acceptodds
Under review as a conference paper at ICLR 2027

Flow Maps for Trajectory Inference to Improve Learned Conditional Laws

10 views

Abstract

Trajectory inference seeks to reconstruct a stochastic process from unpaired samples of its temporal marginals. Since the true process is inherently unidentifiable from the limited observed data, encoding inductive biases into trajectory inference methods is essential to select a plausible approximate process. Many recent methods construct an interpolant process between observed marginals and try to match its probability path with a neural ODE or SDE, but little attention is paid to the learned joint distribution law across time. In many applications, there exists information on the characteristics of this law. In single-cell differentiation, for example, biological knowledge specifies plausible transitions between cell types, and we seek an efficient way to incorporate such knowledge. This would require constraining model-generated trajectories directly, which conventionally involves numerical integration and differentiation through the solver, making training computationally expensive. To address this, we propose CLIFT, a method using (stochastic) flow maps to efficiently impose differentiable constraints on learned conditional laws. A flow map provides parallelisable access to finite-time transitions, creating a natural interface for conditional-law losses whose gradients flow directly to the model parameters. Across single-cell datasets, CLIFT improves adherence to known biological differentiation rules while maintaining competitive marginal distributional fidelity and being significantly more efficient than imposing constraints through numerical rollouts.

Primary area:
generative models
Keywords:
Trajectory inference, flow maps, single cell data

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

In this paper, the authors first highlight a limitation of ODE/SDE approaches for trajectory inference that stems from their Markovian structure. However, their training interpolants are not guaranteed to be Markovian. They then define hierarchical processes, whose transitions between regions are constrained by a known graph, and show their relation to trajectory inference in single-cell data. Finally, they present a framework equipped with flow maps and a novel hierarchical process loss that encourages the trained model to take paths that are valid according to the provided prior knowledge about valid transitions.

Critical strengths
  • Strong motivation. The authors first show that the interpolant is not Markov by a bend depending on the endpoints and ODE/SDE approaches are Markov, the future depends on current state. An interpolant that carries lineage information is not guaranteed to be Markovian. ODE/SDE based approach can infer marginals yet still can take an invalid trajectory, jumping between sibling cell type regions as demonstrated by experiments. Hence, the authors motivate the constraining model and encourage it to produce valid trajectories.

  • Novel hierarchical loss. Cell evolution follows a hierarchical path. A cell belonging to a type cannot evolve into its parent or sibling cell type as time passes. The authors encourage the model to follow a valid hierarchical path by introducing a novel loss based on soft transition compatibility. They first train a classifier that predicts cell types and define a valid transition matrix. Then, they define the compatibility as the multiplication of the classifier's predicted type probabilities at the earlier and later states with the valid transition matrix. The proposed loss is the negative logarithm of this compatibility, where minimizing it pushes the transitions toward valid regions.

  • Runtime. The proposed method does not rely on costly step-by-step simulations by providing the transition with a single model call with flow maps, which makes it practically appealing.

Critical weaknesses
  • The following work is closely related and could be discussed and possibly compared against in terms of both performance and runtime: von Bassewitz, J. P., Pervez, A., Fumero, M., Robinson, M., Karaletsos, T., & Locatello, F. (2026). Learning explicit single-cell dynamics using ODE representations. In International Conference on Learning Representations.

  • The authors demonstrate their method mostly on relatively shallow and simple lineage processes. The paper could be strengthened by an additional dataset with a deeper and more complex lineage tree, for example: Calderon, D., Blecher-Gonen, R., Huang, X., Secchia, S., Kentro, J., Daza, R. M., ... & Shendure, J. (2022). The continuum of Drosophila embryonic development at single-cell resolution. Science, 377(6606), eabn5800.

  • Minor cosmetic issue: Table 5 in Appendix H overflows horizontally.

  • Line 7 of Algorithm 1 defines Lfate\mathcal{L}_\text{fate}, but line 8 uses Lcon\mathcal{L}_\text{con}, which is not defined.

Questions for authors
  • How would ablation variants that use the hierarchical loss without filtering perform? Could the authors add this variant to Table 2?
  • Could the authors provide a sensitivity analysis with respect to the weight of the hierarchical loss?
6 comments
AuthorsOpenReview ↗

Comparison with Cell-MNN

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.


(A) Comparison with Cell-MNN

We appreciate you pointing us to [1], it is indeed a relevant related work, and we have now discussed it and compared to it empirically. Cell-MNN is conceptually related to CLIFT, in that both methods approach trajectory inference using ODEs/SDEs, and both do so with interfaces that allow for sampling arbitrary time jumps along trajectories without numerical integration. Cell-MNN approximates the target process using locally linear ODEs in PCA space, that can be solved analytically to move forward in time, while our CLIFT method uses flow maps, that are a more expressive one-step operator than the linear ODE framework. Cell-MNN is primarily designed to model the marginal evolution of the target process well, in line with most existing trajectory inference methods, and they do not impose any constraints on the conditional laws of the model's process, which is our primary focus with CLIFT.

