acceptodds
Under review as a conference paper at ICLR 2027

Active Learning for Conditional Flow Matching Model in Engineering Design

Abstract

Although the flow matching model has demonstrated powerful capabilities in modern machine learning, its training notoriously relies on an incredibly large scale of high-quality labeled samples. Nevertheless, the acquisition of high-quality labeled datasets is hindered by exorbitant labeling costs in certain fields, notably engineering design. Therefore, selecting the most informative samples for training at minimal cost poses a key challenge in this field. This issue constitutes a central topic in active learning, a subfield of machine learning dedicated to maximizing model performance while minimizing annotation cost. The central challenge involves developing a query strategy to acquire the most informative data samples with minimal labeling effort. This paper presents a pilot study that investigates the application of active learning, which traditionally explored within the context of discriminative models, to flow matching models. By analyzing conditional flow matching through a piecewise-linear closed-form formulation, this work provides insight into how individual data points can affect the diversity and conditional accuracy of the resulting generative model. Building on this analysis, we propose two distinct query strategies: one aimed at enhancing model diversity, and the other designed to improve model accuracy. These strategies highlight a dataset-level trade-off between diversity and accuracy, and motivate a hybrid query strategy that balances the two objectives. Extensive experiments validate the effectiveness of our approach, showing that the proposed query strategies outperform baselines in terms of diversity and accuracy, respectively.

Primary area:
generative models
Keywords:
Flow matching model, Active learning, Generative model, Engineering Design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aWXGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies active learning for conditional flow matching in engineering design, where obtaining condition labels can be substantially more expensive than collecting candidate designs. The authors develop an analytical perspective based on a closed-form flow-matching formulation and a continuous piecewise-linear (CPWL) assumption. From this analysis, they argue that samples with similar conditions mainly improve generative diversity, whereas samples covering distant regions of the condition space improve conditional accuracy. Based on this observation, they propose two query strategies, QDQ_D and QAQ_A, respectively targeting diversity and accuracy, together with a weighted hybrid strategy. The approach is evaluated on one synthetic dataset and three engineering-design datasets using multiple active-learning baselines, 20 random seeds, supplementary statistical tests, proxy-error analyses, and experiments that relax the theoretical assumptions.

Critical strengths
  1. Rather than using generative models to assist conventional active learning, the work explicitly studies active learning for conditional generative models.
  2. The evaluation covers four datasets, multiple classical and recent active-learning baselines, repeated experiments over 20 random seeds, statistical comparisons, computational-cost analysis, proxy-error analysis, and tests under violations of the CPWL and closed-form assumptions.
Critical weaknesses
  1. The theoretical connection between the CPWL assumption and the proposed condition-space interpolation appears insufficiently justified.
  2. The practical query mechanism depends heavily on proxy-predicted labels for the unlabeled pool, which weakens the claimed active-learning interpretation.
Questions for authors
  1. Clarification: Can the authors provide a formal justification for Eq. (2), specifically explaining under what conditions c∗c^\ast and the vertices c0,…,cdc_0,\ldots,c_d are guaranteed to lie in the same affine region of the conditional network for fixed x,tx,t? How is the simplex subdivision of the condition space related to the actual activation-region partition of the trained CPWL network? If this cannot be guaranteed, which theoretical conclusions used to motivate QDQ_D and QAQ_A remain valid?
  2. Clarification: How sensitive are the final active-learning gains to the quality of the label proxy itself? In particular, could the authors compare QD/QAQ_D/Q_A using several proxy models with controlled prediction accuracy, and report performance as a function of proxy error? This would help determine whether the gains originate from the proposed acquisition principle or from the quality of the surrogate label predictor.
  3. Clarification: For the hybrid strategy QHybrid=ωQD+(1−ω)QAQ_{\mathrm{Hybrid}}=\omega Q_D+(1-\omega)Q_A, are the two scores normalized before linear combination? Since their individual terms involve distances and entropy changes with potentially different numerical scales, how stable is the meaning of ω\omega across datasets and acquisition rounds?

The problem is interesting and the experimental effort is substantial, but the main theoretical argument currently relies on assumptions whose connection to the actual conditional network is not sufficiently established, while the practical method depends strongly on surrogate labels that are unavailable by definition in active learning. These issues make me lean against acceptance, although the paper is close enough that stronger theoretical clarification and a more convincing proxy-dependence analysis could materially change my assessment.

6 comments

We thank the reviewer for the constructive comments. We have clarified the theoretical assumptions, expanded the quantitative analysis of the trained model and proxy dependence, and documented the normalization and interpretation of the hybrid strategy.

1. Conditions underlying Eq. (2) and the scope of the theoretical conclusions

Section 2.2 now explicitly states the condition-wise interpolation assumption and its sufficient condition. Appendix C discusses the condensation literature as background for data-dependent interpolation and provides quantitative comparisons between the analytical construction and the trained model.

For fixed (x′,t)(\boldsymbol{x}',t), a sufficient condition for Eq. (2) is that the entire condition simplex

C=conv⁡c0,…,cd\mathcal{C}=\operatorname{conv}\\{\boldsymbol{c}_0,\ldots,\boldsymbol{c}_d\\}

is contained in a single affine region of the map c↦ut(x′,c)\boldsymbol{c}\mapsto\boldsymbol{u}_t(\boldsymbol{x}',\boldsymbol{c}). Then, for

c∗=∑i=0daici,ai≥0,∑i=0dai=1,\boldsymbol{c}^{\ast}=\sum_{i=0}^{d}a_i\boldsymbol{c}_i,\qquad a_i\geq0,\qquad \sum_{i=0}^{d}a_i=1,

we have

ut(x′,c∗)=∑i=0daiut(x′,ci).\boldsymbol{u}_t(\boldsymbol{x}',\boldsymbol{c}^{\ast})=\sum_{i=0}^{d}a_i\boldsymbol{u}_t(\boldsymbol{x}',\boldsymbol{c}_i).

For the trajectory-level analysis, this condition is assumed at the relevant state–time pairs along the sampling trajectory.

