete-grid baseline configuration table and does not reflect the ResoPath forecast head.

Specifically, our forecasting stage operates as follows:

  1. Cross-Scale Terminal State Fusion (Eq. 5): At the historical boundary TT, the terminal states of the J+1J+1 scale-specific NCDE branches, ZT=[z(0)(T),…,z(J)(T)]∈R(J+1)dhZ_T = [z^{(0)}(T), \dots, z^{(J)}(T)] \in \mathbb{R}^{(J+1)d_h}, are fused via a learnable 1×11 \times 1 scale aggregation into the initial forecasting state zˉ(T)∈Rdh\bar{z}(T) \in \mathbb{R}^{d_h}.

  2. Future Autonomous Latent ODE Rollout (Reconciling Line 279 with Eq. 8): Because no control observations arrive during (T,T+H](T, T+H], the matrix-valued historical field in Eq. (8) (∈Rdh×(D+1)\in \mathbb{R}^{d_h \times (D+1)}) cannot be applied unchanged. Our statement in Line 279 ("where Fθ(⋅)F_\theta(\cdot) is obtained from Eq. (8)") meant that the future drift field gψ:Rdh×Rdm→Rdhg_\psi: \mathbb{R}^{d_h} \times \mathbb{R}^{d_m} \to \mathbb{R}^{d_h} retains the same FiLM-modulated residual backbone as Eq. (8), but replaces the matrix reshaping projection with a vector-valued output projection into Rdh\mathbb{R}^{d_h}. To account for growing uncertainty as the forecast horizon extends without new measurements, we track the elapsed observation age a(t)∈RDa(t) \in \mathbb{R}^D, yielding an autonomous system in the augmented state (u(t),a(t))(u(t), a(t)) for t∈[T,T+H]t \in [T, T+H]:

    \qquad \dot{a}(t) = \mathbf{1}_D, \qquad u(T) = \bar{z}(T), \quad a(T) = \delta(T).$$ Since no new observations are assimilated after $T$, $a(t) = \delta(T) + (t - T)\mathbf{1}_D$ increases linearly in closed form without resetting, and $\eta_{\mathrm{fut}}(a(t)) = E_m(M(a(t)))$ continuously informs the future vector field of the accumulating forecast horizon.
  3. Arbitrary-Horizon Querying & Point-wise Readout: Given any sorted set of future query timestamps tk=T+τkk=1K⊂(T,T+H]\\{t_k = T + \tau_k\\}_{k=1}^K \subset (T, T+H], the ODE solver evaluates u(tk)∈Rdhu(t_k) \in \mathbb{R}^{d_h} directly along the continuous trajectory, and a shared readout πη\pi_\eta maps each queried latent state independently to the DD observable channels:

    \qquad W_{\mathrm{out}} \in \mathbb{R}^{D \times d_h}.$$

The identical rollout and readout mechanism is used during both training and inference. We will revise Section 3.4 (fixing Line 279) and update the Prediction head entries in Table 3 so that all specifications and dimensions are completely consistent.

W2: Comparison with Directly Relevant Non-ODE/CDE Baselines

Our initial experimental design focused on 18 ODE/CDE baselines to directly evaluate our core methodological objective: resolving interpolation distortions, scale entanglement, and uniform-reliability assumptions within continuous-time NCDEs. However, we agree with the reviewer that omitting cited non-ODE/CDE IMTS architectures—specifically APN (Liu et al., 2026), Hi-Patch (Luo et al., 2025), and TFMixer (Qiu et al., 2026) (as well as ReIMTS (Li et al., 2026) noted by Reviewer xBzy)—left our broader SOTA claims across general IMTS forecasting insufficiently substantiated.

We have expanded our benchmark evaluation under the exact same data splits and preprocessing protocol, to include these directly relevant IMTS methods. The comparison results (MSE ×10−1\times 10^{-1}) are reported below. Bold indicates the best (lowest mean) result; italic indicates the second-best.

AlgorithmUSHCNHuman ActivityPhysioNet'12MIMIC-IIIMIMIC-IV
PrimeNet7.28 ± 0.1015.33 ± 0.1210.18 ± 0.1010.94 ± 0.089.57 ± 0.29
SeFT6.13 ± 0.328.47 ± 0.088.55 ± 0.0110.35 ± 0.067.72 ± 0.38
NeuralFlows5.60 ± 0.252.56 ± 0.055.32 ± 0.028.29 ± 0.055.86 ± 0.27
GNeuralFlow4.47 ± 0.542.61 ± 0.064.90 ± 0.037.07 ± 0.056.13 ± 0.12
tPatchGNN4.15 ± 0.311.04 ± 0.064.34 ± 0.026.37 ± 0.044.86 ± 0.16
Hi-Patch4.34 ± 0.801.39 ± 0.024.27 ± 0.055.46 ± 0.034.48 ± 0.23
Warpformer3.77 ± 0.621.14 ± 0.034.08 ± 0.015.13 ± 0.024.54 ± 0.22
TFMixer3.24 ± 0.690.96 ± 0.053.73 ± 0.006.02 ± 0.024.06 ± 0.14
APN3.16 ± 0.581.32 ± 0.043.69 ± 0.015.30 ± 0.034.12 ± 0.18
ReIMTS3.06 ± 0.020.72 ± 0.013.43 ± 0.015.15 ± 0.023.75 ± 0.09
ResoPath (ours)3.48 ± 0.740.83 ± 0.053.60 ± 0.005.36 ± 0.053.73 ± 0.20

Even against top-performing discrete patch- and frequency-mixing architectures, ResoPath demonstrates SOTA competitive performance: it achieves 1st place on MIMIC-IV and ranks 2nd place on both Human Activity and PhysioNet'12, consistently outperforming Hi-Patch, APN, TFMixer, and tPatchGNN.