We have added Cell-MNN to our empirical comparisons across the three single cell datasets in Table 2, using the official implementation from [1] and the same hyperparameter optimisation process as CLIFT. We also include its results in comparison to CLIFT$_D$ here:

Maizels:

MethodViol. %W1W_1
Cell-MNN22.44±3.6922.44\pm3.699.98±0.079.98\pm0.07
CLIFT$_D$9.06±0.899.06 \pm 0.899.81±0.069.81 \pm 0.06

Cite:

MethodViol. %W1W_1
Cell-MNN17.61±0.4317.61\pm0.4349.04±0.3549.04\pm0.35
CLIFT$_D$13.16±0.4213.16 \pm 0.4243.96±0.1243.96 \pm 0.12

Multi:

MethodViol. %W1W_1
Cell-MNN23.77±0.2023.77\pm0.2062.63±0.4362.63\pm0.43
CLIFT$_D$18.98±0.2018.98 \pm 0.2053.69±0.3153.69 \pm 0.31

As we can see, CLIFT outperforms Cell-MNN in terms of both the Viol. %\% and W1W_1 metrics on all three datasets. The more expressive flow map framework seems to allow for better modelling of these cellular differentiation processes than the locally linear ODE approach of Cell-MNN. Nevertheless, in terms of runtime, across the Maizels, Cite, and Multi datasets, Cell-MNN training took on average 55, 1414, and 88 minutes respectively, so this method does offer computational gains over CLIFT.

Update: We have added a discussion of Cell-MNN to our Section 4, and added it to our empirical comparisons in Section 5.

[1] von Bassewitz, J. P., Pervez, A., Fumero, M., Robinson, M., Karaletsos, T., & Locatello, F. (2026). Learning explicit single-cell dynamics using ODE representations. In International Conference on Learning Representations.

AuthorsOpenReview ↗

CLIFT ablation without interpolant filtering

(B) CLIFT ablation without interpolant filtering

We have now evaluated an ablation where we use our hierarchical loss on a flow map, but without interpolant filtering. We have added these results to Table 2 of the revised manuscript, labelling this ablation as CLIFT$_{\mathrm{D}}$ (unfilt.) and CLIFT$_{\mathrm{S}}$ (unfilt.). We also report these results, in comparison to the base flow map ablations, and the full CLIFT method, here.

Maizels:

MethodViol. %W1W_1
OT-FMap23.67±1.3223.67 \pm 1.3210.12±0.0310.12 \pm 0.03
CLIFT$_\mathrm{D}$ (unfilt.)14.56±1.4214.56 \pm 1.4210.09±0.0310.09 \pm 0.03
CLIFT$_D$9.06±0.899.06 \pm 0.899.81±0.069.81 \pm 0.06
MethodViol. %W1W_1
OT-SSFM30.27±0.4030.27 \pm 0.4010.14±0.0110.14 \pm 0.01
CLIFT$_\mathrm{S}$ (unfilt.)31.06±0.6331.06 \pm 0.6310.06±0.0110.06 \pm 0.01
CLIFT$_S$15.67±1.3515.67 \pm 1.359.94±0.019.94 \pm 0.01

Cite:

MethodViol. %W1W_1
OT-FMap32.81±1.1232.81 \pm 1.1246.54±0.3146.54 \pm 0.31
CLIFT$_\mathrm{D}$ (unfilt.)27.94±1.8327.94 \pm 1.8346.27±0.1746.27 \pm 0.17
CLIFT$_D$13.16±0.4213.16 \pm 0.4243.96±0.1243.96 \pm 0.12
MethodViol. %W1W_1
OT-SSFM33.62±0.4833.62 \pm 0.4845.58±0.0845.58 \pm 0.08
CLIFT$_\mathrm{S}$ (unfilt.)24.38±0.3524.38 \pm 0.3544.95±0.0344.95 \pm 0.03
CLIFT$_S$15.87±0.3015.87 \pm 0.3044.40±0.0844.40 \pm 0.08

Multi:

MethodViol. %W1W_1
OT-FMap33.92±0.1533.92 \pm 0.1554.90±0.3754.90 \pm 0.37
CLIFT$_\mathrm{D}$ (unfilt.)29.79±0.7329.79 \pm 0.7355.06±0.2155.06 \pm 0.21
CLIFT$_D$18.98±0.2018.98 \pm 0.2053.69±0.3153.69 \pm 0.31
MethodViol. %W1W_1
OT-SSFM36.46±0.5436.46 \pm 0.5455.52±0.0355.52 \pm 0.03
CLIFT$_\mathrm{S}$ (unfilt.)25.53±0.2825.53 \pm 0.2855.01±0.0555.01 \pm 0.05
CLIFT$_S$18.85±0.3218.85 \pm 0.3254.12±0.0854.12 \pm 0.08

From these comparisons, we can generally see, across both the deterministic and stochastic settings, that the CLIFT (unfilt.) ablation tends to improve the Viol % and W1W_1 metrics compared to the base OT-FMap and OT-SSFM variants, but it is generally not as good on either metric as the full CLIFT method.

