acceptodds
Under review as a conference paper at ICLR 2027

Conditional Flow Matching for Transport Between Markov Processes

Abstract

Motivated by sequence-to-sequence transport tasks such as time-series domain adaptation, prompt and query rewriting, and machine translation, we study transport between trajectories of Markov processes. Given one source trajectory and one target trajectory, we develop a conditional flow matching algorithm that learns to transform a fresh source trajectory into a trajectory following the target dynamics. The flow conditions on the current source and generated target states, preserving the target Markov structure under recursive generation. Under suitable regularity and initialization assumptions, the population flow reproduces the target trajectory distribution. For finite-sample learning over a finite realizable velocity class, we establish an error bound that incurs a factor proportional to the mixing time compared with independent transitions, as in classical statistical problems including regression (Nagaraj et al., 2020), principal component analysis (Kumar and Sarkar, 2023), and matrix concentration (Neeman et al., 2024). We complement this result with a lower-bound construction showing that dependence on the mixing time is unavoidable, even for regular Gaussian conditional transitions. We evaluate the method on synthetic data and supervised image retrieval from electroencephalography (EEG) recordings in THINGS-EEG2 (Gifford et al., 2022). For each subject, we use 4,096 calibration images while keeping the ENIGMA encoder and projection network pretrained on other subjects frozen (Kneeland et al., 2026; Alljoined, 2026). Compared with adapting only the subject-specific temporal map, jointly training this map with a preceding conditional flow using 32 previous EEG samples raises mean top-5 retrieval accuracy across ten subjects from 43.87% to 49.05%, an 11.82% relative improvement.

Primary area:
generative models
Keywords:
Conditional flow matching, Markov processes, dependent data, sample complexity, mixing time, trajectory transport, domain adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zCReOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

They investigate how we can transporting a trajectory to a target one when considering both are driven by Markov processes. They propose to learn a conditional flow-matching velocity function on the interpolation of independently drawn next states to do that, conditioned on the source and target states. They use single (long) source trajectories with a single target trajectory as the training data. For the generations, they use recursively via a transition function from x'(i+1) to y(i+1). To evaluate their method, they used synthetic Gaussian and nonlinear-mixture transition recovery with an EEG-to-image retrieval study on THINGS-EEG2. As their results shows a history-conditioned flow inserted before subjects' temporal map increases the mean Top-5 accuracy ~5% across ten subjects.

Critical strengths

The idea of reduction from trajectory transport to a conditional transition problem, and its verification in a Gaussian instance, is an well defined and clear idea. Also, even though the experiments do not fully match or support the paper’s claims, the EEG experiment is interesting and represents an application that is less explored with CFMs.

Critical weaknesses

1-The main issue is EEG experiment does not clearly show that flow matching is the reason for the improvement. Specifically the flow is trained together with the retrieval loss which adds much more parameters than the baseline, and uses a long history that goes beyond the paper’s theory. Also there is no comparison against a simple adapter of the same size, against the same model without the flow-matching loss, or against other cross-subject EEG methods. The synthetic experiments only test one-step accuracy and do not test full trajectory generation.

2-As I understand, the you developed theory only guarantees the output follows the target distribution. It does not explain what the map keeps from the source trajectory while I think this is the most important reason to use transport for domain adaptation. The finite model class assumption makes the bounds of limited use for neural networks. Also several assumptions and propositions are developed in the main body of the paper and left with no testing nor the experiments are sufficient or clearly test them.

Questions for authors

1- How does the EEG model perform without the flow-matching loss? and how does a simple adapter with a similar number of parameters perform?

2- Can you report how errors build up over long generated trajectories atleast in the synthetic experiments?

3- What does the learned map keep the same from the source trajectory beyond generating the correct target distribution?

4- Can the mixing-time factor in the main rate be removed?

7 comments

We thank the reviewer for their careful reading and detailed feedback. We appreciate the questions about the experimental comparisons, the properties of the learned transport map, and our theoretical guarantees. In response, we conducted additional ablation experiments and evaluated trajectory-level transport over longer horizons. We address each point below.

1. EEG experiments and ablation studies

Contribution of the flow-matching loss and comparison with a similarly sized adapter.

We thank the reviewer for suggesting these comparisons. We conducted additional experiments on THINGS-EEG2 using the same subject splits, training data, initialization, and optimization budget as our original experiments.

To isolate the contribution of the flow-matching (FM) objective, we set its loss coefficient to zero while retaining the same flow architecture and supervised retrieval objective. Both the subject-specific temporal map and velocity network remain trainable. We also evaluated a simple two-layer linear adapter with approximately the same number of trainable parameters.

