acceptodds
Under review as a conference paper at ICLR 2027

DriftCF: Generative Modeling of Joint Potential Outcome Distributions via One-Step Particle Drifting

12 views

Abstract

Interventions often change several dependent outcomes, making the treatment-specific joint potential-outcome distribution the relevant object for uncertainty, dependence, and joint risk. Learning this law is difficult because each unit reveals only one factual outcome vector, while the missing counterfactual distribution must preserve its complete dependency structure. Existing expressive generators obtain samples through iterative denoising, numerical integration, or ordered generation. We introduce Drifting-based CounterFactual generation (DriftCF). Causal identification turns the target into conditional joint-distribution matching; DriftCF solves it by attracting complete generated vectors toward factual outcomes and repelling particles according to the model's own concentration. Detached regression amortizes this correction into a non-autoregressive generator that produces one complete vector in one forward pass. We establish the field's smoothed-score form, target fixed point, and amortized transport property. Experiments test joint-law recovery, dependence, the attraction–repulsion mechanism, and one-step efficiency on multi-outcome benchmarks and an online incentive-allocation study. Code is available in the Supplementary Material.

Primary area:
generative models
Keywords:
Counterfactual, generative modeling, joint potential outcome distributions, particle drifting

Reviews

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

This paper proposes DriftCF, a generative model for estimating the joint distribution of multiple potential outcomes under different treatments. Instead of predicting outcomes independently, the method tries to capture their dependencies using an attraction-repulsion mechanism that moves generated samples toward the target distribution. The model is trained to generate the full outcome vector in a single forward pass, avoiding the iterative sampling required by diffusion and flow-based approaches.

The authors provide theoretical analysis of the proposed mechanism and evaluate it on several semi-synthetic benchmarks, including ACIC, MIMIC, IHDP, and News. They also report an industrial online experiment on incentive allocation. The results show improvements in joint distribution estimation and inference efficiency, along with some positive online business metrics.

Critical strengths

The problem is well motivated. Joint potential outcome modeling is useful for decisions involving multiple correlated outcomes, and the paper makes a reasonable case for why estimating marginal distributions separately is insufficient.

The one-step generation design is another practical advantage. It makes the approach potentially more suitable for online decision systems where iterative sampling can be expensive. The authors also evaluate both distribution quality and downstream decision performance, rather than focusing only on prediction metrics.

Critical weaknesses

The main issue for me is the methodological novelty. The attraction-repulsion mechanism and distribution matching formulation have clear connections to existing particle-based generative methods. The paper provides theoretical analysis, but I am still not convinced that the proposed formulation represents a substantial advance over these methods. The authors should make this distinction clearer, ideally with more direct comparisons.

I also find the experimental evidence somewhat insufficient to support the broader claims. The offline results are encouraging, but it is difficult to judge whether the adapted generative baselines are equally well optimized. Since efficient one-step generation is a major contribution, a more systematic quality-versus-computation comparison would be useful.

The online experiment is interesting, but the comparison is against a point-estimate policy. This makes it difficult to isolate the contribution of DriftCF itself from the general benefit of distribution-aware optimization. The reported improvements in watching time and revenue are positive, but without a stronger distributional baseline, the evidence for the specific modeling approach remains limited.

Questions for authors
  1. What is the main technical difference between DriftCF and existing particle-based distribution matching or transport methods? Can the authors clarify which part of the proposed formulation is genuinely new?

  2. Were the generative baselines tuned with comparable computational budgets? It would be useful to see the trade-off between distribution quality and inference cost under consistent settings.

  3. For the online experiment, is there a reason the authors did not compare against another distributional model using the same allocation objective? Such a comparison would help establish whether the improvement comes from DriftCF or from distribution-aware decision-making in general.

4 comments
AuthorsOpenReview ↗

Rebuttal [1 / 2]

We thank the reviewer for the thoughtful, constructive comments on novelty relative to particle matching/transport and comparable computational budgets. We hope the analysis, quality–cost comparisons, and online-scope clarification below address these concerns. These concerns are addressed in the updated PDF. Revisions addressing the reviews are highlighted in blue.

Q1. Technical difference from particle distribution matching and transport

Our treatment-aware conditional-drift formula connects generator stationarity to the identified intervention law. Treatment inputs select the requested law; weighted interactions select factual support; distributional representations retain information for recovery. Under Proposition 3.1, the target is Pt(⋅∣x)=L(Y(t)∣X=x)=P(Y∣X=x,T=t)P_t(\cdot\mid x)=\mathcal L(Y(t)\mid X=x)=P(Y\mid X=x,T=t). We establish when conditioning on (ϕ(X),T)(\phi(X),T) preserves it.

For H=ϕ(X)H=\phi(X), the representation-conditional law is

P(Y∣H,T)=E[P(Y∣X,T)∣H,T].P(Y\mid H,T) =\mathbb{E}[P(Y\mid X,T)\mid H,T].

(Y1)

(Y1) identifies compression-induced conditional-law mixing.