These results show that the hierarchical loss can notably improve modelling performance without any alterations to the interpolants. Nevertheless, the best performance are obtained when combining this loss with the filtered interpolants, which is to be expected as this involves more sophisticated tailoring of the model using the prior knowledge available about the ground-truth process's hierarchy.

Updates: We have added this additional ablation and the relevant discussion to Table 2 and Section 5.

AuthorsOpenReview ↗

Sensitivity analysis of the hierarchical loss weight

(C) Sensitivity analysis of the hierarchical loss weight

The choice of the weight for the hierarchical loss, λh\lambda_h is an important design choice when training CLIFT. Indeed, this is one of the hyperparameters that we optimised over a grid search on the Maizels dataset, by considering the marginal reconstruction on two held-out validation timepoints. We now include these results, in terms of the Viol. %\% and W1W_1 achieved across this hyperparameter sweep, to see how sensitive the performance of CLIFT is to this setting. For the deterministic setting, we swept over λh∈{0,100,350,1000}\lambda_h \in \{0,100, 350, 1000\}, and for the stochastic setting we swept over λh∈{0,1,10,100}\lambda_h \in \{0,1,10,100\}. We list the results for each setting in the table below, and include them in our Appendix L.

λh\lambda_hViol. (%)Test W1W_1Validation W1W_1
015.55±0.9215.55 \pm 0.929.96±0.029.96 \pm 0.0210.69±0.0210.69 \pm 0.02
10013.20±1.5413.20 \pm 1.549.91±0.029.91 \pm 0.0210.65±0.0210.65 \pm 0.02
35011.33±0.7311.33 \pm 0.739.92±0.039.92 \pm 0.0310.67±0.0410.67 \pm 0.04
10009.06±0.899.06 \pm 0.899.81±0.069.81 \pm 0.0610.51±0.0710.51 \pm 0.07
150010.23±0.9210.23 \pm 0.929.94±0.039.94 \pm 0.0310.69±0.0210.69 \pm 0.02
λh\lambda_hViol. (%)Test W1W_1Validation W1W_1
022.06±1.0022.06 \pm 1.009.96±0.019.96 \pm 0.0110.54±0.0110.54\pm0.01
117.03±0.6817.03 \pm 0.689.95±0.019.95 \pm 0.0110.54±0.0110.54 \pm 0.01
1015.67±1.3515.67 \pm 1.359.94±0.019.94 \pm 0.0110.52±0.0110.52 \pm 0.01
10048.28±1.6248.28 \pm 1.6211.89±0.0611.89 \pm 0.0612.52±0.1812.52 \pm 0.18

As we can see from the above, there is an optimal operating point in the middle of the sweeps for both settings. At low constraint weights, the Viol. %\% tends to be higher than this optimal point, as does the test and validation W1W_1, while at too high constraint weights, performance degrades again. There seems to be a strong correlation between both metrics as well, so, in this experiment at least, a user does not have to trade-off between optimising one or the other, as both metrics share the same optimum in both settings.

Update: We have added tables with results across different hierarchy weights for the Maizels experiment in Appendix L.


(D) Further dataset

Thank you for pointing us to this dataset, it does indeed appear to offer a more complex lineage tree setting that would be useful to investigate. While we have not yet had time to run experiments on it, we will try tto do so in the remaining discussion time.


Thank you again for your time and effort in reviewing our submission. We hope that the above amendments have addressed your concerns. We are more than happy to engage with any further comments you may have. We have also corrected the table overflow and the notation inconsistency in Algorithm 1.

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

This work studies trajectory inference from temporal population snapshots without paired samples across time. The authors propose CLIFT, a method that adapts flow maps to incorporate known cell-type lineage relationships into learned transitions. The maps are trained using standard flow map objectives with a hierarchical loss that uses a fixed classifier to penalize transitions violating a prescribed lineage graph. Direct access to finite-time transitions allows this loss to be optimized without differentiating through numerical integration. Experiments on synthetic and single-cell datasets report improved conditional predictions and fewer lineage violations while maintaining competitive marginal distributional accuracy, with lower training cost than imposing the constraint through numerical rollouts.

Critical strengths

The paper demonstrates a practical application of flow maps for imposing lineage constraints without differentiating through numerical integration. Table 3 reports a training-time reduction from 65 to 8 minutes on stochastic CITE. Table 1 reports improved conditional predictions, while Table 2 shows fewer lineage violations beyond filtering alone.