Top-5 test retrieval accuracy (%), averaged over ten subjects and three seeds:

Flow widthTemporal map only (ENIGMA baseline)Markov CFM (ours)Without FM lossSimple adapter with similar trainable parameters
1443.8746.7344.6344.82
6443.8749.0546.8246.58

The above results indicate that adding the flow-matching objective improves Top-5 retrieval accuracy beyond what is achieved by the identical architecture without that loss or the similarly sized adapter.

Comparison with other cross-subject EEG methods.

We thank the reviewer for encouraging a broader comparison. Recent methods such as SATTC [2] and SCORE [3] address cross-subject EEG-to-image retrieval through different approaches. SATTC calibrates retrieval similarities using frozen EEG and image encoders, while SCORE learns representations designed for cross-subject coordinate alignment without labeled target-subject calibration. Our experiment instead studies whether conditional flow matching improves adaptation of a fixed pretrained ENIGMA decoder, using the same subject-specific calibration data, encoder, and training objective across all compared methods. Since the published SATTC and SCORE results use different pretrained models and adaptation protocols, their reported accuracies are not directly comparable to ours. We will expand the related-work discussion to explain these differences and clarify how Markov CFM relates to these approaches.

Use of previous EEG samples.

We thank the reviewer for raising this point. Our construction extends directly to higher-order Markov processes by treating the previous k observations as the current state. Specifically, for an order-k process, define

Y_i^{(k)}=(Y_{i-k+1},\ldots,Y_i).$$ These history processes are first-order Markov. We can therefore learn a conditional flow transporting the next-observation law $P_X^{(k)}(X_i^{(k)},)$ to $P_Y^{(k)}(Y_i^{(k)},)$, conditioned on the two histories, exactly as in our original construction. At generation time, we apply the learned flow to the next source observation and update the target history by appending the generated observation. With the initial target history drawn from the correct joint distribution independently of the source, the same conditional-transport and induction argument as in Proposition 2 gives the correct target trajectory distribution. The finite-sample analysis likewise extends to these augmented history processes under the corresponding mixing and moment assumptions. Our EEG implementation uses k=32 as a modeling choice. We do not assume that the EEG recordings themselves satisfy the Markov or mixing assumptions required by the theoretical guarantees. We will clarify this extension in the revised paper. ### 2. Error accumulation over longer trajectories Please see our response to Reviewer LGtD

3. What properties of the source trajectory are preserved?

Our theoretical objective is to reproduce the target trajectory distribution, and we acknowledge the reviewer's point that this alone does not guarantee preservation of task-specific labels or semantic information from the source.

However, the learned map does retain a precise relationship with the source trajectory. Each generated observation is obtained by transporting the corresponding source observation, conditioned on the current source and generated target states. Under our flow regularity assumptions, the local ODE maps are invertible. Consequently, given the initial source and target states, the source trajectory can be recovered by successively inverting these maps.

In one dimension, our transport preserves conditional quantiles: an observation at the p-th percentile of the source transition distribution is mapped to the p-th percentile of the target transition distribution. As discussed in our response to Reviewer LGtD, recursively applying these maps gives the Knothe–Rosenblatt rearrangement.

Furthermore, in our EEG experiments, source and target recordings corresponding to the same image are paired during training, and the supervised retrieval objective encourages preservation of image-relevant information. Importantly, the flow network transports EEG signals without explicitly conditioning on image labels, since inferring the label from the signal is the task itself. In our case, letting C denote the viewed image, we use a coupling in which X∼PX(⋅∣C))X\sim P_X(\cdot\mid C)) and Y∼PY(⋅∣C)Y\sim P_Y(\cdot\mid C) are independent conditional on C, but generally dependent marginally. Suppressing temporal conditioning for simplicity, the flow-matching objective is