CPWL describes local affine behavior. Alignment between label-defined simplices and condition-affine regions is an additional analytical assumption. Our analytical model combines closed-form vertex predictors with a piecewise-affine extension in the condition variable:

u‾t(x′,c∗)=∑i=0daiutCF(x′,ci).\overline{\boldsymbol{u}}_t(\boldsymbol{x}',\boldsymbol{c}^{\ast})=\sum_{i=0}^{d}a_i\boldsymbol{u}_t^{\mathrm{CF}}(\boldsymbol{x}',\boldsymbol{c}_i).

Here, CF denotes “closed form”: utCF\boldsymbol{u}_t^{\mathrm{CF}} is the closed-form noise predictor in Eq. (1). We retain ut\boldsymbol{u}_t for the trained network output and use the overbar to distinguish the ideal interpolated predictor.

The interpolation identity holds by construction for this analytical model. A trained network realizes this construction when it is affine in the condition variable within the simplex and matches the closed-form vertex outputs.

Appendix B now states the assumptions and arguments for the subsequent conclusions. Let Si\mathcal{S}_i denote the finite nonempty set of training data associated with condition ci\boldsymbol{c}_i. With positive normalized empirical sample weights, fixed barycentric coefficients, and standard Gaussian initialization, the ideal weighted flow converges almost surely to an endpoint in

\mathcal{S}(\boldsymbol{c}^{\ast})=\left\\{\sum_{i=0}^{d}a_i\boldsymbol{s}_i\\;\middle|\\;\boldsymbol{s}_i\in\mathcal{S}_i\right\\}.

This is the candidate set described in Eq. (3). The revised proof directly analyzes the weighted flow, including its convergence and candidate-endpoint characterization.

For Lemma B.2, we clarify the original inverse-function assumption by considering a C2C^2 local right-inverse branch gg, satisfying f(g(c))=cf(g(\boldsymbol{c}))=\boldsymbol{c}, with the participating vertex samples belonging to this same branch. Under the stated Lipschitz and second-derivative bounds, the revised conditional-error bound is

\left\\|f(\boldsymbol{x}^{\ast})-\boldsymbol{c}^{\ast}\right\\|_2\leq L_f\eta+K\max_{0\leq i,j\leq d}\left\\|\boldsymbol{c}_i-\boldsymbol{c}_j\right\\|_2^2,

where

\eta=\left\\|\boldsymbol{x}^{\ast}-\sum_{i=0}^{d}a_i g(\boldsymbol{c}_i)\right\\|_2.

We retain the original constants LfL_f, MM, and KK: LfL_f is the Lipschitz constant of ff, MM bounds the vector-valued second derivative of the local branch gg, and K=Lf(d+1)M/2K=L_f(d+1)M/2. The term LfηL_f\eta accounts for deviation from the same-branch interpolant; when η=0\eta=0, the bound reduces to the original quadratic form.

The candidate-set analysis provides the diversity upper-bound argument motivating QDQ_D, while the local interpolation-error bound motivates QAQ_A’s condition-coverage objective. These analytical conclusions apply under the stated interpolation and regularity conditions. The acquisition scores are computed from the data and proxy-predicted labels, and their empirical performance is assessed through the solver-based evaluations.

To examine the approximation empirically, we have expanded Appendix C with quantitative comparisons in both single-simplex and multiple-simplex settings. The evaluation uses 1,000 Gaussian noise draws with corresponding target labels. Predictor outputs are compared at identical (x′,t,c∗)(\boldsymbol{x}',t,\boldsymbol{c}^{\ast}) using the same noise parameterization. For generated endpoints, we compute each sample’s minimum MSE to its theoretical candidate set and then average over generated samples.

In the original two-condition Airfoil setting, labels 0 and 1 define a single one-dimensional simplex, with multiple training data points at each endpoint condition. The average nearest-candidate MSE is on the order of 10−410^{-4}. Both predictor-output MSE and nearest-candidate MSE are largest near the target label farthest from the observed training labels, near the midpoint between the two endpoint conditions.

The multiple-simplex evaluation uses a conditional flow matching model trained on real data with their observed condition labels. We compare this trained model with two analytical references. The first uses the observed vertex conditions. The second groups multiple samples with nearby labels around each vertex, approximating nearby conditions by a shared vertex label. The second construction gives a smaller overall discrepancy between the analytical and trained predictor outputs, providing a closer empirical approximation to the trained model.

We use the paper’s time convention throughout: t=0t=0 corresponds to Gaussian noise and t=1t=1 to data. In the multiple-simplex comparisons, predictor MSE is approximately 10−310^{-3}–10−210^{-2} at smaller flow times and increases overall as tt approaches one. This trend agrees with the unconditional closed-form comparison in Appendix H.2 under the same time convention. Along the label axis, both error measures exhibit multiple local peaks.

In the label-wise comparisons, average nearest-candidate MSE is numerically of a lower order of magnitude than predictor-output MSE. These measurements quantify agreement at the predictor and generated-endpoint levels, respectively.

2. Dependence on proxy quality and its relationship to acquisition performance

For unqueried candidates, our acquisition rules use performance labels predicted by the current proxy, and the selected samples receive true labels through querying. Predictive models and model-based estimates also guide querying in other active learning methods, including Committee, UEADA, ISP, and EMGM. Even when labels are not predicted explicitly, query decisions may depend on learned feature representations, as in CoreSet and ALIM. Our method uses predicted performance labels to evaluate its condition-space acquisition criteria.

We have expanded the proxy study in Appendix G to include RBF, Random Forest, SVR, and XGBoost, reporting proxy prediction MSE alongside downstream conditional error and diversity at acquisition rounds 1, 3, and 5.

“Random” denotes randomly sampled proxy hyperparameter configurations. “Large proxy error” denotes the ten configurations with the highest proxy MSE among 300 randomly sampled configurations. These high-error configurations provide an offline sensitivity diagnostic.

The results reveal a stage-dependent relationship. For each proxy family, prediction errors are relatively large in round 1, with dispersed downstream performance and a variable association between proxy MSE and the downstream metrics.

