acceptodds
Under review as a conference paper at ICLR 2027

Laminar Flow Matching: Geometric Multiscale Transport for Time-Series Forecasting

Abstract

Real-world time series exhibit stochastic dynamics across multiple temporal scales, posing a challenge for continuous-time generative forecasting. In standard Flow Matching (FM), a single vector field simultaneously models coarse global evolution and fine-scale fluctuations, which increases transport curvature and leads to inefficient dynamics. We propose Laminar Flow Matching (LFM), a geometry-aware framework that decouples multiscale probability transport into two hierarchically conditioned flows. The Macro-Flow learns a low-curvature global transport path in latent space, while the Micro-Flow models conditionally centered stochastic fluctuations in data space. This hierarchical factorization reduces cross-scale interference and improves transport smoothness and numerical efficiency. Extensive experiments on six real-world datasets show that LFM achieves state-of-the-art performance, reducing MSE by up to 15% and CRPS by 5%, while requiring fewer integration steps than existing generative baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Flow Matching, Time-Series Forecasting, Probabilistic Forecasting, Multiscale Transport, Optimal Transport

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

This paper proposes Laminar Flow Matching (LFM), a hierarchical generative framework designed for multivariate time-series forecasting. The authors argue that directly applying standard conditional flow matching to time series forces a single velocity field to simultaneously capture slow, context-driven macro-trends and fast, stochastic micro-fluctuations. In their view, this entanglement creates curved transport trajectories and stiff ODE dynamics, which require many function evaluations (NFEs) to integrate accurately. To resolve this, LFM partitions the time series into temporal blocks and decouples the probability transport into two sequential stages: an outer macro-flow operating in a block-wise latent space along an integration axis τ∈[0,1]\tau \in [0, 1] to capture low-frequency trends, and an inner micro-flow operating in the raw data space along t∈[0,1]t \in [0, 1] to model intra-block high-frequency residuals conditioned on the macro prediction. The authors provide a geometric motivation for scale conflict in single-field transport (Proposition 1) and evaluate LFM across six standard long-term forecasting benchmarks as well as the high-volatility AEMO electricity demand dataset. The empirical results show competitive performance on both point metrics (MSE, MAE) and probabilistic metrics (CRPS) using a total of 15 ODE Euler steps.

Critical strengths

I really like the intuitive design of decoupling time-series transport into a latent block-level macro-flow and a data-space residual micro-flow. In my experience with generative time-series models, trying to model global trend shifts and local noise in a single stage often leads to over-smoothed forecasts or unstable sampling at low step counts.

The architectural choices in Sections 4.2 and 4.3 are thoughtfully engineered to make this two-stage process work in practice. In particular, constructing the dynamic base x~\tilde{x} via a single-step extrapolation of the macro-flow and applying the stop-gradient operator sg(⋅)\text{sg}(\cdot) in Equation 17, alongside injecting Gaussian perturbations into the macro latent condition ϕ~\tilde{\phi} in Equation 18, is a clean and effective way to mitigate exposure bias. The ablation study in Table 3 convincingly demonstrates that both the anchoring loss and the dynamic base are essential to the final performance.

Another strong point of this work is how well it bridges the gap between deterministic accuracy and probabilistic calibration while keeping inference latency low. Historically, diffusion and flow-based forecasters like CSDI, Diffusion-TS, and TSFlow have struggled to match strong discriminative baselines on standard MSE and MAE metrics.

By initializing the prior with a lightweight predictor Fϕ(xh)\mathcal{F}_\phi(x^h) and straightening the transport paths through hierarchical decomposition, LFM achieves state-of-the-art MSE and CRPS across most benchmarks in Tables 1 and 2 with just P+K=15P + K = 15 Euler steps. The power spectral density comparison in Figure 7 and the trajectory straightness metrics in Table 12 also provide reassuring empirical evidence that the proposed decomposition preserves high-frequency energy without sacrificing numerical stability.

Critical weaknesses

My primary technical concern lies in how the inner micro-flow handles stochasticity during inference. In Section 4.3, the inner micro-flow is described as the component responsible for capturing stochastic intra-block fluctuations.

However, looking closely at Equations 17 to 22 and Algorithm 2, once the macro-flow finishes integrating to the terminal latent state ϕ^1(s)\hat{\phi}_1^{(s)}, the micro-flow is initialized deterministically at ψ0(s)=B(ϕ^1(s))\psi_0^{(s)} = \mathcal{B}(\hat{\phi}_1^{(s)}) with condition ϕ~(s)=ϕ^1(s)\tilde{\phi}^{(s)} = \hat{\phi}_1^{(s)}, and then integrated via a deterministic Euler ODE without injecting any independent data-space noise at t=0t = 0.

This implies that, the entire inner micro-flow is a purely deterministic mapping. Consequently, all sample diversity across the SS trajectories {x^f,(s)}s=1S\{\hat{x}^{f,(s)}\}_{s=1}^S originates solely from the initial prior z0(s)∼N(μ,σ2I)z_0^{(s)} \sim \mathcal{N}(\mu, \sigma^2 I), which passes through the block-wise encoder E\mathcal{E} into a lower-dimensional macro latent space RN×dl\mathbb{R}^{N \times d_l}.

Given that the authors note around line 233 that E\mathcal{E} attenuates micro-scale variability, it seems contradictory to claim that the micro-flow models stochastic intra-block distributions when no fresh randomness enters at the micro stage. Furthermore, during training, because multiple ground-truth high-frequency realizations xfx^f can share the same macro anchor x~\tilde{x}, the velocity field vθ(ψ0,0,ϕ~)v_\theta(\psi_0, 0, \tilde{\phi}) at t=0t = 0 is effectively forced to regress toward the conditional expectation E[xf−x~∣ϕ~]\mathbb{E}[x^f - \tilde{x} \mid \tilde{\phi}], rather than transporting an intra-block noise distribution to the true residual distribution.

I also have some reservations regarding the theoretical formulation in Proposition 1 and a few ambiguities in the experimental evaluation protocol. In Section 3.3, Proposition 1 models the trajectory as γ(s)=γ(L)(s)+asin⁡(ωs)e\gamma(s) = \gamma^{(L)}(s) + a \sin(\omega s) e to show that high-frequency oscillations inflate the spatial Jacobian norm ∥∇zv∥op\lVert \nabla_z v \rVert_{\text{op}}.

However, s∈[0,1]s \in [0, 1] here is the flow-matching transport time, not the physical time axis of the forecasting horizon. Under the linear optimal transport conditional path defined in Equation 3, the conditional trajectory with respect to the transport parameter ss is strictly a straight line with zero second derivative.

High-frequency temporal variations exist across the TfT_f coordinates of the state vector z∈RTf×Dz \in \mathbb{R}^{T_f \times D}, not as sinusoidal oscillations along the ODE integration axis ss. Assuming a sin⁡(ωs)\sin(\omega s) term along the transport parameter essentially builds the conclusion into the premise, rather than showing how multi-frequency structure along the temporal horizon induces curvature in the marginal velocity field over ss.

On the empirical side, there are two evaluation details that make a direct comparison somewhat difficult to interpret. First, in Tables 1, 2, and 13 through 15, the authors report CRPS values for purely deterministic baselines such as DLinear, PatchTST, iTransformer, and TimeXer.

Mathematically, the CRPS of a deterministic point forecast reduces strictly to its MAE, yet the reported CRPS numbers for these baselines are substantially different from (and much lower than) their corresponding MAE values. It remains unclear how CRPS was estimated for non-probabilistic models.

Second, Section 5.1 mentions that LFM's point forecast metrics are obtained by setting σ=0\sigma = 0 during inference, whereas probabilistic metrics use σ>0\sigma > 0 with S=100S = 100 samples. Meanwhile, generative baselines like CSDI and TSFlow typically compute MSE using the empirical mean of their sampled trajectories starting from a standard Gaussian prior, without having access to a pretrained DLinear prior Fϕ(xh)\mathcal{F}_\phi(x^h). This makes it hard to discern how much of LFM's MSE advantage stems from the hierarchical flow architecture versus the deterministic prior initialization and the σ=0\sigma = 0 inference branch.

Questions for authors

I would appreciate it if the authors could clarify the following points in their response:

  1. Could you clarify the source of intra-block stochasticity in the inner micro-flow during inference? Since Algorithm 2 initializes ψ0(s)=B(ϕ^1(s))\psi_0^{(s)} = \mathcal{B}(\hat{\phi}_1^{(s)}) and integrates vθv_\theta deterministically without adding data-space noise at t=0t = 0, doesn't the diversity of the generated distribution remain bounded by the compressed macro latent representation ϕ^1(s)\hat{\phi}_1^{(s)}? Did you experiment with injecting local Gaussian noise at ψ0(s)\psi_0^{(s)} or keeping the perturbation on ϕ~(s)\tilde{\phi}^{(s)} at test time?

  2. In Algorithm 1 (line 7), the dynamic base during training is formed via a single-step extrapolation ϕ^1=ϕτ+(1−τ)uθ(ϕτ,τ,c)\hat{\phi}_1 = \phi_\tau + (1 - \tau) u_\theta(\phi_\tau, \tau, c), where ϕτ=(1−τ)ϕ0+τϕ1\phi_\tau = (1 - \tau)\phi_0 + \tau \phi_1 contains the ground-truth future encoding ϕ1\phi_1. When τ\tau is close to 11, ϕ^1\hat{\phi}_1 is nearly identical to ϕ1\phi_1, whereas at inference time the model must integrate from τ=0\tau = 0. Why not extrapolate directly from τ=0\tau = 0 during training to avoid leaking ground-truth macro information into the micro-flow's base?

  3. How exactly is CRPS computed for the deterministic baselines (Autoformer, DLinear, PatchTST, iTransformer, TimeXer) in Tables 1, 2, and 13 to 15, given that their reported CRPS values do not equal their MAE values?

  4. To ensure an apples-to-apples comparison with other generative models, what are the MSE and MAE of LFM when evaluated using the sample mean of the S=100S = 100 probabilistic trajectories (with σ>0\sigma > 0) rather than the separate σ=0\sigma = 0 deterministic pass? Additionally, how do TSFlow or CGFM perform if they are also initialized with the same informative prior N(Fϕ(xh),σ2I)\mathcal{N}(\mathcal{F}_\phi(x^h), \sigma^2 I)?

As a minor suggestion for future revisions, I would encourage the authors to reframe Proposition 1 around the curvature of the marginal velocity field induced by multi-modal or multi-scale target distributions under linear conditional paths, rather than assuming explicit sin⁡(ωs)\sin(\omega s) oscillations along the transport time parameter ss.

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

The work addresses multi-scale challenges for dynamical systems, where capturing the global evolution (macro scale) and fine-scale fluctuations (high-frequency information) of the systems can be hard. The authors argue that existing Flow Matching methods present a geometric mismatch, since the constructed probability paths are straight and smooth (by linear interpolation between gaussians). In addition, a single/global velocity vector field induces scale conflicts in leaning the mutiscale dynamics.

For this reason, the authors adress these problems by proposing a solution that learns macro and micro scale dynamics. The macro scale evolution of the system is modelled by a learned latent dynamical system (outer marco-flow), which captures the smoothed and overall dynamics. Later on, this latent information is used by the "Inner Micro Flow" which model the dynamics on the ambient space, capturing the fine-grained evolution of the system.

They empirically show the effectiveness across several known datasets which include short-term variability and stochastic dynamics, using several metrics for assessing forecast performance, inference speed by number of function evaluations and error accumulation. Moreover they carried an ablation on their method to check the impact of each modelling choice. The experiments of the paper are the most technically sound.

Critical strengths

The most critical strengths of the paper are the experiments results on the forecasting performance, Table 1 and Table 2. According to the table they achieved a stronger lower relative errors against state-of-the-art baselines, even against TimeXer.

The other strength of the method relates to introduction of the latent variable which provides very informative conditions of the dynamics and help the micro-scale velocity model to keep the global dynamics, and learn the fine-grained dynamics. Morover, it is really important to notice how each modelling is combined through different losses and stage/mechanisms to train the model effectively. The ablation study analyses closely the effects of each choice. The contextual prior construction, outer-macro flow and inner micro flow modelling is very also technically sound (make sense), however i feel at the same time that several modelling add complexity in the modelling part.

Critical weaknesses

Weakness 1: Inconsistent arguments and overstated claims

Some of the paper’s arguments are inconsistent, and several statements and interpretations of the method are stronger than the evidence supports.

Geometry claim. One of the main arguments, listed among the contributions, is that the method is geometry-aware and multi-scale. It is unclear how the proposed model uses or models the geometry of the data. What can be said is that a latent representation of the dynamics helps forecasting by preserving the global evolution of the system. However, I could not find any manifold structure or geodesic metric defined or used in the method. The model learns to encode the dynamics and evolve them linearly, in both the latent and ambient space (Eqs. 11 and 19). Similarly, all target probability paths between the source (input) and the target are constructed linearly in time, so linear evolution is at most what the network can learn. The geometric claim is therefore overstated relative to the geometry the method can actually capture, given the limited assumptions made about the constructed targets. Other works address exactly this limitation with non-linear probability paths, for example Longitudinal Flow Matching (Islam et al., 2026) and Curly Flow Matching (Petrović, 2025). These works, together with the trajectory flow matching literature (Zhang et al., 2024), should be discussed in the related work. I would therefore suggest that the authors revise their statements about geometry and curvature, as they appear too strong under the current assumptions.

Theory. I found the proposition and proof for conditional flow matching unnecessary and hard to follow. Moreover, I did not see any result showing that the proposed method achieves a better bound.

Motivation. The references used to motivate multi-scale dynamics and multi-scale conflicts (Cao et al., 2025; Rahaman et al., 2019; Finlay et al., 2020) are weak support. None of them addresses the multi-scale assumption directly for dynamical systems. They focus on regularizing the dynamics or on analyzing gradient and Hessian information, in generative modelling settings where the source and target are artificial (pure noise mapped to the data distribution) rather than the states/trajectory of a real dynamical system.

Weakness 2: Experimental statements and reporting

I am also concerned about some of the statements and the reporting of the experimental results.

Seeds. No results over multiple seeds are reported, whereas other works do this. Since the method involves several loss functions and training stages, it would be very helpful to see its stability under identical hyperparameters.

Calibration claim (AEMO dataset). The method is described as well calibrated, but MSE, MAE and CRPS are not sufficient to support this. If calibration is an objective, the spread of the predictions should be assessed explicitly (e.g., spread-skill ratio, rank histograms, or coverage of prediction intervals). A lower CRPS does not imply a calibrated model, as CRPS is correlated with predictive accuracy. Moreover, where MAE and CRPS are close, this can indicate that the ensemble is overconfident, i.e., under-dispersed, rather than well calibrated.

Deterministic metrics. When evaluating conditional generative models (including baselines) with 100 ensemble members, how are deterministic metrics computed? Are they computed on the ensemble mean? Morover deterministic model be ignored when evaluatin the CRPS. It can be noticed that TimeXer beside being a discriminative model is a really strong baseline with respect the other baselines.

Figure 1. The figure does not state which experiment it comes from. Standard flow matching appears to capture the variability of the system well, and the mean of its ensemble forecast may capture the dynamics entirely. In subfigure (c), however, no ensemble forecast is shown, only a single sample, which makes the comparison hard to interpret.