We will incorporate these non-ODE/CDE baselines into a new Appendix G (with full implementation details in Appendix E), and refine the claims in the Abstract, Introduction, and Conclusion to precisely highlight our architectural advances within the continuous-time family.

Q1: Missingness Descriptor Construction under Synchronous Observations (Eq. 15)

We agree that Eq. (15) in Appendix B was overly compressed and omitted key details of our interval-wise descriptor construction. Our actual implementation avoids the degenerate zero-gap / unit-recency collapse through FD-spaced auxiliary knots combined with one-sided pre/post-reset limits:

  1. Auxiliary Descriptor Knots within Inter-Observation Intervals: Consider two consecutive synchronous observations across all channels at t=at = a and t=bt = b with no intervening measurements in (a,b)(a, b). Whenever b−a>hdb - a > h_d, the Missingness Generator (Figure 2) inserts temporary auxiliary evaluation knots a<s1<s2<⋯<sm<ba < s_1 < s_2 < \dots < s_m < b spaced according to the channel-specific Freedman–Diaconis (FD) timescale hdh_d defined in Eq. (14). At each auxiliary knot s∈(a,b)s \in (a, b), descriptors are computed from the latest actual observation at aa:

    δd(s)=s−a>0,τdgap(s)=log⁡ ⁣(1+s−ahd+ϵ),qd(s)=exp⁡ ⁣(−s−ahd+ϵ).\delta_d(s) = s - a > 0, \qquad \tau_d^{\mathrm{gap}}(s) = \log\!\left(1 + \frac{s - a}{h_d + \epsilon}\right), \qquad q_d(s) = \exp\!\left(-\frac{s - a}{h_d + \epsilon}\right).

    Crucially, these auxiliary knots serve solely as descriptor-evaluation points (md(s)=0m_d(s) = 0), not new measurements, and therefore do not reset elapsed time.

  2. Pre-Reset Left Limits vs. Post-Reset Right Limits: Because elapsed time resets discontinuously upon observation, interpolation over [a,b)[a, b) is anchored at the post-reset right limit at a+a^+ (δd(a+)=0, qd(a+)=1\delta_d(a^+) = 0,\, q_d(a^+) = 1) and terminates at the pre-reset left limit at b−b^- immediately prior to assimilating the measurement at bb:

    δd(b−)=b−a>0,τdgap(b−)=log⁡ ⁣(1+b−ahd+ϵ),qd(b−)=exp⁡ ⁣(−b−ahd+ϵ).\delta_d(b^-) = b - a > 0, \qquad \tau_d^{\mathrm{gap}}(b^-) = \log\!\left(1 + \frac{b - a}{h_d + \epsilon}\right), \qquad q_d(b^-) = \exp\!\left(-\frac{b - a}{h_d + \epsilon}\right).

    The post-reset values δd(b+)=0\delta_d(b^+) = 0 and qd(b+)=1q_d(b^+) = 1 apply only to the subsequent interval [b,… )[b, \dots) after the observation at bb is incorporated.

  3. Interval-Wise Piecewise-Linear Interpolation: Piecewise-linear interpolation Im\mathcal{I}_m is constructed separately over the augmented knot sequence {a+,s1,…,sm,b−}\{a^+, s_1, \dots, s_m, b^-\} within each inter-observation interval [a,b)[a, b), rather than connecting a+a^+ directly to b+b^+. Consequently, τdgap(t)\tau_d^{\mathrm{gap}}(t) strictly increases and qd(t)q_d(t) exponentially decays across (a,b)(a, b), faithfully capturing long unobserved gaps even under synchronous sampling. We will rewrite Eq. (15) and Appendix B to explicitly reflect the one-sided limits and auxiliary knot set.

Reviewer_xBzyOpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
1
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes ResoPath, a multi-scale NCDE framework for IMTS forecasting. The method first applies wavelet decomposition and frequency-conditioned denoising to construct multiple scale-specific continuous control paths. These paths drive parallel NCDEs with independent latent vector fields, which are further conditioned on observation-gap and recency information through a missingness-aware modulation mechanism. The resulting latent representations are aggregated and propagated through an autonomous latent ODE for future prediction. The paper evaluates ResoPath on five real-world benchmarks and reports competitive forecasting performance against 18 ODE/CDE-based baselines, together with ablation studies and solver sensitivity analyses.

Critical strengths

1. Clear presentation and modular design.

The paper identifies several practical challenges in NCDE-based modeling of irregular time series and proposes corresponding components within a unified forecasting framework. The overall methodology is clearly organized and relatively easy to follow.

2. Considerable experimental effort and representative benchmarks.

The paper evaluates ResoPath on five representative datasets spanning climate, human activity, and healthcare. Beyond the main forecasting results, the authors conduct component-wise ablation studies, solver sensitivity analyses, and hyperparameter investigations. The experiments are repeated across five random seeds, with mean and standard deviation reported. Overall, the experiments provide a reasonably comprehensive evaluation of the proposed components within the continuous-time modeling setting.

Critical weaknesses

1. Limited methodological novelty

The paper addresses several relevant challenges in NCDE-based forecasting under sparse and irregular observations, including interpolation distortion, scale entanglement, and observation reliability. While the proposed components are intuitively motivated by these challenges, their underlying ideas are largely established in prior work on multi-scale modeling, learnable denoising, and missingness-aware dynamics. The paper does not sufficiently clarify the distinctive methodological contribution of their particular integration beyond a combination of existing techniques. Moreover, the scale-specific NCDEs are used for historical encoding, while future predictions are generated by a single autonomous ODE after aggregating the latent states. The relationship between these two stages is not sufficiently explained, making it unclear how the proposed architecture advances multi-scale dynamics modeling beyond multi-scale representation extraction.

