acceptodds
Under review as a conference paper at ICLR 2027

Extrapolation-Aware Predictions with Flow Matching

13 views

Abstract

Flow Matching (FM) is a leading approach for probabilistic prediction in scientific and control applications. When conditioned on inputs outside the training support, however, FM models extrapolate, yielding untrustworthy predictions that cannot be distinguished from valid ones. We trace this to the Lipschitz regularity of neural vector fields, which keeps transport energy nearly constant around the data manifold, so off-manifold inputs follow regular transport paths that produce plausible predictions. We introduce Diverging Flows, a training objective that preserves near-optimal transport for on-manifold conditions while making transport deliberately inefficient for off-manifold ones. We show that the resulting energy margin lower-bounds trajectory curvature, so extrapolations can be detected from the sampling trajectory itself, without auxiliary models, with a threshold calibrated by conformal prediction. Across low-dimensional synthetic manifolds, scientific regression on and fields, and disjoint-domain generation on images, a single model reliably detects extrapolations while retaining FM's predictive fidelity, whereas on the same data FM likelihood is near chance.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
flow matching, extrapolation detection, trustworthy predictions, probabilistic regression

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_2Cf2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors introduce Diverging Flows, a training objective that preserves near-optimal transport for on-manifold conditions while making transport deliberately inefficient for off-manifold ones. Compared with the classical FM, an FM model can be trained to exhibit a sharp geometric transition, where valid conditions yield near optimal paths while off-manifold queries yield strongly curved ones. This paper is interesting, the thoeretical theorem and numerical implementations need further improvement.

Critical strengths
  1. This paper is well writen, which is easy to follow.
  2. The theoretical results are interesting, which inspires the model formulation.
Critical weaknesses
  1. Strong requirements about the paths for Proposition 4.1. In Appendix D.1 Equation (8), does this formula Et,xt∥vθ(xt,c)−uop∥−Et,xt∥vθ(xt,c~)−uOT∥=Et,xt[∥vθ(xt,c)−uop∥−∥vθ(xt,c~)−uOT∥]\mathbb E_{t,x_t}\|v^\theta(x_t,c)-u^{op}\|-E_{t,x_t}\|v^\theta(x_t,\tilde c)-u^{OT}\|=\mathbb E_{t,x_t}[\|v^\theta(x_t,c)-u^{op}\|-\|v^\theta(x_t,\tilde c)-u^{OT}\|] hold? If so , we need to ensure the same path for different conditions cc and c~\tilde c.

  2. In Proposition 4.1, the gradient Lipchitz continuity is very strong.

  3. The comparisons in numerical results are not sufficient.

Questions for authors
  1. Strong requirements about the paths for Proposition 4.1. In Appendix D.1 Equation (8), does this formula Et,xt∥vθ(xt,c)−uop∥−Et,xt∥vθ(xt,c~)−uOT∥=Et,xt[∥vθ(xt,c)−uop∥−∥vθ(xt,c~)−uOT∥]\mathbb E_{t,x_t}\|v^\theta(x_t,c)-u^{op}\|-E_{t,x_t}\|v^\theta(x_t,\tilde c)-u^{OT}\|=\mathbb E_{t,x_t}[\|v^\theta(x_t,c)-u^{op}\|-\|v^\theta(x_t,\tilde c)-u^{OT}\|] hold? If so , we need to ensure the same path for different conditions cc and c~\tilde c. Further explanations are needed.

  2. In Proposition 4.1, the gradient Lipchitz continuity is very strong, which is widely-used in convex optimization. Could you give some explanations?

  3. The comparisons in numerical results are not sufficient. The comparisons mainly focus on likelihood-based and ensemble-based methods. It is better to add more comparisons, e.g., training a separator classifier.

1 comment

We thank the reviewer for finding the paper well written and the theoretical results interesting. Since W1–W3 coincide with Q1–Q3, we answer them together.

