acceptodds
Under review as a conference paper at ICLR 2027

LIFT: Local-to-Interaction Flow Matching for Multivariate Time Series Generation

Abstract

Multivariate time-series generators must reproduce both the dynamics of individual channels and the dependencies that make their trajectories jointly coherent. Several existing generators provide temporal context for cross-channel computation, but do not explicitly train the representations used for such interaction through a channel-local velocity prediction objective. We introduce LIFT, a flow-matching framework for unconditional generation that trains channel-local temporal representations and reuses the same states for multivariate refinement. A shared temporal encoder processes each channel independently, and a local head predicts its velocity using only that channel’s trajectory. An interaction module reads the resulting states and predicts an additive correction conditioned on the joint sample. The local and complete velocities are trained jointly against the same flow-matching target; under a joint realizability assumption, the ideal correction is the difference between joint and channel-local conditional velocity estimates. A mean–carrier module combines broadcast mean context with channel-specific readouts from a small set of pooled carriers, yielding computation linear in channel count at fixed width and carrier count. Across six benchmarks and eight baselines, LIFT achieves the lowest three-seed mean in 23 of 24 dataset–metric comparisons. Controlled variants further support temporally informed interaction and the reuse of locally supervised temporal states for joint refinement.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time series generation, Flow matching, Temporal modeling, Cross-channel dependencies, Generative modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q9hXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LIFT, a local-to-interaction flow matching framework for unconditional multivariate time-series generation. The method decomposes the joint velocity field into a channel-local velocity predictor and an additive interaction-conditioned correction. A shared temporal encoder provides representations for both components, while a mean–carrier interaction module facilitates efficient cross-channel information exchane. The local and complete velocity fields are jointly optimized using flow matching objectives. Experiments on six popular benchmarks against eight baselines demonstrate strong generation quality across multiple evaluation metrics.

Critical strengths
  1. The paper proposes an interesting perspective of generative model-based unconditional MTS generation, jointly modeling channel-specific temporal dynamics and cross-channel dependencies. The proposed local-to-interaction architecture is intuitively motivated and clearly presented.
  2. The paper is easy to follow and reproduce
  3. The paper is well-organized and presents a comprehensive experimental evaluation. In particular, LIFT achieves the lowest reported mean in 23 out of 24 dataset–metric comparisons, demonstrating strong empirical performance. Overall, the manuscript is well-structured, with a thorough methodological presentation and extensive experimental validation.
Critical weaknesses
  1. Some incremental methodological novelty. The main architecture combines several established ideas, including channel-independent temporal modeling, cross-channel interaction, auxiliary supervision, and additive residual refinement. To the best of my knowledge, existing methods, such as CSDI, PaD-TS, and DiM-TS, have already explored temporally informed cross-channel modeling, while compact information exchange through shared representations or routing mechanisms has also been investigated in Crossformer. Therefore, the distinguishing contribution of LIFT appears to be the explicit channel-local velocity supervision and refinement strategy. I also found from the ablation results that temporal context and cross-channel interaction are important. However, the largest improvements occur when comparing Full LIFT with Local Only or Pointwise Interaction, which mainly demonstrates the importance of temporally informed cross-channel modeling rather than the specific benefit of local velocity supervision. Therefore, I have reservations about the extent to which this supervision mechanism contributes to the overall performance.

  2. Missing comparisons with directly relevant generation baselines. For the unconditional time-series generation task, TimeMAR [1] and SDformer [2] are highly relevant and competitive generative baselines. The authors should include comparisons with these methods to provide a more comprehensive evaluation and further substantiate the effectiveness of LIFT.

  3. The proposed mean–carrier interaction module fixes the number of carriers to K=3 across all benchmarks, despite substantial differences in channel dimensionality, ranging from 5 to 50. However, the paper does not explain the rationale behind this choice or provide a sensitivity analysis of the carrier count. please clarify.

  4. One more question: Does the interaction residual strategy actually learn information that is unavailable to the channel-local predictor, or does it simply behave as an additional nonlinear velocity refinement network?

[1]. Xu, Xiangyu, Qingsong Zhong, and Jilin Hu. "TimeMar: Multi-Scale Autoregressive Modeling for Unconditional Time Series Generation." Proceedings of the ACM Web Conference 2026. 2026.