Representation and generation errors separate exactly for finite conditional KL quantities:

EX,TKL⁡(PX,T∣QH,T)=I(Y;X∣H,T)+EH,TKL⁡(PH,Tϕ∣QH,T).\mathbb{E}_{X,T}\operatorname{KL}(P_{X,T}\\|Q_{H,T}) ={}I(Y;X\mid H,T)\\\\ +\mathbb{E}_{H,T}\operatorname{KL}(P^\phi_{H,T}\\|Q_{H,T}).

(Y2)

Here Px,t=L(Y∣X=x,T=t)P_{x,t}=\mathcal L(Y\mid X=x,T=t), Ph,tϕ=L(Y∣H=h,T=t)P^\phi_{h,t}=\mathcal L(Y\mid H=h,T=t); expectations use observational contexts. Inserting p(Y∣H,T)p(Y\mid H,T) into the log ratio and iterating expectation proves (Y2). Its terms measure distributional information loss and generation error; ODS zeros the first (Proposition A.3).

The representation objective targets the same distributional object

For adapted outcomes ZZ, multivariate energy score gives

EZ∼PSES(Q,Z)−EZ∼PSES(P,Z)=12DE(P,Q),\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(Q,Z) -\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(P,Z) =\tfrac12D_{\mathrm E}(P,Q),

where

DE(P,Q)=2E∣ZP−ZQ∣−E∣ZP−ZP′∣−E∣ZQ−ZQ′∣.D_{\mathrm E}(P,Q)=2\mathbb{E}\\|Z_P-Z_Q\\| -\mathbb{E}\\|Z_P-Z'_P\\|-\mathbb{E}\\|Z_Q-Z'_Q\\|.

Score expansion proves the identity. With finite first moments, strict propriety [1] yields, for optimal population risks over all conditional laws,

Rϕ∗−RX∗=12EX,TDE(PX,TZ,PH,TZ,ϕ).R_\phi^*-R_X^* =\tfrac12\mathbb{E}_{X,T} D_{\mathrm E}(P^Z_{X,T},P^{Z,\phi}_{H,T}).

(Y3)

The population encoder gap is exactly a full conditional-law discrepancy; the distance-induced kernel gives its RKHS form [2]. This motivates energy-score training and distributional diagnostics.

For centered bivariate Gaussian Y∣X=x,T=tY\mid X=x,T=t with unit variances, correlation xρtx\rho_t, and X∈−1,1X\in\\{-1,1\\} independent of TT, a constant representation preserves means, univariate marginals, and treatment balance but merges distinct joint laws—the information distributional training retains.

The conditional operator and its fixed point

Our treatment-weighted attraction/repulsion formula has population form V=MPkc−MQkcV=M_P^{k_c}-M_Q^{k_c}, with kc=kxkTk_c=k_xk_T inside both separately normalized expectations. Treatment weights determine the corrected distribution as well as conditioning generation. We use

P(dc,dz)=ν(dc)Pc(dz),Q(dc,dz)=ν(dc)Qc(dz),P(dc,dz)=\nu(dc)P_c(dz),\qquad Q(dc,dz)=\nu(dc)Q_c(dz),

with c=(ϕ(x),t)c=(\phi(x),t) and fixed kc=kxkTk_c=k_xk_T. Gaussian outcome mean shift gives [4]

Vh(c,z)=hz2∇zlog⁡ShP(c,z)ShQ(c,z).V_h(c,z) =h_z^2\nabla_z\log\frac{S_hP(c,z)}{S_hQ(c,z)}.

(Y4)

Under Theorem 4.4, zero field implies ShP(c,z)=a(c)ShQ(c,z)S_hP(c,z)=a(c)S_hQ(c,z). The shared context marginal normalizes it:

∫ShP(c,z)dz=Ch∫kc(c,c′)ν(dc′)=∫ShQ(c,z)dz,\int S_hP(c,z)dz =C_h\int k_c(c,c')\nu(dc') =\int S_hQ(c,z)dz,

Integrals cover the full outcome domain; ChC_h is the Gaussian kernel integral. Thus a(c)=1a(c)=1; injectivity gives Pc=QcP_c=Q_c almost everywhere. Lossless decoding, ODS, and identification yield

V_h=0 \Longrightarrow d_\\#Q^Z_{\phi(x),t} =\mathcal L(Y(t)\mid X=x)

(Y5)

almost everywhere on supported contexts.

Frozen representation, normalization, and bandwidths fix each branch's smoothing operator throughout training. Cross-fitting estimates geometry on the complementary fold, implementing the proof's operator/representation conditions.

From particle correction to generator optimization

For stacked particles zθ\mathbf z_\theta, clipped field u\mathbf u, and generator Jacobian JJ, detached regression [3] gives

∇θL=−2ηNJ⊤u,Δz=2αηNJJ⊤u+o(α).\nabla_\theta L=-\frac{2\eta}{N}J^\top\mathbf u, \qquad \Delta\mathbf z =\frac{2\alpha\eta}{N}JJ^\top\mathbf u+o(\alpha).