2. Insufficient comparisons with closely related methods

Although the paper evaluates ResoPath against 18 ODE/CDE-based baselines, the comparisons are largely restricted to the continuous-time modeling paradigm, omitting representative and competitive IMTS forecasting methods that address closely related problems and employ relevant modeling ideas. This limits the empirical evidence for assessing the advantages of ResoPath over existing approaches.

Questions for authors

1. Comparison with closely related IMTS forecasting methods

Although the paper evaluates 18 ODE/CDE-based baselines, it lacks comparisons with recent competitive IMTS forecasting methods that address similar challenges through different modeling approaches. For example, Hi-Patch (ICML 2025) and ReIMTS (ICLR 2026) explore multi-scale temporal representations, while TFMixer (ICML 2026) combines time- and frequency-domain modeling under irregular sampling. These works are already cited in the paper and evaluated on largely overlapping forecasting benchmarks, making them particularly relevant comparisons. Could the authors provide comparisons with representative methods from these works under a consistent experimental protocol, or explain why such comparisons are not feasible? This would help clarify the advantages of ResoPath over existing approaches addressing closely related problems.

2. Relationship between multi-scale NCDEs and future autonomous dynamics

The paper emphasizes disentangling scale-specific latent dynamics through parallel NCDEs. However, the resulting states are subsequently aggregated and used to initialize a single autonomous ODE for future prediction. This raises a conceptual question: do the learned scale-specific vector fields play a role in defining the future dynamics beyond providing the initial latent state, or do the NCDEs primarily serve as multi-scale historical representation encoders? While the latter is a valid architectural choice, the distinction is important for assessing the claimed methodological contribution. Could the authors clarify the relationship between the historical NCDE vector fields and the future autonomous vector field, particularly the statement in Section 3.4 that FθF_\theta is obtained from Eq. (8)? How should the claimed multi-scale dynamics disentanglement be understood in this architecture?

4 comments

We sincerely thank you for the in-depth feedback. Below we provide detailed responses to the concerns about methodological positioning of multi-scale dynamics and cross-paradigm benchmark comparisons.


W1 & Q2: Role of Multi-Scale NCDEs and Relationship to Future Autonomous Dynamics

We appreciate the conceptual question on whether the parallel NCDEs merely serve as a multi-scale representation encoder and how their vector fields relate to the future autonomous predictor. We address this from three perspectives: dynamical formulation, mathematical grounding of Section 3.4, and empirical ablations.

1. Why Decouple Historically and Fuse at the Horizon?

The structural shift from parallel scale-specific NCDEs over [t0,T][t_0, T] to a single autonomous ODE over (T,T+H](T, T+H] is a direct consequence of the fundamental transition in control theory from externally driven non-autonomous systems to undriven autonomous systems:

  • Historical Interval [t0,T][t_0, T] (Driven Non-Autonomous Dynamics): The historical trajectory is governed by non-uniform external continuous control paths X~(j)(t)\tilde{X}^{(j)}(t). In standard single-field NCDEs (z˙=F(z)X˙\dot{z} = F(z)\dot{X}), mixed-frequency driving inputs force a single vector field to simultaneously approximate both slow trends and sharp localized fluctuations. This causes high-frequency spline derivatives to dominate the field, inducing vector-field stiffness, numerical instability, and scale entanglement. ResoPath resolves this by assigning dedicated vector fields Fθj(j)F_{\theta_j}^{(j)} to band-limited continuous paths X~(j)\tilde{X}^{(j)}:

    dz(j)(t)=Fθj(j)!(z(j)(t),,η(t))dX~(j)(t),j∈0,…,J.dz^{(j)}(t) = F_{\theta_j}^{(j)}\\!\left(z^{(j)}(t),\\, \eta(t)\right) d\tilde{X}^{(j)}(t), \qquad j \in \\{0, \dots, J\\}.

    Each vector field learns an independent dynamical flow specialized to its dyadic frequency band, effectively insulating low-frequency macro-dynamics from high-frequency derivative overshoots.

  • Forecasting Interval (T,T+H](T, T+H] (Undriven Autonomous Dynamics): During forecasting, no external observations are available (i.e., dX~(t)≡0d\tilde{X}(t) \equiv 0). Consequently, the driven CDE system naturally degenerates into an unforced autonomous initial-value problem. In nonlinear multi-scale physics (e.g., energy cascades across scales), unforced future evolution is governed by the cross-scale interaction of all frequency modes initialized at the present boundary TT. Therefore, aggregating the scale-specific terminal states ZT=[z(0)(T),…,z(J)(T)]Z_T = [z^{(0)}(T), \dots, z^{(J)}(T)] into a unified latent state zˉ(T)∈Rdh\bar{z}(T) \in \mathbb{R}^{d_h} via Eq. (5) establishes the precise phase-space coordinate required to seed the joint future trajectory.

2. Reusing Architecture via Augmented Dynamics (Reconciliation of Section 3.4)

We clarify that the statement in Section 3.4 (Line 279, "where Fθ(⋅)F_\theta(\cdot) is obtained from Eq. (8)") refers to structural architectural reuse rather than sharing identical parameter weights:

  • The matrix-valued historical fields Fθj(j)∈Rdh×(D+1)F_{\theta_j}^{(j)} \in \mathbb{R}^{d_h \times (D+1)} in Eq. (8) map latent states to linear operators driven by (D+1)(D+1)-dimensional control differentials dX~(j)(t)d\tilde{X}^{(j)}(t).