[2]. Chen, Zhicheng, et al. "Sdformer: Similarity-driven discrete transformer for time series generation." Advances in Neural Information Processing Systems 37 (2024): 132179-132207.

Questions for authors
  1. Please refer to critical weaknesses
1 comment
AuthorsOpenReview ↗

Response to Reviewer q9hX

We sincerely thank the reviewers for their thorough reading and valuable comments. Unfortunately, due to time constraints on our side, we are unable to provide satisfactory responses to the questions raised. We feel sorry for not being able to make full use of your insightful suggestions and apologize for the time you have spent on our paper.

Reviewer_NLZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces LIFT, a flow-matching framework for unconditional multivariate time-series generation. The main idea is to split the velocity field into two parts: a channel-local component that predicts from the target channel's own trajectory, and an interaction component that uses temporal information from all channels to make an additive correction. Both components share a four-layer bidirectional S4D encoder and are trained jointly with the same flow-matching objective. The local component also has an auxiliary loss, whose weight gradually decreases from 1 to 0.25 during training. For cross-channel interactions, the authors design a mean–carrier module. It broadcasts the channel mean as shared context and uses three carriers with softmax-based read/write weights to aggregate and redistribute centered channel representations. This keeps the computational cost linear in the number of channels. Appendix D provides a theoretical analysis showing that, when the auxiliary loss weight is positive and joint realizability holds, the optimal local component learns the conditional expectation given only the target channel, while the interaction component learns the additional contribution from conditioning on all channels. Experiments on six datasets against eight baselines show that LIFT achieves the lowest mean in 23 of 24 dataset–metric comparisons. The authors conduct ablations on temporal context, state reuse, local supervision, and interaction mechanisms, along with fixed-model context interventions and longer-sequence evaluations. Sampling costs and anonymized code are also provided.

Critical strengths
  1. The PWI ablation provides convincing evidence for the importance of temporal context in cross-channel interaction. The performance improvements are substantial across several datasets.

  2. The experimental evaluation is relatively thorough, with multiple training seeds, confidence intervals, and additional analyses of architectural choices and sequence lengths. The fixed-model interventions are particularly useful, providing evidence that the learned vector field is sensitive to cross-channel correspondence. And codes are provided.

Critical weaknesses

1.The paper's main design principle of reusing locally supervised temporal states for joint refinement is not sufficiently motivated or justified. In particular, it is unclear why representations learned through channel-local velocity prediction should be more effective for cross-channel interaction than representations learned directly through joint supervision or separate temporal encoders. The phase-adjustment example in Appendix A motivates the importance of temporal context in general, but does not specifically justify local supervision or state reuse. Table 2 also suggests that most of the improvement comes from temporal context, while state reuse provides relatively modest additional gains (e.g., ETTh1 Context-FID: 0.014 vs. 0.012, with overlapping confidence intervals). The local-loss ablation also shows mixed results. Moreover, comparisons with existing methods involve different architectures and training recipes, making it difficult to attribute the overall gains to the proposed design.

2.The evaluation does not rule out memorization. Models are trained and evaluated against the same corpus, without a held-out test set. This is particularly concerning for Stocks, which contains only 3,662 heavily overlapping windows. Although the authors acknowledge that the reported scores do not reflect generalization, no memorization analysis is provided.

Questions for authors

1.Why does the ablation without the local auxiliary loss report only Context-FID and FDDS? Could the authors clarify why only three-decimal aggregates are available and, if possible, provide the missing Discriminative and MDD scores?

2.Since the models are trained and evaluated against the same corpus, could the authors provide additional evidence to assess potential memorization, particularly on Stocks with only 3,662 heavily overlapping windows?

3.Could the authors provide more intuition for why locally supervised temporal representations should be particularly useful for joint refinement? Specifically, why should supervision from channel-local velocity prediction help capture information relevant to cross-channel dependencies, compared with learning temporal representations directly through joint supervision? The current phase-adjustment example appears to motivate temporal context in general, rather than the proposed local supervision and state reuse.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NLZR

We sincerely thank the reviewers for their thorough reading and valuable comments. Unfortunately, due to time constraints on our side, we are unable to provide satisfactory responses to the questions raised. We feel sorry for not being able to make full use of your insightful suggestions and apologize for the time you have spent on our paper.