Critical weaknesses
  1. The work’s methodological contribution is modest. Flow maps have already been applied to cellular trajectory inference [1], and CLIFT’s main addition is a classifier-based lineage penalty on finite-time transitions. While useful, reduced lineage violations are an expected benefit of this penalty. The missing comparison with OTP-FM [1] would help establish CLIFT’s marginal accuracy and training efficiency relative to an existing flow-map approach.
  2. For Maizels evaluation, Appendix I.1 states that D3.4 and D6 for hyperparameter selection (lines: 1813-1815) but also includes them in evaluation (lines: 1817-1819). If Table 2 includes these days, its W1 mixes validation and test performance, obscuring generalization beyond the times used for model selection. This could also favor CLIFT over baselines not tuned on those days.
  3. The main body should include a conclusion that summarizes the findings and discusses their limitations (Appendix A essentially).

[1] Raghav Kansal, David Crair, Nghia Nguyen, Scott Pope, and Bradley Parry. “Multimarginal Flow Matching with Optimal Transport Potentials.” 2026. arXiv:2606.05327.

Questions for authors
  1. What solvers and integration-step budgets were used to train the constrained ODE/SDE baselines in Table 3? What explains the big training-time difference between CLIFT_D and CLIFT_S? Could differences in model capacity or training budgets also explain part of the performance gains in Table 2?
  2. How sensitive is performance to the inference budget? The single-cell experiments use 50 integration steps. Performance curves comparing CLIFT, the corresponding unconstrained flow maps, and ODE/SDE baselines would clarify whether accuracy has saturated and how well CLIFT performs with fewer evaluations. An example sweep is 1, 2, 4, 8, 16, 32, and 64 steps per observed interval.
  3. How sensitive is CLIFT to imperfect cell-type annotations? Authors could corrupt 0%, 5%, 10%, and 20% of synthetic training labels, retrain the classifier, and compare CLIFT with filtering-only training under the same corrupted annotations, evaluating against clean ground truth.
  4. Could the authors report a curve of the marginal W1 and lineage violations for the hierarchy weights {0, 1, 10, 100} considered for CLIFT_S in Appendix I.3? This would clarify the trade-off between marginal accuracy and lineage adherence.

Suggestions

  1. In Algorithm 1 (line: 1627), I believe there is a typo in the diagonal loss: Φθ(r,r,Ir)\Phi_\theta(r,r,I_r) should be replaced with the instantaneous velocity aθ(r,r,Ir)a_\theta(r,r,I_r). Also, Corollary 2 should say “transform the SDE into an ODE” (lines: 1389-1390).
  2. In Algorithm 1, Lfate\mathcal{L}{\mathrm{fate}} in step 7 should be Lcon\mathcal{L}{\mathrm{con}}, consistent with step 8. Also, just replace C(x,y)\mathcal{C}(x,y) with −log⁡qvalid(x,y)-\log q_{\mathrm{valid}}(x,y) explicitly.
6 comments
AuthorsOpenReview ↗

Novelty (1/2)

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.


(A) Novelty

Thank you for pointing us to the OTP-FM work, it is indeed a relevant related work, and we had not seen it previously. Given this work proposes a method that also utilises flow maps for trajectory inference, we agree that it is important to distinguish our work from it, both conceptually and empirically. We have also changed our wording in the introduction and related works sections, to avoid saying that flow maps in trajectory inference are "unexplored".

Although both works use flow maps, the aim of the OTP-FM paper differs from ours. OTP-FM is positioned similarly to many other trajectory inference works, in that they intend to construct a model for a process that matches it well in terms of its marginal evolution. Our work, on the other hand, asks a different question: can we learn a joint law more truthful to the true process, such that individual trajectories adhere to known transition constraints? These two different objectives lead to different algorithms that utilise the flow map interface in different manners. OTP-FM explores relaxing the dynamical optimal transport problem, to introduce soft constraints, in place of hard constraints, to fit intermediate marginals. Consequently, it changes the interpolant velocity targets in the iMF objective using potential-driven corrections (which can be calculated using samples from the flow map, depending on the potential), while our method introduces an additional objective calculated on and backpropagated through model-generated transitions.

Our algorithm is motivated by our theoretical investigation on how to enforce known constraints on the joint law. In this, we show why the joint law of an interpolant will not necessarily be well learned using matching objectives, encouraging us to impose these constraints directly on the model. Our experiments confirm this, when comparing the OT-FMap (filt.) vs. CLIFT variants, which train on the same interpolants that adhere to the lineage prior, as CLIFT outperforms, demonstrating the additional benefit of imposing that prior directly to model predictions. In doing so, we step away from the common approach in a large amount of related literature of designing interpolants to encode inductive biases on the model. Changing the learning objectives of the model resembles problems in the stochastic optimal control literature, which we elaborate on in Appendix D.