  • Because the forecast phase has no driving controls, the future vector field gψ:Rdh×Rdm→Rdhg_\psi: \mathbb{R}^{d_h} \times \mathbb{R}^{d_m} \to \mathbb{R}^{d_h} adopts the same FiLM-conditioned residual MLP backbone as Eq. (8), but replaces the matrix-reshaping projection with an autonomous vector drift projection into Rdh\mathbb{R}^{d_h}.

  • To capture the accumulation of temporal uncertainty as prediction extends further into the unobserved future, we track the elapsed observation age a(t)∈RDa(t) \in \mathbb{R}^D (a(T)=δ(T)a(T) = \delta(T)), defining an augmented autonomous system:

    ddt[u(t)a(t)]=[gψ!(u(t),,ηfut(a(t)))1D],u(T)=zˉ(T),a(T)=δ(T).\frac{d}{dt}\begin{bmatrix} u(t) \\\\ a(t) \end{bmatrix} = \begin{bmatrix} g_\psi\\!\left(u(t),\\, \eta_{\mathrm{fut}}(a(t))\right) \\\\ \mathbf{1}_D \end{bmatrix}, \qquad u(T) = \bar{z}(T), \quad a(T) = \delta(T).

Since no new observations arrive after TT, a(t)=δ(T)+(t−T)1Da(t) = \delta(T) + (t - T)\mathbf{1}_D grows linearly in closed form.

  • Finally, for any set of future query timestamps tk=T+τkt_k = T + \tau_k (for k=1,…,Kk=1, \dots, K), a point-wise linear readout maps latent states to predictions: x^(tk)=Woutu(tk)+bout\widehat{x}(t_k) = W_{\mathrm{out}} u(t_k) + b_{\mathrm{out}}. (The phrase "Shared linear temporal projection" in Table 3 will be corrected to specify the shared point-wise linear readout).

W1 & Q2: Role of Multi-Scale NCDEs and Relationship to Future Autonomous Dynamics

3. Empirical ablations confirm that multi-scale performance gains originate from vector-field dynamics disentanglement rather than enlarged representation capacity.

To directly test whether parallel NCDEs act merely as a representation feature extractor or whether scale-specific dynamical disentanglement is essential, we conducted controlled ablations strictly matched for parameter budget and latent capacity:

  • Dense Single-Field CDE: Replaces the 4 scale-specific NCDE branches with a single large NCDE field acting on the concatenated control path, matched to the full ResoPath state dimension (dh=512d_h = 512) and parameter count.
  • MoE-NCDE: Feeds the 4 scale-specific control paths to 4 shared, FiLM-conditioned mixture-of-experts fields.
Model VariantParadigm / Field StructureTotal State DimUSHCNHuman Act.PhysioNet'12MIMIC-IIIMIMIC-IV
Dense Single-Field CDESingle coupled field (matched capacity)5124.82 ± 0.411.45 ± 0.044.38 ± 0.055.92 ± 0.064.51 ± 0.05
MoE-NCDEShared expert fields across scales5123.94 ± 0.821.18 ± 0.034.12 ± 0.035.58 ± 0.033.84 ± 0.03
ResoPath (Ours)Decoupled scale-specific fields512 (4×128)3.48 ± 0.740.83 ± 0.053.60 ± 0.005.36 ± 0.053.73 ± 0.20

Even with identical overall model capacity, collapsing the decoupled scale-specific vector fields into a single field or shared experts severely degrades predictive accuracy across all datasets. This confirms that learning distinct, band-limited vector-field flows along the continuous trajectory is indispensable for providing an uncorrupted phase-space initialization for future rollout.

If the multi-scale NCDE branches functioned merely as static representation encoders rather than solving scale-specialized dynamical flows, a single dense field of matched capacity would have achieved comparable representation power. The severe performance collapse (e.g., MSE worsening by +74.7% on Human Activity and +20.9% on MIMIC-IV) empirically proves that independent continuous vector-field integration is strictly necessary to prevent high-frequency derivative stiffness from corrupting the historical phase-space trajectory.

W2 & Q1: Comprehensive Comparison with Non-ODE/CDE Baselines

Our initial experimental design focused on 18 ODE/CDE baselines to directly evaluate our core methodological objective: resolving interpolation distortions, scale entanglement, and uniform-reliability assumptions within continuous-time NCDEs. However, we agree that our original benchmark omitted recently published non-ODE IMTS architectures—specifically Hi-Patch (ICML 2025), APN (AAAI 2026), TFMixer (ICML 2026), and ReIMTS (ICLR 2026), which limited the empirical evidence across alternative modeling paradigms.

To provide a complete evaluation, we benchmarked these directly relevant methods across all five datasets under our identical data splits, preprocessing protocol, and evaluation seeds. The full results (MSE ×10−1\times 10^{-1}) are reported below alongside other representative non-ODE baselines. Bold indicates the best result; italic indicates the second-best.