In Appendix D.1 Equation (8), does this formula [...] hold? If so, we need to ensure the same path for different conditions cc and c~\tilde{c}. (W1, Q1)

The equality holds, and both energies are taken on the same path. The path xt=(1−t)x0+tx1x_t = (1-t)x_0 + t x_1 and its velocity utOT=x1−x0u_t^{OT} = x_1 - x_0 (Section 4.1, before Eq. 2) depend on the condition only through the pairing of x1∼pdata(⋅∣c)x_1 \sim p_{data}(\cdot \mid c) with cc. An off-manifold c~\tilde{c} has no data of its own, so its energy is evaluated on the reference path of the nearby valid cc, and only the condition argument of vθv^\theta changes. The margin losses of Section 4.2 are defined this way, and each boundary sample keeps the pair (x1,c)(x_1, c) of the valid cc it comes from (Section 4.3). With μc\mu_c the law of (t,x0,x1)(t, x_0, x_1) and fc′=∥vtθ(xt,c′)−utOT∥22f_{c'} = \Vert v^\theta_t(x_t, c') - u_t^{OT}\Vert_2^2, linearity under μc\mu_c gives ∣E(c~)−E(c)∣≤Eμc∣fc~−fc∣≤2K(M+U)∥c~−c∥2|\mathcal{E}(\tilde{c}) - \mathcal{E}(c)| \le \mathbb E_{\mu_c}|f_{\tilde{c}} - f_c| \le 2K(M+U)\Vert\tilde{c} - c\Vert_2 (App. D.1). The DOT score needs no reference path, since it compares each trajectory with the straight line between its own endpoints (Section 4.4). The revised PDF states the reference measure after Eq. 2 and in App. D.1.

In Proposition 4.1, the gradient Lipchitz continuity is very strong, which is widely-used in convex optimization. (W2, Q2)

Proposition 4.1 does not assume a Lipschitz gradient. Its statement only requires the vector field to be "locally Lipschitz continuous with respect to the conditioning variable cc" (lines 182–184 of the submission), that is, ∥vtθ(x,c~)−vtθ(x,c)∥2≤K∥c~−c∥2\Vert v^\theta_t(x, \tilde{c}) - v^\theta_t(x, c)\Vert_2 \le K\Vert\tilde{c} - c\Vert_2. The gradient-Lipschitz (LL-smooth) assumption of convex optimization constrains the gradient of an objective, and no gradient of vθv^\theta appears in Proposition 4.1 or its proof (App. D.1). As stated before the proposition (Section 4.1), FM already needs this regularity in xx for a unique ODE solution [3, 4], and standard architectures satisfy it as compositions of locally Lipschitz layers [1, 2], for any size of KK. The proposition explains silent extrapolation, which Diverging Flows counteracts by inducing a locally larger Lipschitz constant around M\mathcal{M} (Section 4.2).

It is better to add more comparisons, e.g., training a separator classifier. (W3, Q3)

Diverging Flows needs only the predictor, yet it outperforms the separate classifier the reviewer suggests (new in the revised PDF) and matches detectors that need their own model or the training data. The submission already compares with a classification head (App. F.1), DiffPath [5] (App. L.1) and the acceleration proxy of FM [6] (App. L.2). We add a separate binary classifier, trained with BCE on the same boundary samples, deep KNN [7] on the raw conditions, and VOS [8], which needs class labels and so runs only on MNIST. We report AUROC, and ★ marks results new in the revised PDF (App. M).

MethodExtra needERA5FMNISTKMNISTMean
Acceleration proxy [6]none0.5710.592★0.545★0.569
Classification headextra head0.530.903★0.831★0.755
Separate classifier★second network0.8600.9500.7750.862
VOS★class labels, second network–0.8330.977–
Deep KNN★training set at test time0.8730.9790.9850.946
DiffPath-1Dsecond diffusion model0.9730.9810.8630.939
DiffPath-6Dsecond diffusion model0.9930.9910.9290.971
Diverging Flowsnone0.9950.9870.8920.958

Trained on the same boundary samples, the separate classifier and the classification head stay well below Diverging Flows, so these samples alone do not explain the detection. Diverging Flows is above deep KNN and DiffPath-1D on average and close to DiffPath-6D. On KMNIST, deep KNN and VOS are stronger, but they need the training set or class labels, while Diverging Flows reads its score from the trajectory of the prediction itself, at ~0.8% overhead (Section 4.4).

We hope these clarifications and baselines lead the reviewer to reconsider the score.