(Y6)

Consequently,

⟨u,Δz⟩=2αηN∣J⊤u∣2+o(α).\langle\mathbf u,\Delta\mathbf z\rangle =\frac{2\alpha\eta}{N}\\|J^\top\mathbf u\\|^2+o(\alpha).

This is particle motion through the generator tangent operator. Population zero semi-gradient gives J∗u=0\mathcal J^*u=0. Tangent coverage ensures u∈range⁡(J)‾u\in\overline{\operatorname{range}(\mathcal J)}; orthogonality then gives u=0u=0. Radial clipping preserves zeros, so Vh=0V_h=0 and (Y5) applies.

Thus stationarity identifies the causal target under tangent coverage, ODS, decoding, and identification.

AuthorsOpenReview ↗

Rebuttal [2 / 2]

Ideal velocity v=∇log⁡p−∇log⁡qv=\nabla\log p-\nabla\log q gives a transport view. With ∂τq+∇⋅(qv)=0\partial_\tau q+\nabla\cdot(qv)=0, smooth positive densities, and vanishing boundary flux,

ddτECKL⁡(QC∣PC)=−EC,Z∼QC∣v(C,Z)∣2.\frac{d}{d\tau}\mathbb{E}_C\operatorname{KL}(Q_C\\|P_C) =-\mathbb{E}_{C,Z\sim Q_C}\\|v(C,Z)\\|^2.

(Y8)

Integration by parts proves ideal score transport (Y8); (Y6) describes implemented motion, whose stationary target follows under tangent coverage. [5] develops the Wasserstein interpretation.

Direct comparisons with related methods

For common conditional Gaussian-kernel means mP,mQm_P,m_Q, half-squared MMD induces ∇mP−∇mQ\nabla m_P-\nabla m_Q [6]; our separately normalized correction is

hz2[∇log⁡mP−∇log⁡mQ].h_z^2\left[\nabla\log m_P-\nabla\log m_Q\right].

Normalization generally changes direction. CGMMD [7] estimates conditional MMD with nearest neighbors; GDR-learners [8] target nuisance robustness through orthogonal risks. Our construction preserves the causal target through representation and equilibrium. At NFE = 1 for both, Table C1 gives DriftCF/CGMMD Dependence 0.07/0.13, supporting the complete-method distinction.

Q2. Comparable budgets and quality–computation trade-offs

Complete methods share splits, tuning budget, evaluation draws, and outcome normalization. Objective controls also fix initialization, representation, particles, factual minibatches, kernels, adapter, architecture, batch construction, and computation budget. Cost is per complete joint draw under common hardware/query settings.

Table C2 gives median latency 2.82/3.17 ms for DriftCF/one-step CGMMD; Tables C1–C2 give full quality–cost results, including p95 latency and memory.

Q3. Online distributional comparisons and model-specific decision value

The online experiment tests the complete upgrade from deployed point estimates to DriftCF's distributional policy. Policies share eligible population, candidate incentives, budget accounting, and serving constraints. Table 5 gives 0.504% relative ad-revenue lift with randomization-unit-level testing, supporting the policy upgrade.

Exploratory fixed-objective comparisons used a separate low-traffic pool and different policy configuration. Table 5 therefore measures policy-level gains, not fixed-objective generator attribution; offline Tables C1–C2 complement it but do not supply that matched online comparison.

A distributional explanation of decision value

For a fixed query and finite candidate treatments, let Pt,QtP_t,Q_t be true/learned outcome laws. Suppose ut∈Hu_t\in\mathcal H, ∣ut∣H≤B\\|u_t\\|_{\mathcal H}\le B, and

sup⁡tMMD⁡k(Pt,Qt)≤ϵ.\sup_t\operatorname{MMD}_k(P_t,Q_t)\le\epsilon.

The reproducing property and Cauchy–Schwarz yield

∣EPtut−EQtut∣≤Bϵ.|\mathbb{E}_{P_t}u_t-\mathbb{E}_{Q_t}u_t|\le B\epsilon.

For true-utility maximizer t∗t^* and model-utility maximizer t^\hat t over the same feasible set, adding/subtracting model utilities gives

VP(t∗)−VP(t^)≤2Bϵ.V_P(t^*)-V_P(\hat t)\le2B\epsilon.

(Y9)

(Y9) links joint-law accuracy to decision value: smaller discrepancy tightens treatment-selection regret for the stated RKHS utilities and common feasible set, motivating dependence-preserving allocation and joint-law evaluation.

References

[1] Tilmann Gneiting and Adrian E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477):359–378, 2007.

[2] Dino Sejdinovic, Bharath Sriperumbudur, Arthur Gretton, and Kenji Fukumizu. Equivalence of Distance-Based and RKHS-Based Statistics in Hypothesis Testing. Annals of Statistics, 41(5):2263–2291, 2013.

[3] Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative Modeling via Drifting. arXiv:2602.04770, 2026.

[4] Erkan Turan and Maks Ovsjanikov. Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective. arXiv:2603.09936, 2026.

[5] Arthur Gretton, