acceptodds
Withdrawn from ICLR 2027

GapFlow: Cross-Variable Priors for Flow-Matching Time-Series Imputation on Contiguous Gaps

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow matching is a strong framework for multivariate time-series imputation. Replacing the Gaussian source with a structured prior that already interpolates the observed entries shortens the transport the network must learn, and yields large gains over diffusion and attention baselines. We show that this design has a systematic failure mode. The prior is fitted *per variable*, so it carries nothing across channels; when missingness comes in a contiguous block rather than scattered points, a variable's own history is severed and the prior degenerates to a straight line through the gap. On three of six benchmarks the resulting imputer is beaten by a plain transformer by up to : in that regime the structured prior does active damage. We introduce GapFlow, which conditions the source prior on the *observed channels* through a ridge-regularised Gaussian conditional, drives the probability path by a two-sided distance to the nearest observation, and refines the sample by self-conditioning. GapFlow attains the lowest MAE against nine baselines in of settings, spanning six datasets, four missingness mechanisms, two ratios and four window lengths. Its gains are mechanism-dependent: – under block missingness against under MCAR, with no exception across six datasets.

Primary area:
generative models
Keywords:
time series imputation, flow matching, missing data, generative models, multivariate time series, source prior

Reviews

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

This paper proposes GapFlow, a flow-matching framework for multivariate time-series imputation, motivated by the limitations of per-variable interpolation priors under contiguous missingness. GapFlow combines a cross-variable conditional prior based on ridge-regularized channel covariance, a probability path adapted to the two-sided distance from observations, and iterative self-conditioning. The paper evaluates the method across six datasets, four missingness mechanisms, two missing ratios, and multiple window lengths, reporting the lowest MAE among the external baselines in all 72 settings, with larger gains over the interpolation-prior counterpart under block missingness than under MCAR.

Critical strengths

The paper motivates using observed peer channels to improve imputation within contiguous gaps, where a variable’s own observations may provide limited information. This focus addresses a meaningful challenge in multivariate time-series imputation and makes the proposed direction worth investigating.

Critical weaknesses

W1. The paper does not sufficiently establish the validity and scope of its central problem formulation. FM-Interp performing worse than a Transformer does not demonstrate that interpolation priors themselves cause “active damage.” The comparison does not isolate the prior’s effect, and the conditions under which this claimed failure occurs remain insufficiently characterized. This weakens the foundation motivating the proposed method.

W2. Section 3.5 describes recomputing the cross-variable conditional prior using pseudo-observations. However, the pseudo-observation branch in sample_selfcond() passes an all-observed mask, disabling cross-variable mixing and overwriting the mean with pseudo-observations. The resulting operation blends predictions with a new prior sample, undermining the stated mechanism and interpretation of its ablation gains.

W3. Setting R2=0R^2=0 recovers the interpolation mean, but adaptive variance, path scheduling, and self-conditioning can remain active. The paper should clarify whether the exact-reduction claim concerns only the prior mean or the complete method, and specify the conditions required for the latter.

Questions for authors

Q1. In sample_selfcond(), the pseudo-observation branch passes an all-observed mask to prior.sample(). This makes the distance and cross-variable mixing weight zero, while the mean is overwritten by pseudo-observations. Thus, this branch returns pseudo-observations plus noise rather than the cross-variable conditional update described in Section 3.5. How do the authors reconcile this discrepancy? Did the reported self-conditioning results use this released implementation or a different version?

Q2. Is there a controlled comparison between Gaussian and interpolation priors with the network, path, training protocol, and sampling budget fixed? What evidence isolates the interpolation prior as the cause of “active damage” under block missingness?

Q3. Does the independence-based reduction claim concern only the mean in Equation (4), or the complete method? If the latter, under what conditions do the adaptive variance, probability path, and self-conditioning also reduce to FM-Interp?

If these concerns, particularly the implementation discrepancy, are satisfactorily addressed, I would consider increasing my rating.

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

This paper considers multivariate time-series imputation under contiguous missingness. It claims to identify a failure mode of the current approaches in which the prior is fitted per variable and develops GapFlow to address it.