Questions for authors

In addition to the weaknesses listed above, I would appreciate the authors’ clarification on the following points, which I believe are important for improving the manuscript.

  1. Seeds. Could the authors report results over multiple seeds (at least 3, ideally 5, within the compute budget)? Given the short training schedule (see question 3), this should be feasible, and it would allow assessing the stability of the method across runs.

  2. ODE stabilization claim. The contributions state that the method helps stabilize ODE integration. What exactly is meant by this? Do existing CFM approaches break down or diverge during integration, or does the method require fewer function evaluations (NFE)? If the claim concerns stability, could the authors provide an evaluation supporting it (e.g., error growth over long rollouts, sensitivity to solver or step size)?

  3. Training budget of the baselines. The networks of the proposed method are trained for only 40 epochs, which I assume takes just a few minutes. How were the baselines trained? Were they trained to convergence, with a comparable budget, or with their original settings? Please clarify to ensure the comparison is fair.

  4. Computational cost comparison (Figure 5). Comparing NFE across methods may not be technically sound if the cost of a single function evaluation differs between models (for example, latent versus ambient-space networks). Reporting wall-clock time (e.g., milliseconds per sample) would be more informative. It is also unclear how the ODE integration was configured for each method. Were the baselines integrated with the same number of steps and the same solver (Euler, RK4, or an adaptive solver such as Dopri5)? Please specify the solver used for each.

  5. Fixed variance in Eqs. 8 and 18. I could not find the values of the fixed variance used in Eqs. 8 and 18. This hyperparameter seems important for the predicted uncertainty. Which values were used in each experiment, and how sensitive are the results to them?

  6. Higher-dimensional experiment (suggestion). If feasible, it would strengthen the paper to include one additional experiment on a higher-dimensional system, such as Navier–Stokes (fluid dynamics) or Gray–Scott (reaction–diffusion) using a grid of 64x64, to show how the method scales on high dimensional problem.

Reviewer_dkpbOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Laminar Flow Matching (LFM) for time series forecasting. Rather than generating the time series in one process, LFM splits this into a macro and micro flow, focusing on trends and volatility, respectively. The macro model has an encoder and decoder, and performs generation in latent space. It is followed by the macro model, which operates in data space. In an empirical evaluation, the authors demonstrate the effectiveness of their approach and ablate every component.

Critical strengths
  • The empirical evaluation is sufficient and includes recent and important baselines. Further, all important components are ablated.
  • Both models can be trained jointly using the sg operator.
  • The code is attached.
Critical weaknesses
  • The setup is not clear to me. Multivariate forecasting is mentioned, but in Eq. 1, it looks like we have one-dimensional target sequences, even though x∈RT×D\mathbf x\in\mathbb{R}^{T\times D}
  • The method itself does not convince me, possibly because some details are unclear to me. How do you ensure that the outer macro-flow focuses on the global regime? Is this a result of the latent model? Also, it is not explained how the target is smoothed for the loss in Eq. 16. What is the difference to just applying two denoising processes sequentially with some additional layers and anchor loss in-between?
  • Some details are missing, e.g., the smoothing mentioned in the previous point, or the architecture of the encoder and decoder.
  • The connections between the main figure and equations are missing. It would be helpful to map certain parts of the figure to the corresponding definition in the text.
  • The decoupling ablation in Table 3 is surprising to me. Why does Macro in data space perform so much worse than the full model? Shouldn’t it have more degrees of freedom?
Questions for authors
  • Could you clarify how similar your approach is to [1]?
  • The architecture used in the model is unclear to me. Can you elaborate on that and clarify how the different channels of the timeseries interact?
  • The scale conflict is unclear to me. What are the implications of Proposition 1, and how does it demonstrate a limitation of previous flow-matching models?

[1] Duggal, Shivam, et al. “End-to-End Training for Unified Tokenization and Latent Denoising.” arXiv preprint arXiv:2603.22283 (2026).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.