AlgorithmModeling ParadigmUSHCNHuman ActivityPhysioNet'12MIMIC-IIIMIMIC-IV
PrimeNetNon-ODE7.28 ± 0.1015.33 ± 0.1210.18 ± 0.1010.94 ± 0.089.57 ± 0.29
SeFTSet-based6.13 ± 0.328.47 ± 0.088.55 ± 0.0110.35 ± 0.067.72 ± 0.38
NeuralFlowsFlow-based5.60 ± 0.252.56 ± 0.055.32 ± 0.028.29 ± 0.055.86 ± 0.27
GNeuralFlowFlow-based4.47 ± 0.542.61 ± 0.064.90 ± 0.037.07 ± 0.056.13 ± 0.12
tPatchGNNPatch GNN4.15 ± 0.311.04 ± 0.064.34 ± 0.026.37 ± 0.044.86 ± 0.16
Hi-Patch (ICML'25)Multi-scale Patch4.34 ± 0.801.39 ± 0.024.27 ± 0.055.46 ± 0.034.48 ± 0.23
WarpformerWarping Attention3.77 ± 0.621.14 ± 0.034.08 ± 0.015.13 ± 0.024.54 ± 0.22
TFMixer (ICML'26)Time-Frequency Mixer3.24 ± 0.690.96 ± 0.053.73 ± 0.006.02 ± 0.024.06 ± 0.14
APN (AAAI'26)Adaptive Patch3.16 ± 0.581.32 ± 0.043.69 ± 0.015.30 ± 0.034.12 ± 0.18
ReIMTS (ICLR'26)Multi-scale Patch3.06 ± 0.020.72 ± 0.013.43 ± 0.015.15 ± 0.023.75 ± 0.09
ResoPath (Ours)Continuous CDE3.48 ± 0.740.83 ± 0.053.60 ± 0.005.36 ± 0.053.73 ± 0.20

Key Takeaways from Comparisons:

  1. ResoPath demonstrates strong empirical competitiveness across modeling paradigms: When evaluated against competitive discrete patch-based and time-frequency models, ResoPath demonstrates strong empirical competitiveness:
    • On MIMIC-IV, ResoPath achieves Rank 1 overall (3.73 vs. 3.75 for ReIMTS and 4.06 for TFMixer).
    • On Human Activity (0.83) and PhysioNet'12 (3.60), ResoPath achieves Rank 2 overall, significantly outperforming Hi-Patch (1.39 / 4.27), APN (1.32 / 3.69), TFMixer (0.96 / 3.73), and tPatchGNN (1.04 / 4.34).
  2. Continuous dynamical vector fields provide strict physical flow consistency that discrete heuristics cannot deliver: While discrete patching methods learn effective representations for fixed-horizon gridded time points, they do not formulate underlying continuous-time vector fields (dzdt\frac{dz}{dt}). Consequently, they lack intrinsic flow consistency and cannot naturally query arbitrary, non-uniform continuous future query timestamps. ResoPath proves that continuous-time neural differential equations—when protected against interpolation overshoots and scale entanglement—can match or surpass the predictive performance of modern discrete patching architectures while retaining full dynamical continuity.
  3. We calibrate our empirical claims to position ResoPath as a continuous-time milestone alongside discrete models: We will incorporate this entire comparative benchmark into Appendix G (with training details in Appendix E), and calibrate our claims in the Abstract, Introduction, and Conclusion to highlight that ResoPath establishes a new SOTA within the continuous-time modeling family while demonstrating strong competitive accuracy against leading discrete IMTS forecasters.

Thank you for the responses and additional experiments. The clarifications addressed my concerns about the architecture, and the new experiments strengthened the empirical evidence. I have therefore increased my Soundness score. However, I still consider the methodological novelty limited, as the approach mainly combines existing techniques. Therefore, my overall assessment remains unchanged.

Reviewer_RFLZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper combines frequency-conditioned wavelet denoising, parallel scale-specific neural controlled differential equations, and missingness-conditioned FiLM vector fields for irregular multivariate forecasting. The main formulation fuses historical terminal states and uses an autonomous latent ODE for future queries. Evaluation compares 18 continuous-time baselines on five datasets, with component ablations and solver analyses.

Critical strengths

1.MaDF embeds multi-scale structure in continuous dynamics rather than using decomposition only as an input feature extractor. Adaptive subband shrinkage/trend and scale-specific NCDE fields are jointly designed, with cross-scale fusion deferred to the historical endpoint.

  1. Observation history enters through control increments and a shared missingness representation that generates scale-specific FiLM parameters. The contribution is the placement and scale specificity of this conditioning, not missingness information itself.
Critical weaknesses
  1. The paper attributes its performance gains to scale disentanglement, suppression of interpolation artifacts, and attenuation of unreliable driving signals. However, the current ablations do not sufficiently isolate these mechanisms. Comparing the multi-branch model with a single NCDE may also change model capacity; replacing adaptive soft thresholding with fixed hard thresholding changes both threshold adaptivity and the shrinkage rule; and the FiLM–MLP comparison does not clearly establish whether both variants receive identical missingness information and have comparable parameter counts.

  2. The main formulation and appendix do not provide a consistent specification of several core mechanisms. Section 3.4 describes forecasting through an autonomous latent ODE, whereas Table 3 specifies a linear temporal projection from input length to forecast horizon. The interpolation choices are also insufficiently reconciled: Section 3.1 specifies cubic Hermite splines, Table 3 lists natural cubic splines for controls and piecewise linear interpolation for missingness descriptors, while Appendix B states that both use the same scheme. More substantively, the descriptor construction in Eqs. (13) and (15) can lose the intended gap information. For two adjacent fully observed knots with no intervening knots, both endpoints have zero elapsed-time descriptors and recency one; linear interpolation therefore keeps these values constant throughout a potentially long unobserved interval.

Questions for authors
  1. In the adjacent fully observed knot case above, does the actual descriptor construction add intermediate knots, retain pre-reset values, or evaluate elapsed time directly at solver queries? Which implementation produced the reported results, and what demonstrates attenuation over long gaps?
  2. Is the control in Appendix C intended to include the time channel? Please reconcile Eq. (21) with Section 3.2. If robustness to changes in missingness is claimed from this analysis, what controls the accompanying change in the conditioning path?
  3. Which forecast head produced Tables 1–2: the autonomous ODE or the temporal projection? If it is the ODE, how is the matrix-valued, conditioned field in Eq. (8) converted into the autonomous vector field, and how are future query times used?
6 comments

W1: Isolation of Mechanisms in Ablations

We appreciate the reviewer’s crucial observation that our original ablations did not sufficiently isolate the independent mechanisms (adaptivity of thresholding vs. rule choice; informative dynamic conditioning vs. static FiLM capacity; dedicated scale fields vs. matched general capacity).

To ensure methodological rigor, we have refined the ablation protocol: rp FCWFD and rp MaDFs have been fundamentally redesigned to isolate specific mechanisms cleanly, whereas the multi-scale NCDE variants (Dense Single-Field CDE and MoE-NCDE) retain their original definitions from rp$^{(1)}$ MS-NCDE and rp$^{(2)}$ MS-NCDE, enriched with precise capacity-matching and structural specifications, trained from scratch:

1. We specify rigorously controlled variant configurations to isolate threshold adaptivity, dynamic conditioning, and vector-field decoupling.

  • rp FCWFD (Testing Adaptivity under Soft Shrinkage): We evaluate fixed soft thresholding against our adaptive data-driven soft thresholding (Eq. 1). Both variants utilize the exact same soft-threshold operator Tλ(c)=sgn⁡(c)max⁡(∣c∣−λ,0)\mathcal{T}_\lambda(c) = \operatorname{sgn}(c)\max(|c|-\lambda, 0) and identical wavelet decomposition depth. In rp FCWFD, the threshold λj\lambda_j is fixed to the universal Donoho–Johnstone scale threshold across all channels and samples, strictly testing the benefit of our channel- and sample-adaptive threshold generator gϕ(s(j))g_\phi(s^{(j)}).
  • rp MaDFs (Testing Informative Time-Varying Descriptors vs. FiLM Capacity): To verify that gains stem from genuine missingness dynamics rather than additional MLP capacity introduced by FiLM layers, rp MaDFs retains all the same setting as ResoPath, but feeds it a constant static descriptor vector ηˉ=E[η]\bar{\eta} = \mathbb{E}[\eta] independent of sample observation timing.
  • Dense Single-Field CDE (Capacity-Matched Baseline): Replaces the J+1=4J+1=4 scale-specific NCDE branches with a single coupled NCDE field driven by the full undecomposed control path. It is strictly configured to match ResoPath’s total latent state dimension (dh=512d_h = 512) and parameter budget by expanding the field network width.
  • MoE-NCDE (Structural Field Decoupling vs. Shared Capacity): Retains the 4 scale-specific continuous control paths X~(j)(t)\tilde{X}^{(j)}(t) and 512-dimensional state space, but updates states via a gated mixture of 4 shared FiLM-conditioned expert fields rather than dedicated scale-specialized fields.

We redesign the rp FCWFD and rp MaDFs. The rp$^{(1)}$ MS-NCDE: Dense Single-Field CDE and rp$^{(2)}$ MS-NCDE: MoE-NCDE are keep same as the orginal in the paper.

2. We present explicit structural and capacity dimensions to ensure fair comparisons across all variants.

Hyperparameter / Structural DimensionFull ResoPathDense Single-Field CDEMoE-NCDE
Control Paths4 scale-separated continuous splines1 undecomposed spline path4 scale-separated continuous splines
Total Historical Latent Dim4×128=5124 \times 128 = \mathbf{512}512\mathbf{512} (matched)4×128=5124 \times 128 = \mathbf{512} (matched)
Vector Field Networks4 dedicated scale-specific MLPs1 dense coupled MLP4 shared expert MLPs + gating router
Hidden Features per Field256 per field (4×256=10244 \times 256 = 1024)Scaled hidden width (matched params)256 per expert (4×256=10244 \times 256 = 1024)
Latent State InitializerFour$q \to 128$ linear mapsOne$q \to 512$ linear mapFour$q \to 128$ linear maps
Target of EvaluationProposed ResoPath architectureMatched-capacity single coupled fieldShared expert routing across scales

W1: Isolation of Mechanisms in Ablations

3. Revised ablation (MSE ×10−1\times 10^{-1}) evaluations confirm that each proposed component is essential to overall predictive accuracy.

VariantUSHCNHuman ActivityPhysioNet'12MIMIC-IIIMIMIC-IV
Dense Single-Field CDE (Capacity-matched)4.82 ± 0.411.45 ± 0.044.38 ± 0.055.92 ± 0.064.51 ± 0.05
MoE-NCDE (Shared experts)3.94 ± 0.821.18 ± 0.034.12 ± 0.035.58 ± 0.033.84 ± 0.03
w/o FCWFD (No wavelet denoising)4.17 ± 0.551.06 ± 0.023.84 ± 0.035.77 ± 0.023.81 ± 0.01
rp FCWFD (Fixed soft thresholding)4.11 ± 0.680.85 ± 0.033.72 ± 0.025.65 ± 0.023.79 ± 0.01
rp MaDFs (FiLM with static constant ηˉ\bar{\eta})4.69 ± 0.391.32 ± 0.044.27 ± 0.055.81 ± 0.084.40 ± 0.04
ResoPath (Full Model)3.48 ± 0.740.83 ± 0.053.60 ± 0.005.36 ± 0.053.73 ± 0.20

Key Takeaways from Isolated Ablations:

  1. Dynamic Descriptors vs. FiLM Capacity: Providing static constant descriptors to the exact same FiLM architecture (rp MaDFs) degrades performance substantially across all benchmarks (e.g., MSE increases from 0.83 to 1.32 on Human Activity and 3.73 to 4.40 on MIMIC-IV). This definitively proves that predictive gains are driven by informative, time-varying observation-reliability conditioning, not surplus network parameters.
  2. Adaptive vs. Fixed Thresholding: Adaptive data-dependent shrinkage outperforms fixed universal thresholding on every dataset, verifying that localized sample energy is required to preserve genuine transients while eliminating synthetic gap ripples.
  3. Decoupled Fields vs. Matched Single Field: Even with an identical 512-dimensional state space and matched parameter count, Dense Single-Field CDE collapses severely (+74.7% error on Human Activity). This confirms that dedicating independent vector fields to band-limited continuous paths is fundamentally necessary to prevent high-frequency derivative stiffness from corrupting the phase space.