References

[1] Virmaux and Scaman. Lipschitz Regularity of Deep Neural Networks. NeurIPS 2018

[2] Kim et al. The Lipschitz Constant of Self-Attention. ICML 2021

[3] Coddington and Levinson. Theory of Ordinary Differential Equations. 1955

[4] Lipman et al. Flow Matching Guide and Code. arXiv 2024

[5] Heng et al. Out-of-Distribution Detection with a Single Unconditional Diffusion Model. NeurIPS 2024

[6] Rao et al. The Geometry of Flow-Matching Uncertainty. arXiv 2026

[7] Sun et al. Out-of-Distribution Detection with Deep Nearest Neighbors. ICML 2022

[8] Du et al. VOS: Learning What You Don't Know by Virtual Outlier Synthesis. ICLR 2022

Reviewer_KR6oOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Diverging Flows, a neat adaptation of the flow matching loss function, which tackles silent extrapolation -- the phenomenon of neural networks generalizing to OOD data which prevents detection of model uncertainty. The introduction contains compelling motivations and examples of how silent extrapolation has had damaging effects in real life applications. The key trick is to define a manifold of acceptable conditions (data points) for which the velocity field should create straight paths, while generating dramatically more curved or diverging trajectories for conditions outside of the manifold. This contrastive learning adaptation of the FM loss (via regularizing terms) does not require a negative dataset.

Critical strengths

This paper has multiple critical strengths.

  1. The Intro section was a delight to read. It is offers a terrific motivation of why the problem at hand is important without ever relying on jargon terminology, while providing non-generic pointers to reports where issues with extrapolations have or are anticipated to cause damage. As such it strongly separates itself from a many other ML conference papers. This doesn't only provide a positive reading experience, it heavily affects the informativeness of the paper, why I score it as a 3 in informativeness.

  2. The extrapolation problem in FM is neatly formulated in Sec. 4.1. This formalization of the extrapolation is not only useful for understanding and motivating the contribution in this submission, but might also be useful as a basis for other researchers to think about alternative solutions to the extrapolation problem. For Diverging Flows, this formulation allows the construction of a manifold boundary, which in turn means that the method does not require a negative dataset -- this is very nice, and well-motivated in Sec. 4.3: the most difficult samples are on the manifold boundary, and, if I understand correctly, the constructed energy barrier messes with the smoothness bias of NNs, disallowing silent extrapolation / NN generalization outside the manifold. As far as I see, this seems clever and useful. An algorithm for constructing boundary examples is given in the section, too.

Critical weaknesses

Although I find the paper very well-written and understandable in general, I had critical issues understanding cc. See the two comments below.

  1. I suggest, in the top of Sec. 3, introducing and exemplifying the manifold object and its interplay with the conditions. Definition 3.1 is quite difficult to grasp given only the information provided at that point.

  2. I thought that cc was an initial condition for the ODE, i.e. x0x_0, but as I got to Eq. (1) I got confused. In FM as introduced in the pointer to Lipman, the neural vector field takes as input time and a state xx. In Eq. (1) unconditional flow model ϕt\phi_t is used (this is ok!) without introduction (not so nice), and the vector field additionally takes a condition cc. Since the extra condition is not standard, I would expect an explanation or a reference here. And then, adding to my confusion, in line 157, I am not able to understand ϕ1(c)\phi_1(c). The flow model as I have come to understand it is defined as xt=ψt(x)x_t = \psi_t(x), which raises one issue and one confusion: issue this is not a trajectory, as stated in the sentence in line 157, and confusion x1=ψ1(x0)x_1 = \psi_1(x_0), which leads me to believe that cc is in fact x0x_0, but then I am confused by extra input to the vector field? Is c=x0c = x_0? In summary, the FM paragraph needs to contain a more precise mathematical description of the objects introduced in Eq. 1.

If these issues are addressed or sufficiently explained, it would strengthen the manuscript and I would consider raising my score.

Minor:

In Fig. 1 you have additional space to explain the ID subindex abbreviation.

Questions for authors

Is Diffusion Flows applicable to a mixture of vector fields? E.g., in medical applications one might want to train one vector field per patient, á la Rohbeck et al. (2025). If one introduces a new input that is on the condition manifold of one of the vector field networks, but all other nets disagree with the new data point, all except one network would produce curves that would "trigger the OOD alarm". However, again, the data point actually matches the condition manifold of one of the nets. From the mixture perspective, the new data point then comes from a valid condition, but will a mixture of Diffusion Flows be interpretable in any other way than classifying the new point as an anomaly?

References

Rohbeck et al. 2025, Modeling Complex System Dynamics with Flow Matching Across Time and Conditions.

3 comments

We thank the reviewer for the careful reading and the kind words on the introduction and on the formulation in Section 4.1. The reviewer's understanding of Section 4.3 is correct. The boundary samples raise a barrier of high transport energy around the valid conditions, and recovering low energy beyond it would require the vector field to vary even more sharply, which the smoothness bias of neural networks disfavors.

I suggest, in the top of Sec. 3, introducing and exemplifying the manifold object and its interplay with the conditions. (W1)

In the revised PDF, Section 3 opens with this introduction and an example, before Definition 3.1. The new text reads as follows.

Let C=Rk\mathcal{C} = \mathbb{R}^k denote the conditioning space, i.e., all inputs the model accepts, and let training pairs (x,c)(x, c) be drawn from pdata(x,c)p_{data}(x, c). We define the manifold of valid conditions M⊂C\mathcal{M} \subset \mathcal{C} as the support of the condition marginal pdata(c)p_{data}(c), which we assume to be lower-dimensional. For instance, if the conditions are generated by m<km < k underlying factors z∈Z⊂Rmz \in \mathcal{Z} \subset \mathbb{R}^m through a smooth map gg, i.e., c=g(z)c = g(z), then M=g(Z)\mathcal{M} = g(\mathcal{Z}) is an mm-dimensional manifold in Rk\mathbb{R}^k, and a perturbed input c+δc + \delta, with δ\delta normal to M\mathcal{M}, is accepted by the model but lies off M\mathcal{M}. The target conditional pdata(⋅∣c)p_{data}(\cdot \mid c) is defined only for c∈Mc \in \mathcal{M}, so a prediction for c∉Mc \notin \mathcal{M} has no ground truth and is necessarily an extrapolation. [...] The full notation is summarized in Appendix O. Definition 3.1 therefore asks for fidelity where a target exists and for a detectable signal where it does not.

Is c=x0c = x_0? [...] In summary, the FM paragraph needs to contain a more precise mathematical description of the objects introduced in Eq. 1. (W2)

No, cc and x0x_0 are two separate inputs. The source sample x0x_0 is the initial state of the ODE, while cc is an extra input to the vector field that stays fixed during integration, as in standard conditional FM [1, 2] and Algorithms 1 and 2 (App. I). In line 157 we slightly abused the notation, since ψ1(i)(c)\psi_1^{(i)}(c) denotes the endpoint ψ1(x0(i);c)\psi_1(x_0^{(i)}; c) of the sampling trajectory from x0(i)x_0^{(i)}, not the trajectory itself. The revised Flow Matching paragraph of Section 3 defines both objects and reads as follows.

Each prediction is generated from two inputs, a source sample x0∼N(0,I)x_0 \sim \mathcal{N}(0, I) from the noise prior p0p_0, drawn independently of cc, and the condition cc, through the flow ODE ddtψt(x0;c)=vtθ(ψt(x0;c),c)\frac{d}{dt}\psi_t(x_0; c) = v_t^\theta(\psi_t(x_0; c), c), ψ0(x0;c)=x0\psi_0(x_0; c) = x_0 (Eq. 1), where vtθv_t^\theta is a neural vector field that pushes the noise prior p0p_0 to the target density p1(⋅∣c)p_1(\cdot|c). For a fixed condition cc, Eq. 1 defines the flow ψt(⋅;c):Rd→Rd\psi_t(\cdot; c): \mathbb{R}^d \to \mathbb{R}^d, t∈[0,1]t \in [0,1], which is unique when vθv^\theta is locally Lipschitz in xx (Section 4.1) [...]. The sampling trajectory of x0x_0 is the curve of states x^t=ψt(x0;c)\hat x_t = \psi_t(x_0; c), t∈[0,1]t \in [0,1], and its endpoint x^1=ψ1(x0;c)\hat x_1 = \psi_1(x_0; c) is a sample of pθ(⋅∣c)p_\theta(\cdot|c) [...]. Since x0x_0 is independent of cc, the FM loss supervises vθv^\theta on Rd×M\mathbb{R}^d \times \mathcal{M} only.