In rounds 3 and 5, the relationship within each proxy family becomes clearer: larger proxy MSE is generally associated with worse downstream performance—higher true conditional MSE for QAQ_A and lower generated-sample diversity for QDQ_D.

We report the results separately by proxy family and acquisition round. Within-family comparisons characterize sensitivity to prediction error, while comparisons across families examine variation associated with the proxy model. Prediction accuracy and local error structure can influence candidate rankings. The downstream results therefore reflect the acquisition rule together with the quality of the proxy information used to implement it.

3. Normalization of hybrid scores and interpretation of the hybrid weight

At each acquisition round, the complete scalar scores sD(x)s_D(\boldsymbol{x}) and sA(x)s_A(\boldsymbol{x}) for QDQ_D and QAQ_A are separately min–max normalized over the current input pool X\boldsymbol{X} before combination. Section 2.4 documents this implementation detail alongside Eq. (7) and distinguishes acquisition scores from selected queries.

Larger ω\omega places greater relative weight on the diversity-oriented score, while smaller ω\omega emphasizes the accuracy-oriented score. The weight-sweep results in Section 3.2 and Appendix D.4 show dataset-dependent changes in the resulting metrics.

Accordingly, ω\omega specifies the direction of the acquisition preference rather than quantitatively controlling the achieved diversity–accuracy trade-off. Its effect depends on the score distributions in the current candidate pool and the dataset. Section 4 identifies hybrid strategies providing quantitatively calibrated control of this trade-off as a direction for future work.

Reviewer_tcQmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies active learning for conditional flow matching models in engineering design, where obtaining labeled data can be expensive. The authors propose a theoretical framework to analyze how training samples affect the diversity and accuracy of generated designs. Based on this analysis, they introduce two query strategies, one targeting diversity and the other accuracy, along with a hybrid strategy that balances both. Experiments on one synthetic and three engineering design datasets show that the proposed strategies outperform several existing active learning methods on their respective objectives.

Critical strengths

Practical and interesting problem: The paper addresses the problem of expensive data labeling in engineering design. The idea of developing separate active learning strategies for diversity and accuracy is interesting and useful.

Strong experimental evaluation: The authors evaluate their methods on four datasets and compare them with several existing active learning methods. The results are promising, and the additional ablation studies and statistical tests strengthen the experimental findings.

Critical weaknesses

Theoretical justification: The theoretical analysis relies on strong assumptions, and some steps are not clearly justified. In particular, the interpolation arguments in Lemmas B.1 and B.2 need more explanation. Since these results motivate the proposed query strategies, this raises concerns about the theoretical foundation of the method.

Limited diversity evaluation: The paper mainly measures diversity using pairwise distances between generated samples. However, larger distances do not necessarily mean more useful or meaningful diversity. It would be helpful to evaluate diversity specifically among generated designs that satisfy the target conditions.

Questions for authors

Theoretical justification: Could the authors provide more details on the proofs of Lemmas B.1 and B.2? In particular, how is the interpolation behavior justified when the flow fields differ, and how does the analysis handle cases where multiple designs share the same performance label?

Diversity evaluation: Since larger pairwise distances do not necessarily indicate better designs, could the authors evaluate diversity only among generated samples that satisfy the target conditions within a certain error threshold? This would help demonstrate whether the increased diversity is meaningful in practice.

5 comments

We thank the reviewer for the constructive comments. We have expanded Appendix B to explain interpolation with distinct vertex fields and multiple training samples sharing a label. Appendix C provides quantitative comparisons with the trained model, and Appendix D.7 reports diversity after filtering by true condition error.

1. Proofs of Lemmas B.1 and B.2 and multiple samples sharing a performance label

Section 2.2 explicitly states the condition-wise affine interpolation assumption used in the analytical model and its sufficient condition. For a fixed target condition,

c∗=∑i=0daici,ai≥0,∑i=0dai=1.\boldsymbol{c}^{\ast}=\sum_{i=0}^{d}a_i\boldsymbol{c}_i,\qquad a_i\geq0,\qquad \sum_{i=0}^{d}a_i=1.

Throughout, t=0t=0 is the Gaussian-noise endpoint and t=1t=1 is the data endpoint. We use pp for the sample-space dimension and dd for the condition-space dimension.

Let Si⊂Rp\mathcal{S}_i\subset\mathbb{R}^{p} be the finite nonempty set of training data at condition ci\boldsymbol{c}_i, with empirical weights πi,s>0\pi_{i,\boldsymbol{s}}>0 and ∑s∈Siπi,s=1\sum_{\boldsymbol{s}\in\mathcal{S}_i}\pi_{i,\boldsymbol{s}}=1. The closed-form posterior retains every sample:

w_{i,\boldsymbol{s}}(t,\boldsymbol{x})\propto\pi_{i,\boldsymbol{s}}\exp\left[-\frac{\left\\|\boldsymbol{x}-t\boldsymbol{s}\right\\|_2^2}{2(1-t)^2}\right],\qquad \sum_{\boldsymbol{s}\in\mathcal{S}_i}w_{i,\boldsymbol{s}}(t,\boldsymbol{x})=1.

Define the posterior mean

mi(t,x)=∑s∈Siwi,s(t,x)s.\boldsymbol{m}_i(t,\boldsymbol{x})=\sum_{\boldsymbol{s}\in\mathcal{S}_i}w_{i,\boldsymbol{s}}(t,\boldsymbol{x})\boldsymbol{s}.

The noise predictor in Eq. (1) can then be written as

utCF(x,ci)=x−tmi(t,x)1−t.\boldsymbol{u}_t^{\mathrm{CF}}(\boldsymbol{x},\boldsymbol{c}_i)=\frac{\boldsymbol{x}-t\boldsymbol{m}_i(t,\boldsymbol{x})}{1-t}.

Here, CF denotes “closed form.” Multiple samples sharing a label remain separate posterior components. We use wiw_i and πi\pi_i to denote the corresponding posterior and prior probability distributions over Si\mathcal{S}_i.

For Lemma B.1, the ideal interpolated predictor and its sampling velocity are

u‾t(x,c∗)=∑i=0daiutCF(x,ci),\overline{\boldsymbol{u}}_t(\boldsymbol{x},\boldsymbol{c}^{\ast})=\sum_{i=0}^{d}a_i\boldsymbol{u}_t^{\mathrm{CF}}(\boldsymbol{x},\boldsymbol{c}_i),

v‾t(x,c∗)=x−u‾t(x,c∗)t=∑i=0daimi(t,x)−x1−t,0<t<1.\overline{\boldsymbol{v}}_t(\boldsymbol{x},\boldsymbol{c}^{\ast})=\frac{\boldsymbol{x}-\overline{\boldsymbol{u}}_t(\boldsymbol{x},\boldsymbol{c}^{\ast})}{t}=\frac{\sum_{i=0}^{d}a_i\boldsymbol{m}_i(t,\boldsymbol{x})-\boldsymbol{x}}{1-t},\qquad 0<t<1.

We analyze this velocity directly, allowing the vertex fields to differ. With finite sample sets, positive normalized weights, fixed barycentric coefficients, and standard Gaussian initialization, the ideal trajectory converges almost surely, with its endpoint in

\mathcal{S}(\boldsymbol{c}^{\ast})=\left\\{\sum_{i=0}^{d}a_i\boldsymbol{s}_i\\;\middle|\\;\boldsymbol{s}_i\in\mathcal{S}_i\right\\}.

Thus, all same-label samples enter the vertex predictor, and the endpoint candidates combine samples from the respective vertex sets.

Appendix B provides the following argument. First choose t0∈(0,1)t_0\in(0,1). On [0,t0][0,t_0],

Dxmi(t,x)=t(1−t)2Cov⁡wi(s).D_{\boldsymbol{x}}\boldsymbol{m}_i(t,\boldsymbol{x})=\frac{t}{(1-t)^2}\operatorname{Cov}_{w_i}(\boldsymbol{s}).

Finite sample sets give a uniform covariance bound. The sampling velocity extends smoothly to t=0t=0, is globally Lipschitz in the state on this interval, and has at most linear growth. Its finite-time flow is a global C1C^1 diffeomorphism. Gaussian initialization therefore yields an absolutely continuous distribution at t0t_0.

We then introduce

z=xtt,τ=log⁡t1−t,ε=e−2τ,τ0=log⁡t01−t0.\boldsymbol{z}=\frac{\boldsymbol{x}_t}{t},\qquad \tau=\log\frac{t}{1-t},\qquad \varepsilon=e^{-2\tau},\qquad \tau_0=\log\frac{t_0}{1-t_0}.

The distribution of z(τ0)=xt0/t0\boldsymbol{z}(\tau_0)=\boldsymbol{x}_{t_0}/t_0 is also absolutely continuous. In these transformed coordinates, define

w_{i,\boldsymbol{s}}(\boldsymbol{z},\varepsilon)\propto\pi_{i,\boldsymbol{s}}\exp\left[-\frac{\left\\|\boldsymbol{z}-\boldsymbol{s}\right\\|_2^2}{2\varepsilon}\right],\qquad \boldsymbol{m}_i(\boldsymbol{z},\varepsilon)=\sum_{\boldsymbol{s}\in\mathcal{S}_i}w_{i,\boldsymbol{s}}(\boldsymbol{z},\varepsilon)\boldsymbol{s}.

The transformed trajectory satisfies

z′=Mε(z)−z,Mε(z)=∑i=0daimi(z,ε),\boldsymbol{z}'=\boldsymbol{M}_{\varepsilon}(\boldsymbol{z})-\boldsymbol{z},\qquad \boldsymbol{M}_{\varepsilon}(\boldsymbol{z})=\sum_{i=0}^{d}a_i\boldsymbol{m}_i(\boldsymbol{z},\varepsilon),

where a prime denotes differentiation with respect to τ\tau.

Define

\Psi_{\varepsilon}(\boldsymbol{z})=-\varepsilon\sum_{i=0}^{d}a_i\log\left[\sum_{\boldsymbol{s}\in\mathcal{S}_i}\pi_{i,\boldsymbol{s}}\exp\left(-\frac{\left\\|\boldsymbol{z}-\boldsymbol{s}\right\\|_2^2}{2\varepsilon}\right)\right].

Normalized priors give Ψε≥0\Psi_{\varepsilon}\geq0, and

\frac{d}{d\tau}\Psi_{\varepsilon(\tau)}(\boldsymbol{z}(\tau))=-\left\\|\boldsymbol{z}'\right\\|_2^2-2\varepsilon\sum_{i=0}^{d}a_i\operatorname{KL}(w_i\\|\pi_i)\leq-\left\\|\boldsymbol{z}'\right\\|_2^2.

The posterior means lie in fixed compact convex hulls. Variation of constants therefore bounds the transformed trajectories and velocities uniformly over bounded initial sets, while the energy inequality gives finite kinetic energy.

Let A=i:ai>0A=\\{i:a_i>0\\}, and let Ni(z)N_i(\boldsymbol{z}) denote the nearest-sample set in Si\mathcal{S}_i. Every accumulation point belongs to

\mathcal{E}=\left\\{\boldsymbol{z}:\boldsymbol{z}\in\sum_{i\in A}a_i\operatorname{conv}N_i(\boldsymbol{z})\right\\}.

Near a point outside E\mathcal{E}, posterior mass outside its nearest-sample sets vanishes uniformly as ε→0\varepsilon\to0, giving a positive late-time speed lower bound. Bounded speed provides a uniform residence time for visits to a smaller neighborhood, so repeated visits contradict finite kinetic energy.

The set E\mathcal{E} is finite. For each nearest-sample pattern, let WW span the within-group sample differences at active vertices. The equal-distance constraints have direction space W⊥W^{\perp}, while the affine hull of the weighted nearest-sample convex hull has direction space WW. Their intersection contains at most one point, and there are finitely many patterns. The accumulation set of a bounded continuous trajectory is connected; its inclusion in the finite set E\mathcal{E} therefore establishes convergence.

To locate the endpoint, write B=S(c∗)B=\mathcal{S}(\boldsymbol{c}^{\ast}) and define

V(τ)=∑i=0daitr⁡Cov⁡wi(s).V(\tau)=\sum_{i=0}^{d}a_i\operatorname{tr}\operatorname{Cov}_{w_i}(\boldsymbol{s}).

Let ΦT(z0)=z(T;τ0,z0)\Phi_T(\boldsymbol{z}_0)=\boldsymbol{z}(T;\tau_0,\boldsymbol{z}_0) be the transformed flow map. Its Jacobian determinant satisfies

JT(z0):=det⁡DΦT(z0)=exp⁡[∫τ0T(V(τ)ε(τ)−p)dτ].J_T(\boldsymbol{z}_0):=\det D\Phi_T(\boldsymbol{z}_0)=\exp\left[\int_{\tau_0}^{T}\left(\frac{V(\tau)}{\varepsilon(\tau)}-p\right)d\tau\right].

Independent posterior draws from the vertex sets give

\operatorname{dist}(\boldsymbol{z},B)^2\leq\left\\|\boldsymbol{z}'\right\\|_2^2+\sum_{i=0}^{d}a_i^2\operatorname{tr}\operatorname{Cov}_{w_i}(\boldsymbol{s})\leq\left\\|\boldsymbol{z}'\right\\|_2^2+V(\tau).

If the endpoint lies outside BB, this inequality and finite kinetic energy give a positive lower bound on VV outside a set of finite time measure. Since ε=e−2τ\varepsilon=e^{-2\tau}, the integral of V/εV/\varepsilon grows faster than the linear term in the Jacobian exponent, yielding JT(z0)→∞J_T(\boldsymbol{z}_0)\to\infty.

For a bounded initial set K\mathcal{K}, let AKA_{\mathcal{K}} contain the initial states whose endpoints lie outside BB. All flow images of K\mathcal{K} remain uniformly bounded. The change-of-variables formula gives

∫AKJT(z0),dz0=Leb⁡(ΦT(AK))≤CK,\int_{A_{\mathcal{K}}}J_T(\boldsymbol{z}_0)\\,d\boldsymbol{z}_0=\operatorname{Leb}(\Phi_T(A_{\mathcal{K}}))\leq C_{\mathcal{K}},

with CKC_{\mathcal{K}} independent of TT. The endpoint map is measurable as a pointwise limit of continuous finite-time flow maps. Fatou’s lemma implies that AKA_{\mathcal{K}} has zero Lebesgue measure; a countable exhaustion extends this conclusion to the whole space. Absolute continuity of z(τ0)\boldsymbol{z}(\tau_0) then gives the almost-sure result. Finally, xt=tz(τ)\boldsymbol{x}_t=t\boldsymbol{z}(\tau) has the same limiting point.

We separately state a sufficient condition for posterior concentration: every b∈S(c∗)\boldsymbol{b}\in\mathcal{S}(\boldsymbol{c}^{\ast}) has a unique nearest sample in each active vertex set. Endpoint convergence then gives a fixed positive squared-distance gap δi\delta_i for tt sufficiently close to one, yielding

wi,s(t,xt)wi,si(t,xt)≤πi,sπi,siexp⁡[−δi2(1−t)2]⟶0,s≠si.\frac{w_{i,\boldsymbol{s}}(t,\boldsymbol{x}_t)}{w_{i,\boldsymbol{s}_i}(t,\boldsymbol{x}_t)}\leq\frac{\pi_{i,\boldsymbol{s}}}{\pi_{i,\boldsymbol{s}_i}}\exp\left[-\frac{\delta_i}{2(1-t)^2}\right]\longrightarrow0,\qquad \boldsymbol{s}\neq\boldsymbol{s}_i.

Under this additional condition, each active posterior concentrates almost surely on one sample. Appendix B states this concentration corollary separately from the candidate-endpoint theorem.

For Lemma B.2, we use a C2C^2 local right-inverse branch gg, defined on an open neighborhood of the condition simplex, with f(g(c))=cf(g(\boldsymbol{c}))=\boldsymbol{c}. The participating vertex samples g(ci)∈Sig(\boldsymbol{c}_i)\in\mathcal{S}_i lie on this same branch. We assume that ff is LfL_f-Lipschitz on its domain containing x∗\boldsymbol{x}^{\ast} and g(c∗)g(\boldsymbol{c}^{\ast}), and that

\left\\|D^2g(\boldsymbol{c})[\boldsymbol{v},\boldsymbol{v}]\right\\|_2\leq M\left\\|\boldsymbol{v}\right\\|_2^2

for every condition in the simplex and every direction v∈Rd\boldsymbol{v}\in\mathbb{R}^{d}. Taylor expansion and the barycentric identity give

\left\\|\sum_{i=0}^{d}a_i g(\boldsymbol{c}_i)-g(\boldsymbol{c}^{\ast})\right\\|_2\leq\frac{M}{2}\sum_{i=0}^{d}a_i\left\\|\boldsymbol{c}_i-\boldsymbol{c}^{\ast}\right\\|_2^2.

Consequently,

\left\\|f(\boldsymbol{x}^{\ast})-\boldsymbol{c}^{\ast}\right\\|_2\leq L_f\eta+K\max_{0\leq i,j\leq d}\left\\|\boldsymbol{c}_i-\boldsymbol{c}_j\right\\|_2^2,\qquad \eta=\left\\|\boldsymbol{x}^{\ast}-\sum_{i=0}^{d}a_i g(\boldsymbol{c}_i)\right\\|_2,\qquad K=\frac{L_f(d+1)M}{2}.

This retains the original constants and quadratic geometric term, with η\eta measuring deviation from the specified same-branch interpolant. Lemma B.1 incorporates all same-label samples, while Lemma B.2 analyzes conditional error on a common smooth branch.

Appendix C complements these proofs with comparisons between the trained and analytical predictors in single-simplex and multiple-simplex settings. The evaluation uses 1,000 Gaussian noise draws with corresponding target labels. Predictor outputs are compared at identical (x′,t,c∗)(\boldsymbol{x}',t,\boldsymbol{c}^{\ast}) with the same noise parameterization. For generated endpoints, each sample’s minimum candidate MSE is computed and then averaged.

In the two-condition Airfoil setting, labels 0 and 1 define one simplex with multiple training samples per endpoint condition. Average nearest-candidate MSE is on the order of 10−410^{-4}. Both predictor-output and nearest-candidate MSE are largest near the target label farthest from the observed training labels, near the midpoint between the endpoint conditions.

The multiple-simplex evaluation uses a conditional flow matching model trained on real data with their observed condition labels. We compare this trained model with two analytical references: one based on observed vertex conditions, and another grouping multiple nearby-label samples around each vertex. The second construction gives a smaller overall discrepancy between the analytical and trained predictor outputs.

Predictor MSE is approximately 10−310^{-3}–10−210^{-2} at smaller flow times and increases overall as tt approaches one, consistent with the unconditional closed-form comparison in Appendix H.2. Along the label axis, both error measures exhibit multiple local peaks. In the label-wise comparisons, nearest-candidate MSE is numerically of a lower order of magnitude than predictor-output MSE. The candidate distance assesses the full candidate set, while η\eta concerns the specified same-branch interpolant.

2. Diversity among samples satisfying the target conditions

We have recomputed the diversity metric in Eq. (8) using generated samples that satisfy the requested conditions. Each performance-label coordinate is normalized to [0,1][0,1], and samples are retained when

\left\\|f(\boldsymbol{x}^{\ast})-\boldsymbol{c}^{\ast}\right\\|_{\infty}\leq0.05.

Here, f(x∗)f(\boldsymbol{x}^{\ast}) is the true solver-evaluated label on the normalized scale, so every coordinate must have absolute error at most 0.05.

We apply the same tolerance rule across all evaluated methods and datasets. For each target condition, each method uses all its retained samples to calculate average pairwise L2 diversity. We use the same target conditions and aggregation procedure as in Eq. (8). Appendix D.7 reports the protocol and the filtered diversity comparisons for the original and additional baselines across all four datasets.

Averaged across evaluated methods and datasets, the relative decreases in diversity are:

Acquisition iterationMean relative decrease in diversity
03.96%
11.70%
21.05%
30.98%
40.93%
50.98%

Each decrease is the difference between unfiltered and filtered diversity divided by unfiltered diversity. The method rankings remain largely unchanged after filtering. Across the four datasets, QDQ_D retains the highest or near-highest filtered diversity among the acquisition methods at later iterations. These comparisons directly evaluate diversity among generated samples that satisfy the requested target conditions within the common tolerance.

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

This paper studies active learning for conditional flow matching in engineering design, where obtaining performance labels requires expensive simulations. The goal is to select shapes from an unlabeled pool under a limited labeling budget to improve both diversity and conditional accuracy. The authors use an idealized analysis combining a closed form optimal flow matching model with piecewise-linear interpolation in condition space to motivate two query strategies. Q_D aims to improve diversity by selecting different shapes with similar predicted labels and balancing the number of samples across label clusters. Q_A aims to improve accuracy by selecting shapes whose predicted labels are far from existing labels, filling gaps in label space. An RBF network predicts the labels needed for selection. A weighted combination of the strategies allows the balance between diversity and accuracy to be adjusted. The evaluation covers one synthetic dataset and three engineering datasets, with comparisons against random sampling and several active learning baselines over 20 random seeds. Diversity is measured through pairwise distance between generated samples, while accuracy is measured by the error between the target condition and the actual label, obtained through simulation for the engineering tasks. Q_D achieves the highest diversity and Q_A the best accuracy among the compared query strategies across all four datasets, each at the cost of the other metric. The hybrid offers intermediate trade-offs. Additional experiments include ablations, hyperparameter sensitivity, proxy error analysis, and tests under deviations from the theoretical assumptions.

Critical strengths
  1. The empirical results seem very solid to me. For the engineering tasks, the authors run simulations on the generated shapes to check how well they match the requested performance. The evaluation therefore uses the actual performance labels rather than predicted ones, which is more expensive but means that the reported accuracy is not effected by the error of a surrogate model and physically infeasible shapes are excluded automatically. Across all four datasets, each strategy (Q_D, Q_A) is consistently the best on its target metric compared to the benchmark query strategies, with results reported over 20 random seeds.
  2. I like that the paper approaches data selection by considering what the generative model should improve, rather than simply applying uncertainty-based criteria from discriminative models, like related research tends to do, as the authors argue. The resulting principle, that samples with similar labels mainly add diversity and samples with new labels mainly add accuracy, is simple and easy to apply. In my opinion, this gives the paper value beyond the specific query selection rules, as it offers a useful way to think about data selection for different goals in conditional generation.
Critical weaknesses
  1. The writing quality is below what I would expect for a submission. The paper contains grammatical errors and typos throughout the paper (e.g. “which traditionally explored” in the abstract, “which transforming the open querying problem” in Section 1.1, and “to avoid the polished results is differ from their original meaning” in the AI use statement). Also inconsistent terminology appears (e.g. sub region / subregion) and formatting issues (e.g. “01.6” in Table 3). Appendix I also contains an undefined reference “[1]” and the sentence “We included this method as one of the baselines in our original manuscript” looks like a leftover text from a previous revision. Individual errors are rather small but together they make parts of the paper harder to follow.
  2. To me, the connection between the theoretical analysis and the trained model is not clear enough. The strategies are based on the idealized model with closed form flow matching and linear interpolation between training labels. The authors themselves describe this as a “conceptual motivation rather than a hard constraint”, and the evidence that the trained model follows this behavior is mainly qualitative, as shown in Appendix C. Since the analytical framework is presented as one of the main contributions, I would expect a clearer connection to the trained model. This makes the explanation of the results a bit less convincing to me.
Questions for authors
  1. Appendix C gives some qualitative examples of the assumed interpolation behavior. Is there more direct evidence that the trained model behaves as described by the idealized model in Section 2.2? This would help address my main concern about the connection between the analysis and the trained model.
  2. Appendix H.1 reports similar results when using SiLU and GELU, which is explained by the selection logic being independent of the architecture. How does this relate to the piecewise-linear assumption used in the analysis?
  3. Which aspects of Q_D and Q_A are specific to flow matching, and which could also apply to conditional generative models in general?
  4. How are the generated shapes for which the solver returns NaN handled when calculating diversity? Are they excluded, and how often does this happen for Q_D, Q_A and the baselines?
5 comments

We thank the reviewer for the constructive comments and the positive assessment of the solver-based evaluation and experimental study. We have expanded the quantitative comparison between the analytical model and the trained network, clarified the roles of the theoretical assumptions, and revised the presentation.

1. Direct quantitative evidence connecting the trained model to the analytical model

We have expanded Appendix C with quantitative comparisons in both single-simplex and multiple-simplex settings. These comparisons examine predictor outputs and generated endpoints separately.

For the predictor comparison, we evaluate the trained noise predictor ut(x′,c∗)\boldsymbol{u}_t(\boldsymbol{x}',\boldsymbol{c}^{\ast}) and the analytical predictor at identical (x′,t,c∗)(\boldsymbol{x}',t,\boldsymbol{c}^{\ast}), using the same noise parameterization. The analytical predictor combines the closed-form outputs at the simplex vertices:

u‾t(x′,c∗)=∑i=0daiutCF(x′,ci),c∗=∑i=0daici,ai≥0,∑i=0dai=1.\overline{\boldsymbol{u}}_t(\boldsymbol{x}',\boldsymbol{c}^{\ast})=\sum_{i=0}^{d}a_i\boldsymbol{u}_t^{\mathrm{CF}}(\boldsymbol{x}',\boldsymbol{c}_i),\qquad \boldsymbol{c}^{\ast}=\sum_{i=0}^{d}a_i\boldsymbol{c}_i,\qquad a_i\geq0,\qquad \sum_{i=0}^{d}a_i=1.

Here, CF denotes “closed form,” and utCF\boldsymbol{u}_t^{\mathrm{CF}} is the closed-form noise predictor in Eq. (1). The evaluation uses 1,000 Gaussian noise draws with corresponding target labels. We report the MSE between the trained and analytical predictor outputs as a function of flow time and target label.

For the endpoint comparison, we compute each generated sample’s minimum MSE to the theoretical candidate set S(c∗)\mathcal{S}(\boldsymbol{c}^{\ast}) in Eq. (3), then average these sample-wise minimum errors over generated samples. This measures the proximity of generated endpoints to the candidate combinations predicted by the analysis.

In the original two-condition Airfoil setting, labels 0 and 1 define a single one-dimensional simplex, with multiple training data points at each endpoint condition. The average nearest-candidate MSE is on the order of 10−410^{-4}. Both predictor-output MSE and nearest-candidate MSE are largest near the target label farthest from the observed training labels, near the midpoint between the two endpoint conditions.

The multiple-simplex evaluation uses a conditional flow matching model trained on real data with their observed condition labels. We compare this trained model with two analytical references. The first uses the observed vertex conditions. The second groups multiple samples with nearby labels around each vertex, approximating nearby conditions by a shared vertex label. The second construction gives a smaller overall discrepancy between the analytical and trained predictor outputs, providing a closer empirical approximation to the trained model.

We use the paper’s time convention throughout: t=0t=0 corresponds to Gaussian noise and t=1t=1 to data. In the multiple-simplex comparisons, predictor MSE is approximately 10−310^{-3}–10−210^{-2} at smaller flow times and increases overall as tt approaches one. This trend agrees with the unconditional closed-form comparison in Appendix H.2 under the same time convention. Along the label axis, both error measures exhibit multiple local peaks.

In the label-wise comparisons, average nearest-candidate MSE is numerically of a lower order of magnitude than predictor-output MSE. These measurements quantify agreement at the predictor and generated-endpoint levels, respectively.

Section 2.2 now explicitly states the condition-wise affine assumption and its relationship to the trained network. Appendix B provides the candidate-endpoint and branch-wise conditional-error arguments, while Appendix C quantifies the relationship between the analytical construction and the trained model.

2. Relationship between the CPWL assumption and the SiLU/GELU results

The analytical model in Section 2.2 combines closed-form vertex predictors with a piecewise-affine extension in the condition variable. Within each simplex, the condition-wise affine assumption specifies the interpolation identity used in the analysis. The CPWL network setting motivates this structure, and the manuscript explicitly states the additional correspondence required between condition simplices and network affine regions.

The acquisition rules operate on the dataset and proxy-predicted labels. QDQ_D combines label proximity, label-cluster balance, and input-space novelty, while QAQ_A promotes condition-space coverage. Their scores can therefore be computed independently of the conditional generator’s activation function. The generator architecture enters through the model trained on the selected data.

Appendix H.1 examines this distinction by replacing LeakyReLU with SiLU and GELU. The similar performance trends across these activations show that the selection rules remain effective for the tested smooth-activation networks.

Section 2.2 and Appendix H.1 now distinguish these roles explicitly: the condition-wise piecewise-affine model supplies the analytical motivation, and the activation-function experiments evaluate the selection rules under alternative network parameterizations.

3. Flow-matching-specific analysis and applicability to other conditional generators

The flow-matching-specific component is the analytical connection between training data and conditional sampling dynamics. Equation (1) expresses the closed-form noise predictor through posterior weights over the empirical training data. Combining these vertex predictors with condition-wise affine interpolation gives the weighted flow analyzed in Appendix B and its candidate-endpoint characterization.

This connects same-label sample multiplicity and condition-space interpolation to the sampling behavior of conditional flow matching.

The acquisition expressions in Eqs. (4) and (6) use quantities with broader applicability: label distances, label-cluster entropy, and input-space distances. These quantities can also be computed when selecting training data for other conditional generators. The branch-wise interpolation-error bound likewise uses general smoothness arguments.