\mathbb E_{X,Y\sim P_X(\cdot\mid C)\times P_Y(\cdot\mid C)} \mathbb E_t \left[||v_t(Z_t)-(Y-X)||^2\right], \qquad Z_t=(1-t)X+tY.$$ If, for example, the conditional interpolation distributions $P(Z_t\mid C=c)$ have disjoint supports across image labels for almost every t, then, under suitable flow regularity, the population flow transports $P_X(\cdot\mid C=c)$ to $P_Y(\cdot\mid C=c)$ for each c. Although we do not assume this condition holds for EEG, same-image pairing and the supervised retrieval loss encourage preservation of task-relevant information. Thus, our experimental objective incorporates image-level correspondence, while our theoretical guarantees concern distributional transport under the stated assumptions. We will clarify this distinction in the revised paper. ### 4. Can the mixing-time factor be removed? We thank the reviewer for raising this important question. Proposition 3 establishes that dependence on the mixing time is unavoidable in the sample complexity under our assumptions. Specifically, we construct two instances with Gaussian conditional transitions and a common class of only two candidate velocities. If the training trajectory is too short relative to the mixing time, it may fail to visit the region where the two models differ, making them statistically indistinguishable. This construction follows the lower-bound approach of Nagaraj et al. [6]. For an estimator selecting from this class to achieve the stated risk threshold on both instances with failure probability $(0,1/2\pi)$, Proposition 3 requires $$m\geq \frac{\tau_{\mathrm{mix}}}{3}\log\left(\frac{1}{2\pi\delta}\right),$$ where $\tau_{\mathrm{mix}}$ is the mixing time at tolerance $1/4$. Thus, even for regular Gaussian transitions and a class containing only two velocities, the required sample size must scale at least linearly with the mixing time for fixed confidence. It may be possible to improve the mixing-time dependence in the estimation error bound of Theorem 1. However, Proposition 3 shows that no such improvement can eliminate the mixing-time dependence of the sample-size requirement in general. We would welcome any concrete suggestions the reviewer may have for sharpening the upper bound.

5. Finite model class and assumptions

We appreciate the reviewer's concern about the applicability of our assumptions to neural-network estimators. Our main contribution is an algorithm for learning transport between time-series distributions from limited, temporally dependent observations, together with theoretical guarantees under explicit assumptions. We clarify these assumptions below.

Finite model class. Finite hypothesis classes are a classical setting in statistical learning theory, including the analysis of neural networks (Anthony and Bartlett [11]). Covering-number arguments provide a standard approach to extending finite-class bounds to infinite function classes, under suitable approximation and regularity conditions.

Moreover, as explicitly noted by Anthony and Bartlett [11, Section 2.2], neural networks with finitely many possible parameter values form finite function classes. In particular, for a fixed neural-network architecture with p parameters, each represented using at most q possible values, the corresponding function class satisfies ∣V∣≤qp,log⁡∣V∣≤plog⁡q.|\mathcal V|\leq q^p,\qquad \log|\mathcal V|\leq p\log q.

This finite-precision interpretation is also used in the analysis of score matching for diffusion models by Kumar, Nagaraj and Sarkar [7]. Gupta et al. [8] similarly establish finite-class score-estimation bounds and extend them to neural networks using covering-number arguments. Our guarantees can therefore naturally extend to neural-network based function classes through such techniques.

Other assumptions. The assumptions in our analysis serve distinct purposes and have precedents in statistical learning and Markov-process analysis.

The conditional sub-Gaussian noise assumption (Assumption 2, lines 298–307) controls regression fluctuations. Appendix C verifies it for bounded displacements, bounded additive noise, and Gaussian transition kernels with uniformly bounded covariance.

The fourth-to-second moment comparison (Assumption 3, lines 324–328) is also used in score-estimation analysis [7]. In Appendix B, we explicitly verify realizability and the moment comparison for our Gaussian example

For finite-sample rollout guarantees, the coverage condition (Eq. (41), lines 1193–1195) controls the difference between the training distribution and the distribution of generated state pairs, analogous to concentrability conditions in offline reinforcement learning [9].

The contraction assumption (Theorem 2, lines 1195–1198; Assumption 5, lines 1242–1248) controls how local transition errors accumulate over time, as in perturbation analyses of Markov processes [10]. Our additional trajectory experiments evaluate the behavior of recursive generation, while the theoretical guarantees hold under the stated assumptions. We will clarify their role and applicability in the revised paper.

References

[1] Kneeland et al., ENIGMA: EEG-to-Image in 15 Minutes Using Less Than 1% of the Parameters, 2026.

[2] Huang and Zhu, SATTC: Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval, CVPR 2026.

[3] Cui et al., SCORE: Subject Coordinate Recovery for Label-Free Cross-Subject EEG-to-Image Retrieval, 2026.

[4] Rowland et al., Orthogonal Estimation of Wasserstein Distances, AISTATS 2019.

[5] Courty et al., Joint Distribution Optimal Transportation for Domain Adaptation, NeurIPS 2017.

[6] Nagaraj et al., Least Squares Regression with Markovian Data: Fundamental Limits and Algorithms, NeurIPS 2020.