Our results show that CLIFT learns conditional laws that better respect the prescribed lineage constraints, while remaining competitive in marginal reconstruction to baselines. In many cases, CLIFT even improves marginal reconstruction relative to methods trained with the same interpolant targets. Although Lh\mathcal L_h is designed to reduce lineage violations, and, as you state, this is an expected benefit of our model, this reduction remains meaningful. A model that frequently predicts biologically inadmissible transitions has limited utility for studying individual cellular trajectories, even if it accurately reproduces population-level distributions. Moreover, the training objective Lh\mathcal L_h and Viol. % evaluation metric are distinct. The training loss uses a classifier fitted only to the observed training marginals, whereas evaluation uses a separate classifier trained across all available time points, including those withheld from trajectory-model training. Evaluation also measures discrete violations along trajectories starting from held-out cells. The reported reduction in Viol. % by CLIFT models therefore reflects that Lh\mathcal L_h can generalise lineage adherence to this more accurate held-out cell-type classifier at unseen cell positions. Again, these gains can be achieved while maintaining competitive marginal fidelity, and often performing best amongst a suite of baselines, particularly in the stochastic setting. The synthetic and LARRY experiments give further evidence that CLIFT learns more faithful conditional distributions over cell futures, beyond reducing the specific lineage violations penalised during training, as it reduces the distance between the learned and ground-truth conditional laws, and improves clone-conditional Wasserstein distance.

AuthorsOpenReview ↗

Novelty (2/2) and more

Another point of novelty to our work is that we show that strong stochastic flow maps are capable methods for trajectory inference, extending flow map applications here to the stochastic setting. It is often more desirable to have a trajectory inference model that can estimate a distribution over the future of a given particle, rather than condensing this to a deterministic trajectory. In biological contexts, this can allow the model to capture heterogeneity amongst similar cells, and allow for applications such as the estimation of cellular fate probabilities. OTP-FM, on the other hand, is a deterministic method.

We have now added an empirical comparison to OTP-FM in Table 2, and we report it in comparison to CLIFT$_D$ here as well. In terms of runtime, we see that OTP-FM trains in approximately 7, 7, and 8 minutes on Maizels, Cite, and Multi, respectively. This is faster than CLIFT$_D$, which is largely slowed by the calculation of the filtered interpolants, as further elaborated on in response (D). CLIFT$_S$ is largely competitive with these OTP-FM runtimes.

MethodMaizels Viol. %Maizels W₁Cite Viol. %Cite W₁Multi Viol. %Multi W₁
OTP-FM19.45 ± 0.489.82 ± 0.0925.92 ± 0.2845.76 ± 0.2228.89 ± 0.6253.19 ± 0.23
CLIFT_D9.06 ± 0.899.81 ± 0.0613.16 ± 0.4243.96 ± 0.1218.98 ± 0.2053.69 ± 0.31

We can see that OTP-FM is a competitive method with CLIFT$_D$ in terms of W$_1$ in two of the three datasets, and it is much worse in terms of Viol. % on all, as expected.

Update: We have expanded our Section 4 and Section 5 to discuss and compare with OTP-FM.


(B) Leakage from hyperparameter selection on Maizels data

Thank you for raising this important point that helps improve the soundness of our results. The W1W_1 results initially displayed in Table 2 did indeed average over all non-training days, including the two days used for hyperparameter selection, and this is a form of information leakage. However, we do note that all methods, include CLIFT, the ablations, and all baselines, were tuned on the Maizels data in the same manner, so this leakage was unlikely to preferentially favour CLIFT. We have updated our Maizels evaluation now to exclude the timepoints used for hyperparameter optimisation, and we have updated the Maizels W1W_1 column in Table 2 with the new results.

With this update, the takeaways on this data remain similar to before. In terms of W1W_1, CLIFT$_D$ and CLIFT$_S$ score best amongst the deterministic and stochastic methods, respectively, with clear improvements on top of the base flow map and the filtering-only ablations.

Update: We have edited our results in Section 5 for the Maizels experiment to exclude the held-out days used for hyperparameter selection from the mean W1W_1 evaluation metric.


(C) Discussion

We have now moved our discussion to the main body of the manuscript, as Section 6. It recaps the takeaways of the work, and highlight its limitations and potential future work directions.


AuthorsOpenReview ↗

Training details

(D) Training details

For the constrained ODE/SDE models we compare CLIFT to, rollouts are used to calculate Lh\mathcal L_h. To compute Lh\mathcal L_h, the constrained ODE baseline uses explicit forward Euler integration, and the constrained SDE baseline uses differentiable Euler–Maruyama-style rollouts of its learned local model. Both are implemented directly in JAX. We use a maximum integration step size of 0.01 on the globally normalised time axis: a transition from ss to tt uses N=max⁡(1,⌈(t−s)/0.01⌉)N=\max(1,\lceil(t-s)/0.01\rceil) equally spaced steps. This corresponds to 100 steps for a full transition from t=0t=0 to t=1t=1. Gradients are backpropagated through the discrete rollout.

Table 3 was intended to compare CLIFT with rollout-based methods, rather than report fully optimised runtimes. The results show that CLIFT is much faster than these methods. We investigated why CLIFT$_D$ trains more slowly than CLIFT$_S$ and identified two main factors that increase the training time per step for CLIFT$_D$. First, interpolant filtering is more computationally expensive for CLIFT$_D$. For each mini-batch, we compute the OT coupling and then check whether the resulting interpolants violate the hierarchical rules using classifier values along multiple points along their trajectories. In the stochastic case, we only check whether the paired endpoints represent an admissible transition according to the hierarchy. This difference in computing the filtering interpolant accounts for around 50% of the deterministic per-step runtime. We also found that the deterministic implementation computes diagnostic metrics more frequently. Reducing this overhead can save a further 20% of the time per step.