Sections 2.2–2.4 now clarify this distinction: flow matching provides the sampling mechanism underlying our analysis, while the resulting acquisition scores express more general preferences for conditional data selection. Our experimental evidence concerns conditional flow matching; applying and evaluating these criteria with other conditional-generator families is a direction for future work.

4. Solver NaNs and the diversity evaluation

All generated shapes undergo minor geometric adjustments before solver evaluation to satisfy the solver’s geometric input requirements. Across the reported evaluations, no solver NaNs were observed for QDQ_D, QAQ_A, or the baselines. The observed NaN incidence was therefore zero, and no samples were excluded because of solver NaNs.

Section 3.1 now describes this preprocessing step and reports the observed solver outcomes. We have also clarified that validity refers to the geometric input requirements and solver evaluation used in this study.

We additionally report diversity among samples satisfying the requested conditions in Appendix D.7. Each performance-label coordinate is normalized to [0,1][0,1], and samples are retained when

\left\\|f(\boldsymbol{x}^{\ast})-\boldsymbol{c}^{\ast}\right\\|_{\infty}\leq0.05.

Here, f(x∗)f(\boldsymbol{x}^{\ast}) is the true solver-evaluated label on the normalized scale. The absolute error must therefore be at most 0.05 in every label coordinate. For each target condition, we recompute the average pairwise L2 diversity using all retained samples, with the same target conditions and aggregation procedure as Eq. (8).

Averaged across evaluated methods and datasets, the relative decreases in diversity at acquisition iterations 0–5 are 3.96%, 1.70%, 1.05%, 0.98%, 0.93%, and 0.98%, respectively. Each decrease is measured relative to the corresponding unfiltered diversity. The method rankings remain largely unchanged. Across the four datasets, QDQ_D retains the highest or near-highest filtered diversity among the acquisition methods at later iterations.

5. Writing, notation, and presentation

We have corrected the grammatical issues identified in the abstract, Section 1.1, and the AI Use Statement, and performed an additional language-editing pass throughout the manuscript. These revisions address word order, articles, sentence structure, and terminology while preserving the technical content.

We have standardized “subregion,” corrected the typographical issues in the hyperparameter tables, and unified vector notation and symbol definitions.

In Appendix I, we have replaced the undefined reference “[1]” with the corresponding bibliographic citation and revised the sentence referring to an “original manuscript” to describe the baseline comparison directly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Summary

We thank the reviewers for their careful reading and constructive discussion.

This work presents a pioneering exploration of active learning for conditional flow matching models in engineering inverse problems, where obtaining performance labels is costly. We investigate how actively acquiring a limited number of performance labels can improve the generative model’s diversity and conditional accuracy. Starting from the geometry of the training data, we develop an analytical framework connecting data distributions to conditional generation behavior and derive acquisition strategies targeting these two objectives. Our main experiments cover four datasets, classical and recent active learning baselines, and 20 random seeds, providing an analytical perspective and methodological foundation for active data acquisition in this setting.

During the discussion, we strengthened the paper in the following respects.

1. Clarifying the theoretical assumptions and quantifying the connection between the analytical construction and trained models.

In Section 2.2 and Appendix B, we clarified the assumptions required for conditional interpolation and the subsequent theoretical results. For Lemma B.1, we supplemented the proof for the ideal weighted closed-form flow, covering vertex-specific closed-form predictors and multiple training samples sharing the same condition. For Lemma B.2, we explicitly stated the assumption that the samples participating in interpolation belong to a common smooth local right-inverse branch.

Appendix C now includes single-simplex experiments and multi-simplex experiments using models trained on real data. These experiments compare trained-model outputs with interpolated closed-form outputs and measure the distance from generated samples to the theoretical candidate set. They quantify the discrepancy between trained models and the ideal analytical construction, providing empirical support for the theoretical interpretation in the evaluated settings.

2. Evaluating both condition satisfaction and diversity among samples satisfying the target conditions.

In Appendix D.7, we filter generated samples using their true evaluated labels. Each label coordinate is normalized to [0,1], and a sample passes when the absolute error in every coordinate is at most 0.05. The method rankings remain largely unchanged after filtering, and QDQ_D continues to achieve the highest or near-highest diversity in later iterations.

We additionally report condition satisfaction curves from iteration 0 through iteration 5 and provide the complete final-iteration statistics in our response. At the final iteration, all evaluated methods achieve mean condition satisfaction rates above 92% on all four datasets. Every evaluated case contains at least two passing samples, and no target conditions or random seeds are excluded because of insufficient passing samples. These additions jointly characterize the proportion of samples satisfying the requested conditions and the diversity among those samples.

3. Further analyzing dependence on proxy models and their hyperparameters.

Appendix G expands the comparison to four proxy families: RBF, Random Forest, SVR, and XGBoost. It also includes first-round rankings against all baselines, with QDQ_D ranked by diversity and QAQ_A ranked by accuracy.

Pooling configurations across the four proxy families, QDQ_D achieves a Top-3 ranking in 92.71% of randomly sampled, untuned configurations and 87.50% of high-error configurations. The corresponding rates for QAQ_A are 26.04% and 5.00%. These results show that QDQ_D remains competitive even with random hyperparameters, whereas QAQ_A requires careful selection and tuning of proxy hyperparameters. Proxy hyperparameter tuning is standard practice; the random-configuration experiments further reveal the different sensitivities of the two acquisition objectives to proxy configurations. Improving proxy estimation and acquisition reliability with limited initial data is a direction for future work.

4. Improving implementation details and presentation.

We clarified that the hybrid strategy applies min-max normalization separately to the two acquisition scores over the candidate pool at each iteration. Its weight expresses a directional preference for diversity or accuracy. Results across datasets also show that this weight does not directly provide quantitative control over the resulting trade-off. Developing methods that provide such quantitative control is future work.

We also systematically addressed the reviewer-identified issues in grammar, spelling, terminology, symbol definitions, references, and table formatting to improve clarity and consistency.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.