[7] Kumar, Nagaraj and Sarkar, Dimension-free Score Matching and Time Bootstrapping for Diffusion Models, NeurIPS 2025.

[8] Gupta et al., Improved Sample Complexity Bounds for Diffusion Model Training, NeurIPS 2024.

[9] Zanette and Wainwright, Bellman Residual Orthogonalization for Offline Reinforcement Learning, NeurIPS 2022.

[10] Rudolf and Schweizer, Perturbation Theory for Markov Chains via Wasserstein Distance, Bernoulli 2018.

[11] Anthony and Bartlett, Neural Network Learning: Theoretical Foundations, Cambridge University Press, 1999.

Reviewer_LGtDOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a method for learning a map that transforms a source trajectory into one following the target dynamics. Specifically, it learns a family of conditional transport maps Φx,y\Phi_{x,y} that push forward PX(x,⋅)P_X(x,\cdot) to PY(y,⋅)P_Y(y,\cdot), which is constructed implicitly by integrating velocity fields learned through flow matching conditioned on the current source and target states. The authors provide theoretical guarantees and evaluate the method on synthetic and real-world data.

Critical strengths
  • The central problem and its motivation are stated clearly. The supporting analysis is sound; the proofs are rigorous, self-contained, and verified on explicit examples.
  • The algorithm is simple and simulation-free to train. Experiments cover both synthetic and real-world data, and the scope and results are reported transparently.
Critical weaknesses
  • The method combines established components, such as conditional flow matching and recursive transport of transition laws, which limits its methodological novelty.

  • Problem (9) generally admits multiple solutions. The proposed construction selects a particular solution, but its advantages and disadvantages relative to alternative transport maps remain unclear. A discussion of the properties and the suitability in applications of the proposed solution would strengthen the paper.

Questions for authors
  • In the 1D settings, the flow matching coincides with the optimal transport between the conditional transition laws. Conditional on the initial states, its recursive application should yields the Knothe-Rosenblatt rearrangement (see e.g. [1]) of the future-trajectory distributions. This observation is more general than the Gaussian example in Proposition 6. Could the authors add a brief discussion of this?

  • The experiments report conditional next-state distribution error, but do not directly assess the quality of trajectories transport. Could the authors report trajectory-level distribution error across increasing rollout lengths, using, for example, the trajectory Wasserstein distance in (47)?

[1] Optimal Transport for Applied Mathematicians. Filippo Santambrogio.

2 comments

We thank the reviewer for their careful reading of our paper, positive assessment of the theoretical results, and constructive suggestions. We address each point below.

1. Novelty and choice of transport map

We agree that conditional flow matching is an established method for learning velocity fields [1]. To our knowledge, ours is the first finite-sample analysis of conditional flow matching for transport between Markov processes using temporally dependent training trajectories. We establish conditions under which recursively applying conditional transport maps generates the correct target trajectory distribution, and prove finite-sample estimation bounds together with a lower bound showing that dependence on mixing time is unavoidable.

As the reviewer points out, Eq. (9) admits multiple transport maps. Applications with a particular cost requirement may prefer a different solution to Eq. (9), depending on the problem constraints, which can potentially be encouraged by appropriate augmented loss functions. For e.g. in our case, we augment the flow matching loss with a retrieval loss and train jointly. We will discuss these properties in the revised paper. Please also see our response to Reviewer zCRe for further details.

2. Connection to the Knothe–Rosenblatt rearrangement

We thank the reviewer for pointing out this connection. Indeed in the one-dimensional setting, our construction corresponds to the Knothe–Rosenblatt rearrangement.

Under our flow regularity assumptions, the scalar ODE flow preserves order. Thus, for continuous conditional transition distributions, its time-one map is the increasing quantile transport, given by

Φx,y(a)=FY∣y−1(FX∣x(a)),\Phi_{x,y}(a) = F_{Y\mid y}^{-1}\bigl(F_{X\mid x}(a)\bigr),

where FX∣xF_{X\mid x} and FY∣yF_{Y\mid y} are the conditional cumulative distribution functions. Conditional on fixed initial states, recursively applying these maps in temporal order gives the Knothe–Rosenblatt rearrangement between the corresponding future-trajectory distributions [2, 3]. This generalizes Proposition 6 beyond Gaussian transitions. Related work on monotone triangular transport and conditional sampling includes Baptista et al. [5] and the conditional flow-matching approach of Zhai et al. [6]. We will add this connection and the relevant references to the revised paper.

3. Trajectory-level Wasserstein evaluation

We conducted additional experiments measuring the sliced Wasserstein distance (SW2) [4], a standard metric for comparing probability distributions, between generated and target trajectories over increasing rollout lengths. We apply this metric to entire trajectories using one-dimensional projections, providing a computationally tractable measure of trajectory-level distribution error.

We reused the trained checkpoints from our dimension-10 Gaussian and nonlinear-mixture experiments, with 32,768 training transitions. For each method, we generated 4,096 fresh trajectories and computed full-path sliced Wasserstein distance using 256 shared random projection directions. We evaluated horizons of H∈{1,4,16,32,64}H \in\{1,4,16,32,64\} transitions and averaged results across three seeds. We compared Markov CFM with the ablation without the previous source state, marginal flow, and the oracle using the true transition kernels.

Full-path sliced Wasserstein distance at 64 transitions (lower is better):

BenchmarkMarkov CFM (ours)Without previous source stateMarginal flowOracle (true kernels)
Gaussian0.06050.07660.08090.0370
Nonlinear mixture0.04220.20380.06530.0378

At 64 transitions, Markov CFM has the lowest error among the learned methods on both benchmarks. These results complement our existing one-step transition experiments and provide direct empirical evidence that conditioning on the previous source state improves recovery of the target trajectory distribution. We will include these additional trajectory-level results in the revised paper.

References

[1] Lipman et al., Flow Matching for Generative Modeling, ICLR 2023. Paper.

[2] Santambrogio, Optimal Transport for Applied Mathematicians, Birkhäuser, 2015.

[3] Villani, Optimal Transport: Old and New, Springer, 2009.

[4] Rowland et al., Orthogonal Estimation of Wasserstein Distances, AISTATS 2019. Paper.

[5] Baptista et al., Conditional Sampling with Monotone GANs: From Generative Models to Likelihood-Free Inference, SIAM/ASA JUQ, 2024.

[6] Zhai, Jeong, and Ročková, Conditional Flow Matching for Bayesian Posterior Inference, AISTATS 2026.

I thank the authors for clarifying the connection to the Knothe-Rosenblatt rearrangement and for adding the trajectory-level evaluation. These additions address my main concerns. I find the finite-sample analysis for dependent trajectories a substantive contribution and will raise my score.

Reviewer_JG6GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, the authors proposed a new method, which utilizes conditional flow matching to transports the trajectories followings Markovian processes. The aim is to preserve the target traj's transition structure and answert the question of "can we learn, from one source and one target trajectory, a map that transforms a fresh source trajectory into one that follows the target dynamics?". The extensive experiments of this work on synthetic dataset and cross-subject electroencephalography (EEG), functional magnetic resonance imaging (fMRI) signals adaptation show improved domain adaptation performance.

Critical strengths
  1. The paper is well-written and has a concrete flow. The problem formulation of this work is well-stated in the Introduction Section.
  2. The proposed methodology with Population correctness and Sample complexity bounds grounded by theorem is solid for a submission to ICLR. The detailed derivations in the appendix A is overall make sense to me.
Critical weaknesses
  1. There seems to be no standard name for the proposed conditional flow-matching framework.
  2. It's good to see that the proposed framework has been evaluated on both synthetic dataset and EEG, fMRI real-world datasets, while the experimental results in Section 4 is a bit not abudant.
  3. The related-work Section 1.2 now could be more expanded to have more generative approaches to transit trajectories, such as ERDiff [1]. It uses diffusion (score-based) probabilistic models to preserve spatio-temporal structure when aligning latent dynamics across source-target trajectories.

[1] Extraction and Recovery of Spatio-Temporal Structure in Latent Dynamics Alignment with Diffusion Models, NeurIPS 2023.

Questions for authors

I have no more questions at this point here. Please refer to my weaknesses section to see my other questions/concerns.

1 comment

We thank the reviewer for their positive assessment of our problem formulation and theoretical contributions, and for their constructive suggestions. We address each point below.

1. Naming of the proposed method

We will consistently refer to our method as Markov Conditional Flow Matching (Markov CFM) throughout the revised paper.

2. Additional experimental evaluation

We appreciate the reviewer's suggestion to broaden the experimental evaluation. We conducted an additional experiment on sleep-stage classification using the EESM19 dataset, which contains simultaneous recordings of brain electrical activity from electrodes placed on the scalp (scalp EEG) and in or around the ears (ear-EEG).