Beyond these overheads, CLIFT$_D$ and CLIFT$_S$ build on different flow-map methods and their official implementations. CLIFT$_D$ uses the deterministic flow map of Boffi et al. (2026), whose Lagrangian self-distillation loss requires backpropagating through the flow map’s time derivative. CLIFT$_S$ uses the strong stochastic flow map of McCallum et al. (2026), whose semigroup loss avoids this additional derivative. Their different objectives may also lead to different convergence behaviour, although both use early stopping with the same patience.

Architectures are matched within the deterministic and stochastic flow-map ablations. CLIFT$_D$ and its FM/FMap baselines use three hidden layers of width 1024. CLIFT$_S$ and its SSFM/SDE baselines share a three-layer width-512 drift network, a two-layer width-512 diffusion network, and a small uncertainty network. CLIFT adds no trainable parameters to the corresponding flow model; therefore, increased model capacity cannot explain its improvements over these baselines. Architectures differ across the broader set of methods in Table 2, although we attempt to keep model capacities similar. OT-MFM, MMFM and MM-GAN also use velocity networks with three hidden layers of width 1024, OTP-FM uses a flow map network with three hidden layers of width 1024, while SF²M and BranchSBM use multiple networks. These comparisons do not give CLIFT a systematic advantage in parameter count.

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

The paper discusses the limitations of current flow-based approaches to trajectory inference and proposes a new algorithm that approximates the dynamics of the underlying process with flow maps.

Critical strengths

I believe the main strength of the paper is its discussion of the possible limitations of ODE/SDE-based trajectory inference algorithms, supported by an attempt to explain these limitations theoretically. In addition, the authors propose a flow-map-based algorithm for trajectory inference.

Critical weaknesses
  1. The presentation of the theory in Sections 3.1 and 3.2 is somewhat unclear. I struggle to see how this theory relates to a) real-world processes and b) the algorithm presented later. I suggest first giving an informal overview of the idea with a schematic example, and only then moving to a more rigorous theoretical exposition. Moreover, it is not clear how the theory motivates the training procedure of CLIFT. Finally, I do not think the paper explains why the proposed algorithm, which uses flow maps, does not suffer from the same problems as SDE/ODE-based methods.

  2. Trajectory inference methods typically require first defining an interpolant together with its training procedure, and only then training a flow model on this interpolant. If I understand correctly, this paper proposes using flow maps in place of flow matching models. However, I could not find an explicit discussion of how the interpolant is trained. For instance, Kapuśniak et al. (2024) use metric learning, Rohbeck et al. (2025) use splines, and Kviman et al. (2026) use GANs. I suggest discussing these different ways of training an interpolant and explaining whether your method can be combined with each of them. Such a discussion could significantly improve the clarity of the paper.

  3. Table 2 presents results on several benchmarks in terms of Vol% and W1W_1. However, there is no clear evidence of an advantage of CLIFT in terms of W1W_1. It is also difficult to understand the importance of the Vol% metric, since it is not a standard metric and is not sufficiently explained: lines 414–417 are, unfortunately, not enough to fully appreciate its significance.

Questions for authors

Suggestions:

  1. Please add pseudocode for the proposed algorithm to the main text. Moreover, improving the writing of Section 3 would greatly benefit the clarity of the paper. I suggest providing an intuitive explanation of the theory, linking it directly to the algorithm, and describing the algorithm itself in more detail.

  2. Please justify the importance of the Vol% metric by explaining what it measures in practice.

  3. I also suggest including a discussion of Rohbeck et al. (2025) and Kviman et al. (2026) in the paper, together with a numerical comparison of the proposed algorithm with these methods.

5 comments
AuthorsOpenReview ↗

Overview (1/2)

Thank you for your thoughtful comments and suggestions, which have helped us significantly improve our manuscript. Please see the revised PDF, where the changes are written in blue. We also directly respond to your points below.


(A) Presentation of theory

Trajectory inference aims to model the path law PP of a stochastic process (Xt)t∈[0,1](X_t)_{t\in[0,1]} using samples from some KK marginals ptk,1≤k≤Kp_{t_k}, 1 \leq k \leq K.. The objective is to use this sparse information to propose a good model P^\hat P. Such a model can be used to understand both the macroscopic evolution of the process, through the marginals p^t\hat p_t, and the future of specific particles, through the joint laws of P^\hat P. It is this second use case that our work is primarily concerned with. The model's prediction for the future of a given particle x0x_0 (for example, a stem cell) is governed by the conditional law P^( ⋅∣X0=x0)\hat P(\,\cdot \mid X_0 = x_0). Modelling this law properly is what allows the prediction of which fates a given cell can reach and with what probability. We investigate how prior knowledge on the joint laws of PP can be used to improve this aspect of P^\hat P.