In Fig. 1 you have additional space to explain the ID subindex abbreviation. (Minor)

In the revised PDF, the caption of Figure 1 states that ID stands for in-distribution and that CID\mathcal C_{\mathcal{ID}} is the manifold of valid conditions M\mathcal{M}.

Will a mixture of [Diverging] Flows be interpretable in any other way than classifying the new point as an anomaly? (Q1)

Yes. The mixture rejects a query only when all components reject it, so the new point is accepted and the matching component makes the prediction. Each component kk shapes the boundary of its own manifold Mk\mathcal M_k, so its DOT score indicates whether c∈Mkc \in \mathcal M_k, and the valid conditions of the mixture are the union M=⋃kMk\mathcal{M} = \bigcup_k \mathcal M_k.

In the reviewer's example, the rejecting components only say that cc lies outside their own manifolds, while the accepting component identifies the patient. The conformal guarantee of Section 4.4 and App. G carries over without correction. If each component is calibrated on its own valid conditions at level α\alpha, then for c∈Mjc \in \mathcal M_j component jj rejects with probability at most α\alpha, and so does the mixture. When manifolds overlap, the conformal p-values put the components on a common scale, and one can select the component with the highest p-value.

We thank the reviewer for the thorough feedback and hope that, with these additions, the reviewer will consider increasing the score.

References

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

[2] Lipman et al. Flow Matching Guide and Code. arXiv 2024

Thank you for an excellent and concise response. I only have one syntactic issue with the new Sec 3 example: "c+δc+\delta with δ\delta normal to M\mathcal{M}". What does δ\delta normal to M\mathcal{M} mean?

Other than that, the example and the more spelled out notations resolve my concerns. The new experimental results in the response to 2Cf2 also strengthen the paper that was potentially lacking in number empirical comparative results.

I'll raise my score as I think this is well-suited for ICLR.

Thank you for the careful reading. By "δ\delta normal to M\mathcal{M}" we meant a direction orthogonal to the manifold at the point cc. Precisely, δ\delta is a small nonzero vector orthogonal to the tangent space TcMT_c\mathcal{M}, so moving from cc along δ\delta leaves M\mathcal{M} instead of sliding along it. The input c+δc + \delta is still a valid vector in Rk\mathbb{R}^k that the model accepts, but no training condition lies there. We have rephrased the sentence this way in the first paragraph of Section 3.1 of the revised PDF.

Thank you again for the valuable feedback and the constructive discussion.

Reviewer_TP9YOpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
3(+2)
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper considers the problem of out of domain (OOD) generalization of flow matching models. A deployed flow matching model can be evaluated on conditions that are out of domain, and the flow matching model might not generalize well to them. Therefore, we need a qay to test whether the point on which the FM model is evaluated on is OOD or not. The paper proposes a new training loss that allows to test this once the model is trained.

Critical strengths

S1. The problem of OOD detection is important, as highlighted in the introduction. S2. The proposed method is principled

Critical weaknesses

W1. The paper is hard to follow. A short background section introducing the notation would be very helpful. W2. The paper does not distinguish various kinds of conditionings. For example, the most widely used conditional FM models are class-conditioned, but where the model goes from pure noise to data-manifold. In that case, I do not think that this method is relevant. The experiments focus on trajectory inference, where the p_0 and p_1 are both data-manifold. In that setting I am confused on what is the distinction between the conditioning variable c and the source point x_0. It would be helpful to distinguish cases where the space (x0, c) is a meaningful submanifold or not.

Questions for authors

See W2

6 comments

We thank the reviewer for finding the problem important and the method principled.