The task is to predict a person's sleep stage from EEG recordings, with each 30-second segment assigned a sleep-stage label. We investigate whether a classifier trained on scalp EEG can be adapted to classify sleep stages from ear-EEG recordings of an unseen participant. Since signals recorded at the ears differ from those recorded on the scalp, we use Markov CFM to transform ear-EEG features into a representation suitable for the frozen scalp classifier.

We used a leave-one-participant-out evaluation across 20 participants. For each held-out participant, we trained the scalp classifier on labeled recordings from the other 19 participants and then froze it. The adapters were trained using paired ear-EEG and scalp EEG recordings from these same 19 participants. At test time, the adapted classifier received only ear-EEG recordings from the held-out participant. We evaluated 71,055 thirty-second epochs and averaged accuracy equally across all 20 participants.

MethodSleep-stage accuracy
Frozen scalp classifier, no adapter (scalp normalization)43.2%
Frozen scalp classifier, linear adapter62.1%
Frozen scalp classifier, Markov CFM66.1%
Classifier trained directly on labeled ear-EEG (supervised reference)70.0%

Markov CFM improves accuracy by 4.0 percentage points over the linear adapter with same number of parameters. It also improves accuracy by 22.9 percentage points over directly applying the frozen scalp classifier to ear-EEG recordings. These results complement our existing EEG-to-image retrieval experiments and provide further evidence of the usefulness of Markov CFM for neural signal domain adaptation. We will include this experiment in the revised paper.

3. Comparison with ERDiff

We thank the reviewer for pointing us to ERDiff [1]. ERDiff fits a diffusion model to source neural dynamics and uses its likelihood to align a target encoder. Our approach instead learns a conditional transport map between source and target transition distributions, conditioned on the current source and target states. The map is applied recursively to generate a trajectory following the target transition law, for which we establish population correctness and finite-sample guarantees under mixing assumptions. We will expand Section 1.2 of the revised paper to discuss ERDiff and clarify its relationship to our approach.

References

[1] Wang et al., Extraction and Recovery of Spatio-Temporal Structure in Latent Dynamics Alignment with Diffusion Models, NeurIPS 2023. Paper.

[2] EESM19 dataset, OpenNeuro, ds005185, version 1.0.0. Dataset.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We sincerely thank all three reviewers for their careful reading, constructive feedback, and thoughtful assessment of our work. We appreciate their recognition of our problem formulation and theoretical contributions. Their suggestions have helped us strengthen the experimental evaluation and clarify the scope of our theoretical results. We summarize our main responses below.

New experimental evaluation on cross-device EEG adaptation.: To further evaluate the applicability of our algorithm Markov Conditional Flow Matching (Markov CFM), we conducted an additional experiment on the EESM19 sleep-stage classification dataset. We investigated whether a classifier trained on scalp EEG could be adapted to classify sleep stages from ear-EEG recordings of unseen participants. Across 20 held-out participants, Markov CFM achieves 66.1% accuracy, compared with 43.2% without an adapter and 62.1% with a simple linear adapter with same number of trainable parameters.

Theoretical extensions and clarification of the transport construction.: We explain how our approach extends naturally to general kk-th order Markov processes by treating the previous kk observations as an augmented first-order Markov state, under the corresponding assumptions. Based on the helpful pointer by Reviewer LGtD, we also establish that our recursively applied conditional flows coincide with the Knothe–Rosenblatt rearrangement, a classical triangular transport map constructed through successive conditional quantile transformations. This identifies the particular trajectory coupling selected by our method and extends our scalar Gaussian example to more general transition distributions. Furthermore, we explain the role of our finite-sample estimation and rollout assumptions, provide examples where they hold, and connect them to established results in statistical learning theory.

Direct evaluation of trajectory transport and temporal dependence.: In response to the request for trajectory-level experiments, we evaluated sliced Wasserstein distance between generated and target trajectory distributions over increasing rollout lengths. Markov CFM achieves lower trajectory-distribution error than both learned baselines on both synthetic benchmarks.

We also conducted additional ablation experiments on our setting, comparing Markov CFM against the same flow architecture trained without the flow-matching loss and a similarly sized simple adapter. These experiments provide further evidence of the contribution of the flow-matching objective beyond the additional trainable parameters. Our individual responses provide the experimental details and mathematical arguments addressing each reviewer's concerns. We will incorporate these additional results and clarifications into the revised manuscript. We thank the reviewers again for their valuable suggestions.

Dear Reviewers, the authors’ responses are now available. If you haven’t already, please review them and share your comments before the discussion period closes. Thank you for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.