Critical strengths

See summary

Critical weaknesses

Technically, the paper claims they identify a failure mode in which multivariate time-series imputation methods construct the prior per variable. This is not really true, like in [1], the prior is constructed considering the spatiotemporal structure of the data. Also, the paper considers multivariate time-series imputation but only compares with methods that were developed for univariate time-series imputation. The discussion of related work on multivariate time-series imputation is also insufficient. [2] is cited in the paper but not compared in experiments; why? This should be an important baseline as it is designed for multivariate time-series imputation. There are also a lot of works on generative models designed specifically for multivariate time-series imputation that are missing (like [3,4]), making the main claim and contribution of the paper questionable.

The other main problem with the paper is the writing; many parts feel AI-written, without much human assistance. Like in the introduction, there are many sentences that do not carry much meaningful information or are hard to understand. Some parts like

'Fidelity of the reimplementations. To check that the reimplemented arms reproduce their published behaviour rather than a degraded variant, we run one method at its originally published epoch budget and compare against its published number on a shared setting: we obtain 0.12084 against 0.12083, a relative difference of 0.0035%. The harness sets no random seed anywhere on that code path, so the agreement reflects the code path rather than seed alignment.'

are clearly AI-written, and even the number of relative difference in it is wrong. This is just a simple example, but from the introduction to the conclusion, similar problems appear. I believe we should treat this only as an unfinished draft in its current form, and there is still a lot of work to be done to make it publishable.

[1] Spatiotemporal Imputation with Graph-Informed Flow Matching, ICML 2026

[2] PriSTI: A Conditional Diffusion Framework for Spatiotemporal Imputation, ICDE 2023

[3] Filling the Missings: Spatiotemporal Data Imputation by Conditional Diffusion, IJCAI 2025

[4] CoSTI: Consistency Models for (a Faster) Spatio-temporal Imputation, Knowledge-Based Systems 2025

Questions for authors

See weakness

Reviewer_7EMhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Gap-adaptive Flow matching (GAPFLOW) which addresses the limitations of flow-matching methods for multivariate time-series imputation, particularly when missing values occur in long, contiguous gaps. It incorporates cross-variable information to improve imputation when individual time series lack sufficient local observations. It proposes a cross-variable conditional prior, a probability path based on two-sided distances to the nearest observations, and self-conditioned refinement to improve imputation over contiguous gaps. Experiments on six datasets across four missingness mechanisms show that GAPFLOW outperforms nine published baselines in all 72 settings, particularly for long contiguous gaps, though with higher inference costs and weaker performance on some probabilistic calibration metrics.

Critical strengths

S1: The paper evaluates GAPFLOW on six datasets under four missingness mechanisms and two missing ratios, outperforming nine published baselines across all 72 settings. Beyond reporting aggregate results, it provides detailed analyses, including ablation studies, gap-depth analysis, computational costs, and probabilistic metrics, which strengthen the credibility and interpretability of the findings. S2: GAPFLOW integrates cross-variable conditional priors with self-conditioned refinement, enabling the model to exploit correlations among variables when imputing contiguous missing segments. This provides a targeted and technically motivated solution to the limitations of interpolation-based priors, particularly for long gaps.

Critical weaknesses

W1: Although the paper evaluates GAPFLOW on six datasets, experiments with different window lengths are limited to ETTm1. It therefore remains unclear whether the reported improvements generalize to substantially longer sequences and datasets with higher dimensions, where cross-variable dependencies and computational costs may differ considerably. W2: The experiments evaluate GAPFLOW at only two missing ratios, 30% and 50%. Although these settings cover moderate levels of missingness, they do not sufficiently establish the robustness of the method under substantially lower or higher missing ratios. Additional experiments at lower and higher missing ratios would help clarify how the performance of GAPFLOW changes with the severity of missingness and whether its advantages over competing methods remain consistent across a broader range of missing conditions.

Questions for authors

Q1: The evaluation primarily uses synthetically generated missingness patterns. Have the authors also evaluated GAPFLOW on time series containing naturally occurring missing values, such as those caused by sensor failures or communication interruptions? If not, could the authors clarify whether such data were considered and discuss how the absence of this evaluation affects the assessment of the method's practical applicability?

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

Related papers

Discussion (0)

Sign in to comment.