The most widely used conditional FM models are class-conditioned, but where the model goes from pure noise to data-manifold. In that case, I do not think that this method is relevant. The experiments focus on trajectory inference, where the p0p_0 and p1p_1 are both data-manifold. (W2)

In all experiments the source p0p_0 is Gaussian noise, so none of them transports one data distribution to another, and the setting is the noise-to-data conditional FM the reviewer has in mind. Section 3 introduces vtθv_t^\theta as the vector field that "pushes a noise prior p0p_0 to the target density p1(⋅∣c)p_1(\cdot|c)" (line 155 of the submission), and both algorithms, which every experiment follows, start from noise, "prior noise x0∼p0x_0 \sim p_0" and "Initial noise x0∼p0x_0 \sim p_0" (Algorithms 1 and 2, lines 1517 and 1535). The condition is a separate input that the paper specifies for each task, "the current atmospheric state as the conditioning input" on ERA5 (lines 385–386), the MNIST digit "(c)" that is mapped to an SVHN image "(x)" in style transfer (lines 436–438), and a point on the spiral (Section 5.1). This is the standard conditional FM setting [1, 2], with a continuous condition cc in place of a class label.

We agree that class-conditioned FM does not need Diverging Flows, and the paper targets continuous conditions, as in the forecasting and robotics predictors that motivate it (lines 53–55). A class-conditioned model only receives one of finitely many labels, each of them valid, so it cannot be given an off-manifold condition. A continuous condition, such as the atmospheric state in weather forecasting [3] or the observation of a robot policy [4, 5], can take any value in Rk\mathbb R^k, and inputs that no real system produces still reach the model. For these inputs the learned vector fields "construct smooth paths between almost any input and output", so the model "neither degrades nor diverges" but transports the noise to a plausible target (lines 61–65). Conditional FM and diffusion models are now displacing prior predictors in exactly these domains (lines 101–107).

It would be helpful to distinguish cases where the space (x0,c)(x_0, c) is a meaningful submanifold or not. (W2)

Only cc can be invalid, since x0x_0 is the noise sample. Definition 3.1 defines validity through the condition alone, and x0x_0 is drawn from the same noise prior for every cc (line 155), so (x0,c)(x_0, c) is valid exactly when c∈Mc \in \mathcal{M}.

A short background section introducing the notation would be very helpful. (W1)

Section 3 already introduces the conditioning space and the manifold of valid conditions (lines 138–147) and the Flow Matching components (lines 149–157). The revised PDF extends this opening with an example and adds a notation table (Appendix O).

All concerns in W2 come from reading x0x_0 as a data point, while the submission defines it as noise throughout (line 155, Algorithms 1 and 2). Diverging Flows therefore works in the standard noise-to-data conditional FM that the reviewer describes, with the continuous conditions of forecasting and robotics. To state this clearly, the revised PDF adds a sentence to the introduction (Section 1, fourth paragraph). Since the reviewer also found the method principled and the soundness concerns minor, we would be grateful if they would reconsider the rating.

References

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

[2] Lipman et al. Flow Matching Guide and Code. arXiv 2024

[3] Price et al. Probabilistic Weather Forecasting with Machine Learning. Nature 2025

[4] Black et al. π0: A Vision-Language-Action Flow Model for General Robot Control. arXiv 2024

[5] Rouxel et al. Flow Matching Imitation Learning for Multi-Support Manipulation. Humanoids 2024

I thank the authors for their rebuttal. As the reviewer with the lowest score, I reread the paper thoroughly. The rebuttal clarified several points, and I find the problem setting very interesting and the experimental results strong and convincing. In have therefore updated my presentation and informativeness scores. However, several aspects remain unclear, and I do not think the paper in its current form meets ICLR’s standards. I encourage the authors to improve its clarity, as I think this could become a very good paper.

(x1,c)(x_1,c) constraints vs cc constraints

I remain unclear about the distinction between constraints on cc and constraints on the pair (x1,c)(x_1,c). My understanding is that x1x_1 and cc are coupled, while x0x_0 is independent of both. An x1x_1 can therefore be compatible with cc but incompatible with c′c'. Consider an inverse problem with c=Ax1+bc=Ax_1+b: several values of x1x_1 may correspond to the same cc, while, under suitable support and rank assumptions, valid conditions can fill Rk\mathbb{R}^{k}. Formally, a joint constraint set M~⊂Rd×Rk\widetilde{\mathcal{M}}\subset\mathbb{R}^{d}\times\mathbb{R}^{k} can have a full-dimensional projection πc(M~)\pi_c(\widetilde{\mathcal{M}}). I think it is important to distinguish this from the paper’s assumption that valid conditions lie on a lower-dimensional manifold M=CID\mathcal{M}=\mathcal{C}_{\mathrm{ID}}.