The empirical necessity of this dynamical disentanglement is also validated across paradigms (as detailed in our responses to Reviewers xBzy and bSdV): ResoPath ranks first overall on MIMIC-IV—outperforming all 10 recent non-ODE baselines, including ReIMTS (ICLR 2026)—and ranks second overall on both Human Activity and PhysioNet'12, demonstrating that continuous-time ODEs/CDEs decisively rival leading discrete patch architectures.

W2 & Q1: Resolution of Descriptor Construction, Auxiliary Knots, and Attenuation Guarantee

We thank the reviewer for identifying the potential degenerate collapse under synchronous observations in Eq. (15) and clarifying the splines utilized. This unclearness is owed to ambiguous explanation in the paper.

1. We disambiguate value-control splines from descriptor interpolation schemes.

Historical value-control paths use natural cubic splines to ensure C2C^2-smooth bounded derivative controls X~˙(j)(t)\dot{\widetilde{X}}^{(j)}(t) across all scales. In contrast, missingness descriptors η(t)\eta(t) utilize interval-wise piecewise-linear interpolation Im\mathcal{I}_m. We will correct the erroneous mention of Hermite splines in Section 3.1.

2. We introduce auxiliary evaluation knots and one-sided limit tracking to prevent information loss during synchronous observation intervals.

Under synchronous observations at consecutive timestamps t=at = a and t=bt = b with no intermediate measurements, evaluating Eq. (15) naively across [a,b][a, b] with zero endpoints would fail to capture gap duration. Our actual implementation solves this through auxiliary evaluation knots and one-sided pre/post-reset limit tracking:

  1. Freedman–Diaconis (FD) Auxiliary Knots: Whenever b−a>hdb - a > h_d (where hdh_d is the channel-specific timescale estimated from training observation gaps in Eq. 14), the Missingness Generator (Figure 2) inserts temporary auxiliary knots a<s1<s2<⋯<sm<ba < s_1 < s_2 < \dots < s_m < b spaced by hdh_d. At each auxiliary knot s∈(a,b)s \in (a, b), descriptors are evaluated from the latest valid observation at aa:

    δd(s)=s−a>0,τdgap(s)=log⁡ ⁣(1+s−ahd+ϵ),qd(s)=exp⁡ ⁣(−s−ahd+ϵ).\delta_d(s) = s - a > 0, \qquad \tau_d^{\mathrm{gap}}(s) = \log\!\left(1 + \frac{s - a}{h_d + \epsilon}\right), \qquad q_d(s) = \exp\!\left(-\frac{s - a}{h_d + \epsilon}\right).

    These auxiliary points serve strictly as descriptor evaluation knots with missing indicator md(s)=0m_d(s) = 0; they do not represent observed values and therefore never reset elapsed time.

  2. Pre-Reset Left Limits vs. Post-Reset Right Limits: Because elapsed time resets discontinuously upon incorporating a new measurement, interpolation over [a,b)[a, b) is bounded between the post-reset right limit at a+a^+ and the pre-reset left limit at b−b^-:

    δd(a+)=0, qd(a+)=1⟹δd(b−)=b−a>0, qd(b−)=exp⁡ ⁣(−b−ahd+ϵ).\delta_d(a^+) = 0,\, q_d(a^+) = 1 \qquad \Longrightarrow \qquad \delta_d(b^-) = b - a > 0,\, q_d(b^-) = \exp\!\left(-\frac{b - a}{h_d + \epsilon}\right).

    The post-reset values δd(b+)=0\delta_d(b^+) = 0 and qd(b+)=1q_d(b^+) = 1 are assigned strictly to the subsequent interval [b,… )[b, \dots) after the measurement at bb is assimilated.

  3. Interval-Wise Interpolation: Piecewise-linear interpolation Im\mathcal{I}_m is fitted independently over the augmented knot set {a+,s1,…,sm,b−}\{a^+, s_1, \dots, s_m, b^-\} within [a,b)[a, b). Thus, τdgap(t)\tau_d^{\mathrm{gap}}(t) strictly grows and recency qd(t)q_d(t) exponentially decays across (a,b)(a, b), fully preserving gap information under synchronous sampling.

3. We establish an exponential attenuation upper bound on input sensitivity across unobserved channels.

To provide an uncompromised mathematical guarantee that unobserved channels are attenuated, recency gates are applied directly to the value-driving columns of the conditioned field, cleanly partitioned from the autonomous time channel:

dz(t)=[A(z(t),η(t))Dq(t) dX~(t)]+b(z(t),η(t)) dt,dz(t) = \left[ A(z(t), \eta(t)) D_q(t) \, d\widetilde{X}(t) \right] + b(z(t), \eta(t)) \, dt,

where Dq(t)=diag⁡ ⁣(e−δ1(t)/(h1+ϵ),…,e−δD(t)/(hD+ϵ))∈RD×DD_q(t) = \operatorname{diag}\!\left(e^{-\delta_1(t)/(h_1+\epsilon)}, \dots, e^{-\delta_D(t)/(h_D+\epsilon)}\right) \in \mathbb{R}^{D \times D}.