We argue that information about the joint laws of PP can often be available, especially in the prominent trajectory inference setting of studying cell differentiation. Here, single-cell transcriptome datasets describe the differentiation of stem cells into diverse cell types, and it is often known which transitions between types are admissible or inadmissible. Differentiation is commonly pictured through Waddington’s landscape and summarised by lineage trees, which define hierarchical relationships between progenitor cells and their specialised progeny [1,2,3]. An endoderm progenitor cell that starts to differentiate towards, for example, a liver cell type can specialise further into sub-types of liver cell, but it cannot become a brain cell. Lineage trees therefore imply that once a cell is in a certain region of gene expression space, it cannot move to regions that the tree makes inadmissable. We consider cases when we have this kind of information about the joint law of PP, and we want to use it to improve our model P^\hat{P}.

We formalise these concepts through the definition of hierarchical processes (Definition 1). Informally, these are stochastic processes whose state space is partitioned into regions related by a directed acyclic graph of admissible transitions. Once a particle from a hierarchical process has entered a region, it can only stay there or move to that region’s descendants in the graph. As such, knowledge about the graph dictating a hierarchical process is coarse-level knowledge about that process's joint laws (i.e. knowledge of restrictions on which states can be reached from a given state). We consider how to model such hierarchical processes when a user does have prior knowledge about the graph of admissable transitions. The definition applies to any system whose states are organised in such a hierarchy, and it was motivated by the cellular differentiation setting.

A natural question is then: why not use the usual trajectory inference approach to incorporate this prior knowledge on the joint law? As you noted, the usual pipeline involves choosing a coupling π\pi between the marginals, designing a stochastic interpolant ItI_t between coupled samples (x0,x1)∼π(x_0,x_1)\sim\pi, and training a neural velocity field by flow matching, using the interpolant’s velocity as the regression target. The model P^\hat P is then defined by the ODE/SDE started from p0p_0 and driven by the learned velocity field. At the optimum of this objective, P^\hat P has the same marginal laws as the interpolant. However, there is no guarantee that P^\hat P has the same joint law as the interpolant. In other words, prior knowledge about the joint law of PP encoded in the interpolant is not necessarily transmitted to the learned model P^\hat P. This is what we show in Proposition 1.

Continued in Overview (2/2).

AuthorsOpenReview ↗

Overview (2/2)

We therefore argue that encoding prior knowledge of the joint law in the interpolant alone is not sufficient. We propose instead to act directly on the neural ODE/SDE model, after the interpolant has been constructed, by penalising model-generated trajectories that violate the prior knowledge on the ground truth joint law. This is the idea behind CLIFT. Orthogonal to choosing an interpolant, which can be linear or more sophisticated, we add a training objective for the velocity field that encourages the process governed by P^\hat P to respect the same hierarchy as the ground truth process. The hierarchical loss Lh\mathcal L_h of Definition 2 is designed so that a model with a small loss has a small probability of forbidden transitions according to this graph. Proposition 2 and Corollary 1 show that the probability that a transition (Xs,Xt)(X_s, X_t) of the model violates the hierarchy is bounded by this loss. If the hierarchical loss is close to 00 (and the required classifier is well trained), model transitions that violate the hierarchy are therefore rare.

Flow maps are what make calculating and optimising Lh\mathcal L_h feasible. Sampling trajectories from a neural ODE or SDE model can require many solver steps and Lh\mathcal L_h must then be backpropagated through all of them. This is computationally intensive and can make training less stable, as demonstrated by our results in Tables 2 and 3. The flow map interface permits direct access to these finite-time transitions of the model, permitting efficient and effective optimisation with this objective. Furthermore, we show that this training scheme does improve adherence to the specified hierarchy better than simply designing the interpolant process to have only interpolants consistent with the graph (via filtering) in Tables 2 and 4.

[1] Waddington, C.H., 2014. The strategy of the genes. Routledge.

[2] Saez, M., Briscoe, J. and Rand, D.A., 2022. Dynamical landscapes of cell fate decisions. Interface focus, 12(4), p.20220002.

[3] Weinreb, C., Rodriguez-Fraticelli, A., Camargo, F.D. and Klein, A.M., 2020. Lineage tracing on transcriptional landscapes links state to fate during differentiation. Science, 367(6479), p.eaaw3381.

Update: We have revised Section 3 to more clearly introduce our setting of interest before formalising it. We have followed the structure you suggested, first giving an informal motivating overview of the types of biological processes we are interested in, then moving to the theoretical exposition.

AuthorsOpenReview ↗

Interpolants

(B) Interpolants

As you pointed out, trajectory inference methods require defining a stochastic interpolant to act as the training target. In our work, we only considered using a linear interpolant, or a filtered version of the linear interpolant, where, during flow training, any interpolants that violate the specified hierarchy are rejected and re-sampled. Consequently, there is no training phase for the interpolants CLIFT is trained on. This is primarily because of the intuition we derived from our Proposition 1: that if we want to influence properties of the joint laws of P^\hat P, interpolant design is not generally sufficient.

Nevertheless, since the interpolant used in training CLIFT is a design choice that can be changed, we have now run additional experiments using learned interpolants as well, as you suggested. Specifically, we have used the interpolant processes from Rohbeck et al. (2025) and Kviman et al. (2026) to replace the filtered linear interpolants in CLIFT training. We train deterministic and stochastic flow maps using CLIFT's objective, including the hierarchical loss on model-generated transitions, on these interpolants, and we report these results in Appendix A. Furthermore, as suggested, we compare numerically to the full methods of Rohbeck et al. (2025) and Kviman et al. (2026), which train velocity fields on their interpolant targets in Table 2. We also report all of these results below.

MethodMaizels: Viol. % (↓)Maizels: W₁ (↓)Cite: Viol. % (↓)Cite: W₁ (↓)Multi: Viol. % (↓)Multi: W₁ (↓)
Deterministic Methods
MMFM23.69±1.4723.69 \pm 1.4712.82±0.0412.82 \pm 0.0432.82±0.3032.82 \pm 0.3049.58±0.1449.58 \pm 0.1435.01±0.3035.01 \pm 0.3060.46±0.1160.46 \pm 0.11
MM-GAN32.50±3.1332.50 \pm 3.1311.28±0.0711.28 \pm 0.0728.28±0.7128.28 \pm 0.7147.87±0.0947.87 \pm 0.0933.52±0.4833.52 \pm 0.4853.49±0.2053.49 \pm 0.20
MMFM + CLIFT$_{\text{D}}$8.38±0.608.38 \pm 0.6012.58±0.0712.58 \pm 0.0714.06±0.2614.06 \pm 0.2647.92±0.1747.92 \pm 0.1718.10±0.2818.10 \pm 0.2858.50±0.2158.50 \pm 0.21
MM-GAN + CLIFT$_{\text{D}}$17.92±1.2017.92 \pm 1.2011.02±0.0111.02 \pm 0.0124.47±0.6724.47 \pm 0.6747.66±0.0847.66 \pm 0.0831.48±0.3731.48 \pm 0.3753.52±0.1553.52 \pm 0.15
Stochastic Methods
MMFM + CLIFT$_{\text{S}}$9.77±0.629.77 \pm 0.6212.47±0.0212.47 \pm 0.0213.74±0.3413.74 \pm 0.3448.24±0.0548.24 \pm 0.0515.66±0.2115.66 \pm 0.2158.55±0.0358.55 \pm 0.03
MM-GAN + CLIFT$_{\text{S}}$24.13±0.8524.13 \pm 0.8510.84±0.0510.84 \pm 0.0527.30±0.2027.30 \pm 0.2048.29±0.0448.29 \pm 0.0433.02±0.2133.02 \pm 0.2154.21±0.1554.21 \pm 0.15

For both MMFM and MM-GAN, combining their interpolant constructions with CLIFT reduces lineage violation rates relative to the original baselines across all three datasets, with substantial reductions in several settings. These combinations also yield lower W1W_1 distances in most configurations, suggesting that CLIFT can improve both adherence to known biological lineage constraints and marginal reconstruction. These results support the compatibility of CLIFT with alternative interpolant constructions and suggest that directly regularising the learned conditional laws can provide additional benefits.

Updates: We have added additional experiments using learned interpolants in CLIFT training, and comparing to the methods of Rohbeck et al. (2025) and Kviman et al. (2026), in Table 2 and Appendix A.

Comments on the paper

1 comment

Dear AC,

Thank you very much for your efforts in overseeing the review of our work. We are writing to summarise our discussion period. We have made a significant effort to address each point raised by the reviewers, and we believe in doing so our paper has substantially improved. Over the discussion period, we have made the following key changes to the manuscript:

  • Additional baselines. We have added empirical comparisons to more trajectory inference models, including OTP-FM, Cell-MNN, MMFM, and MM-GAN, in Section 5 and Appendix A.
  • Additional ablations and sensitivity analyses. We have compared to a further ablation of CLIFT, without interpolant filtering, in Section 5, and examined performance across constraint weights, data corruption, and sampling budgets in Appendices L, M, and N, respectively.
  • Additional dataset. We have conducted an additional experiment on a new Drosophila development dataset, with a more complex lineage graph, in Appendix O.
  • Revised Section 3. We have made changes to the presentation of Section 3, by introducing the cellular development setting we are interested in before our formalisation of these processes as hierarchical processes. We have also added a pseudocode algorithm for CLIFT to the main text.
  • Added a discussion section. We have moved our discussion from the appendix to Section 6.

We are glad that Reviewer nkJe has appreciated these changes, and increased their rating to a 4. Unfortunately, Reviewers wWjy and pvpx have not yet acknowledged our responses, so we will not have any chance for further clarifications. We believe that we have thoroughly addressed their points, and we hope that you agree. Thanks again for your time and effort in this process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.