This distinction also matters for Eq. 2. I understand from the revision that x1x_1 is sampled under a valid condition cc, and that E(v;c~)\mathcal{E}(v;\tilde{c}) uses the same reference path while changing the condition supplied to the velocity field. However, the quantity then depends on the original cc, not only on c~\tilde{c}. This dependence is absent from the notation, and I still find its interpretation as a quantity associated with c~\tilde{c} alone unclear.

Relatedly, in Eqs. 3–4, perturbing cc while holding x1x_1 fixed can make the pair incompatible without making the new condition invalid. In an inverse problem with CID=Rk\mathcal{C}_{\mathrm{ID}}=\mathbb{R}^{k}, the same sampling procedure could still label perturbed conditions as negatives. I appreciate that this might fall outside the paper’s manifold assumption, but clarifying this limitation would help. I would also appreciate a clearer explanation for why training on boundary samples suffices to detect conditions far from the boundary.

Figure and presentation

I find Figure 2 difficult to interpret. In Figure 2, the difference between the probabilistic regression and conditional generation tasks is not apparent from the panels. It would help to make clearer which elements represent cc, x0x_0, and the generated x1x_1, which manifold belongs to the conditioning space, and how off-manifold conditioning affects the depicted trajectories.

The PDF now clarifies the cc and x0x_0 dinstinction. But I still recommend separating a background from Sec 3 to distinguish established formalism from the contributions.

Theory

Does DiFlo change the learned probability path for in-distribution data, and under what conditions do the theoretical guarantees of FM remain valid? The appendix discusses possible changes to valid trajectories, but I do not see a guarantee that the additional losses preserve the optimal in-distribution conditional distribution. A statement of this kind would strengthen the method.

Experiments

My concern with the AUROC and FPR comparisons is the interpretation of the underlying score. DOT is designed to expose the behavior encouraged by DiFlo; being well-defined for FM does not by itself make it a meaningful OOD score for standard FM. Comparing FM and DiFlo using DOT is useful as an ablation, but poor performance of FM with this score does not establish that FM generally fails at OOD detection. The additional baselines are useful, but the conclusions should remain limited to the detectors and settings evaluated. To me, the central empirical result is that DiFlo maintains predictive performance, for example in MSE, while providing useful OOD detection, which the experiments demonstrate. I think this is enough of a contribution for a good paper.

It would also be interesting to compare the predictive generalization of FM and DiFlo on valid conditions from held-out regimes (OOD evaluation). I would expect FM to perform better, although this remains a hypothesis. This would not necessarily undermine the method, but it could reveal a limitation: ideally, we would identify OOD inputs without reducing predictive performance on them.

Taking a step back, is the intended conclusion that DiFlo can be used in safety-sensitive applications while FM cannot, or that predictions on OOD inputs require an additional reliability assessment? I would also appreciate a clearer justification for using one model for both generation and OOD detection, rather than a separate detector that can reject an OOD condition before querying the generative model.

We thank the reviewer for rereading the paper. We are glad that the rebuttal settled the distinction between cc and x0x_0, on which W2 was based. Several of the new points raised by the reviewer are already addressed in the paper, and we give the location of each. The revised PDF also includes the clarifications below (Section 3.2, Figure 2, Proposition D.3), and line numbers refer to it.

Formally, a joint constraint set M~⊂Rd×Rk\widetilde{\mathcal{M}} \subset \mathbb{R}^d \times \mathbb{R}^k can have a full-dimensional projection πc(M~)\pi_c(\widetilde{\mathcal{M}}). [...] In an inverse problem with CID=Rk\mathcal C_{ID} = \mathbb{R}^k, the same sampling procedure could still label perturbed conditions as negatives.

At inference the model receives only x0x_0 and cc, so the pair (x1,c)(x_1, c) is never an input and only cc can be invalid. The compatibility of x1x_1 with cc is pdata(x1∣c)p_{data}(x_1 \mid c), which is what FM learns and what fidelity requires (Definition 3.1). If the valid conditions fill Rk\mathbb{R}^k, as in the inverse problem, there is no invalid condition to detect. Detection requires that the valid inputs do not fill the input space, and for high-dimensional conditions the standard form of this assumption is the manifold hypothesis [1, 2], stated in Section 4.1 (lines 189–190). Under it, a perturbed condition generically leaves M\mathcal{M}, so the boundary samples of Eqs. 3–4 are invalid conditions.