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

This paper proposes LIFT, a flow-matching framework for unconditional multivariate time series generation. A shared temporal encoder processes each channel independently and is trained through a channel-local velocity prediction objective. The same temporal states are reused by an interaction module that predicts an additive correction to the local velocity. A compact mean-carrier mechanism exchanges information across channels. The paper evaluates generation quality on six benchmarks and studies the effects of temporal context, local supervision, representation sharing, and different interaction modules.

Critical strengths
  1. The paper studies a clear and useful question: whether temporal representations explicitly trained for channel-local prediction can also provide a good basis for modeling multivariate interactions. The local and joint prediction objectives give the architecture a coherent motivation.

  2. The controlled evaluation is relatively comprehensive. The separate-encoder controls, removal of local supervision, direct-output variant, and alternative interaction modules address several plausible explanations for the observed gains.

  3. The reported generation performance is consistently strong across several datasets and metrics. The paper also distinguishes corpus fidelity from held-out generalization and clearly states the assumptions behind the conditional-velocity interpretation.

Critical weaknesses
  1. From an information-flow perspective, LIFT's Mean–Carrier module resembles the router-based interaction mechanisms used in existing methods such as Crossformer, where representations from (C) variables are aggregated into a small number of intermediate representations and subsequently distributed back to the variables. The main differences lie in the centering of variable representations, the specific value transformation used to construct the carriers, and the introduction of flow-time-conditioned gating and residual updates. However, the paper does not sufficiently explain what substantive advantages these modifications offer over existing router-based interaction mechanisms. Although the authors compare the proposed module with alternatives such as channel attention and STAR-style aggregation, they do not provide a direct comparison with a standard router-based aggregation-and-distribution mechanism. Consequently, the current experiments do not sufficiently establish the methodological novelty of the proposed interaction mechanism.
  2. The paper employs three carriers for cross-variable information exchange, but the rationale for this particular choice is insufficiently explained, and a systematic sensitivity analysis over different numbers of carriers is missing. Furthermore, centering, carrier write and read weights, carrier-message gating, and flow-time-dependent scaling are all specific design choices within the proposed module. However, the current ablations primarily compare the complete module against overall architectural variants such as Mean Only, channel attention, and STAR-style aggregation, without independently evaluating the contributions of these individual components.
  3. The paper trains the models on the full data corpus and uses the same corpus as the reference distribution when evaluating generated samples. Consequently, the experiments primarily measure how well the models fit the observed data distribution, rather than their generalization to unseen data. Although the authors acknowledge this limitation, it also means that the main experimental results do not directly establish that the proposed method maintains its generation advantages under broader or previously unseen data distributions. The paper should more clearly distinguish distributional fidelity to the observed training corpus from generalization to unseen data and ensure that its claims remain within the scope supported by the evaluation protocol.
  4. Throughout the main text and appendices, the paper repeatedly introduces explicit qualifications regarding what individual results do not establish. Examples include clarifying that mean rankings are not significance tests, that the experiments do not demonstrate generalization to unseen data, and that certain diagnostic results should not be interpreted as improvements in generation quality or measurements of isolated interaction cost. Although such qualifications are not inherently inappropriate, they are dispersed across different sections and are not consistently integrated into a clear argument linking the paper's main contributions to the corresponding empirical evidence. In addition, the relationships between the motivation for the core interaction module, its specific parameterization, the theoretical interpretation, and the ablation results are not sufficiently clear. Readers must repeatedly consult different sections to determine which claims are actually supported by each experiment. This weakens the clarity of the paper's central methodological argument and makes it difficult to assess whether its conclusions are adequately supported by the presented evidence.
Questions for authors
  1. What substantive advantage does Mean-Carrier exchange offer over a standard router-based aggregation-and-distribution mechanism under comparable modeling and training conditions?

  2. Which design choices within Mean-Carrier exchange are necessary for its gains, and how sensitive are the results to the number of carriers?

  3. What evidence supports extending the reported generation advantages to unseen data, beyond fidelity to the training corpus?

1 comment
AuthorsOpenReview ↗

Response to Reviewer NppN

We sincerely thank the reviewers for their thorough reading and valuable comments. Unfortunately, due to time constraints on our side, we are unable to provide satisfactory responses to the questions raised. We feel sorry for not being able to make full use of your insightful suggestions and apologize for the time you have spent on our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.