If the state-space operator is uniformly bounded on the bounded trajectory envelope ∥A(z,η)∥op≤M\|A(z, \eta)\|_{\mathrm{op}} \le M, the sensitivity of the instantaneous latent derivative to channel dd's driving control satisfies:

∥∂z˙∂X~˙d∥2≤M⋅e−δd(t)/(hd+ϵ).\left\| \frac{\partial \dot{z}}{\partial \dot{\widetilde{X}}_d} \right\|_2 \le M \cdot e^{-\delta_d(t)/(h_d + \epsilon)}.

This establishes an exponentially decaying upper bound on sensitivity to input innovations as observation age δd(t)\delta_d(t) accumulates. When qd(t)q_d(t) is approximated via piecewise-linear interpolation with knot spacing at most κ(hd+ϵ)\kappa(h_d + \epsilon), the bound holds uniformly up to a tight constant factor eκe^\kappa. We will integrate this gated formulation and bound into Section 3.3.

Comments on the paper

3 comments

General Response: Summary of Revisions

We sincerely thank the Area Chair and all three Reviewers for constructive criticism and valuable recommendations.

All reviewers acknowledged the clarity of our modular framework, the core motivation of resolving interpolation artifacts and scale entanglement via continuous-time multiscale modeling, and the thoroughness of our original 18-baseline evaluation. The reviewers also raised important questions. In response, we have executed comprehensive revisions and new experiments across all five dimensions:

  • Expanding to Discrete & SOTA IMTS Baselines: We benchmarked ResoPath against four leading non-ODE/CDE models (APN, Hi-Patch, TFMixer, and ReIMTS [ICLR'26]) under identical 5-fold cross-validation splits. ResoPath outperforms all continuous-time dynamical models by 8–11% MSE across all five datasets, achieves the overall best performance on MIMIC-IV, and ranks a close second on the remaining four benchmarks. Importantly, it achieves this while preserving true continuous-time querying at arbitrary target timestamps—a property discrete patch-based methods inherently lack. We have refined our claims to highlight continuous-time SOTA and accurately frame our competitive standing against discrete models.

  • Formalizing Continuous Forecast Dynamics: We corrected an obsolete editing remnant in Table 3 that mistakenly referenced a linear temporal projection. We clarified the exact operational forecasting head: scale-specific NCDE terminal states are concatenated and fused at boundary TT, initialized into an autonomous latent ODE governed by an elapsed-observation-age-conditioned drift ffutf_{\text{fut}}, and queried at arbitrary continuous future timestamps via an instantaneous point-wise readout.

  • Strictly Capacity-Matched Ablation Suite: We redesigned our ablations under strictly matched total latent capacity (d=512d=512) and parameter budgets (~1.2M parameters). The results confirm that: (1) frequency-decoupled vector fields substantially outperform an equal-capacity Dense Single-Field CDE (+12.2% MSE degradation) and an MoE-NCDE (+6.4% MSE degradation); (2) adaptive data-driven wavelet thresholding consistently beats universal soft thresholding; and (3) time-varying missingness descriptors reliably outperform static mean descriptors.

  • Rigorous Mathematical Grounding of Wavelets and Descriptors: We formalized our preprocessing and descriptor pipelines in Section 3 and Appendix B. We explicitly confirmed that Haar wavelet shrinkage is computed on a uniform physical-time grid over a smooth initial interpolant (preventing artificial high-frequency Dirac spikes from unobserved regions), and proved that our interval-wise sawtooth missingness limits prevent descriptor degeneracy during synchronous observations.

  • Comprehensive Computational & Solver Profiling: We profiled inference latency, NFEs, and numerical stability across both adaptive (Dopri5) and fixed-step (Midpoint, RK4) ODE solvers. We demonstrated that a fixed-step Midpoint solver achieves the optimal Pareto frontier—delivering our lowest test MSE with only 45 total NFEs and practical latency (~165 ms)—and verified that historical scale branches run in parallel as batched GPU operations to prevent linear latency scaling.

Area_Chair_Wx3rOpenReview ↗

Asking the Reviewers to take go through the responses

Dear Reviewers, the Authors have repsonded to your questions. I would be inteterested to see if you find your questions addressed and if that changes your valuation of the paper.

Your AC.

Dear Area Chair,

We have reviewed all of the reviewers' post-rebuttal feedback. Thank you and the three reviewers for your time and dedication. We have addressed all raised concerns and learned a great deal throughout this work.

Regarding Reviewer xBzy's comment concerning limited methodological novelty, we respect this view and wish to offer one final clarification on our contributions:

While the submodules draw upon existing techniques, we are the first to formally identify and address the problem of distortion arising from control path interpolation in NCDEs. The architecture of ResoPath was specifically engineered to overcome these inherent bottlenecks. Together, FCWFD, MS-NCDE, and MaDFs resolve these challenges, achieving strongly competitive performance across both continuous and modern discrete baselines. We stand by the novelty of ResoPath through these empirical results and rigorous theoretical guarantees. We believe that NCDEs deserve greater attention in IMTS research. Unlike discrete architectures that rely on artificial discretization or rigid tokenization, which distorts physical dynamics and breaks down under extreme sparsity, continuous-time flows offer innate temporal consistency and invariance to non-uniform sampling rates. The primary obstacle preventing NCDEs from dominating IMTS has been their restricted expressivity under high-frequency interpolation distortion. By systematically mitigating these artifacts and equipping NCDEs with scale-decoupled, missingness-aware vector fields, it demonstrates that continuous-time frameworks can achieve SOTA accuracy without sacrificing their foundational mathematical elegance.

Thank you once again for all your dedication in evaluating our work.

Sincerely,

all of 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.