However, the quantity then depends on the original cc, not only on c~\tilde{c}.

The dependence on cc is part of the definition (lines 197–199), and the quantity that flags a query depends on c~\tilde{c} alone. E\mathcal{E} serves training, where the margins raise it at boundary samples, and it is never used to score a query. At inference, Theorem 4.2 and the DOT score (Eq. 7) use the realized trajectory and its own endpoint x^1\hat x_1, which the model produces from x0x_0 and c~\tilde{c}.

I would also appreciate a clearer explanation for why training on boundary samples suffices to detect conditions far from the boundary.

Section 4.3 explains this (lines 300–307), and Appendix F tests it. Under FM, low energy off M\mathcal{M} is inherited from the smooth extension of the on-manifold field (Proposition 4.1), since the FM loss never acts there. The margins interrupt this extension at the boundary, and recovering low energy beyond it would require an even sharper field, which the spectral bias of neural networks disfavors [3]. Trained only with boundary samples, Diverging Flows reaches an AUROC of 1.000 on far off-manifold conditions, against 0.65 to 0.73 for the FM likelihood.

But I still recommend separating a background from Sec 3 to distinguish established formalism from the contributions.

Section 3 already separated them, and the revised PDF makes the split explicit. Section 3.1 states the problem, and the new Section 3.2, Flow Matching Preliminaries (lines 162–183), recalls the established formalism [4]. Our contributions are Definition 3.1 and Section 4.

I do not see a guarantee that the additional losses preserve the optimal in-distribution conditional distribution.

The FM optimum on M\mathcal{M} is preserved whenever the margins can be met. This follows from a few-line derivation based on the squared loss of FM, and we added it as Proposition D.3 (Appendix D.4). The FM loss constrains the field only on valid conditions, while the margins can be met by its values at boundary samples off M\mathcal{M}. A field can therefore equal the FM optimum v⋆v^\star on M\mathcal{M} and take any value off it, and the minimizer of the full objective can lose on M\mathcal{M} at most the margin penalty that such a field still pays. Formally, for the minimizer v^\hat v of Eq. 5 and any field v∘v^\circ in the model class that equals v⋆v^\star on M\mathcal{M},

E∥v^−v⋆∥2≤λRrepel(v∘)+βRcurve(v∘),\mathbb{E} \lVert \hat v - v^\star \rVert^2 \le \lambda R_{repel}(v^\circ) + \beta R_{curve}(v^\circ),

with RrepelR_{repel} and RcurveR_{curve} the expected margin losses and the expectation over valid data. For bounded data the repulsion margin can always be met. If the curvature margin can be met as well, v^=v⋆\hat v = v^\star on M\mathcal{M} and the FM guarantee pθ(⋅∣c)=pdata(⋅∣c)p_\theta(\cdot \mid c) = p_{data}(\cdot \mid c) [4] holds unchanged, and otherwise the deviation on M\mathcal{M} is bounded by the remaining curvature penalty, weighted by β\beta. This is consistent with our experiments, where the fidelity of Diverging Flows matches FM in every task (Tables 1–3).

References

[1] Bengio et al. Representation Learning: A Review and New Perspectives. IEEE TPAMI 2013

[2] Pope et al. The Intrinsic Dimension of Images and Its Impact on Learning. ICLR 2021

[3] Rahaman et al. On the Spectral Bias of Neural Networks. ICML 2019

[4] Lipman et al. Flow Matching for Generative Modeling. ICLR 2023

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their feedback. We have uploaded a revised PDF, with new or modified content shown in purple. In response to reviewer 2Cf2, we added new baselines in Appendix M. In response to reviewers KR6o and TP9Y, we expanded the background on the problem setting and on Flow Matching in Section 3 and added a notation table in Appendix O. Minor changes are described in the individual replies. So that the references in the reviews remain valid, Appendices A to L keep their submitted letters, while the experiment details and the qualitative results move to Appendices N and P.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.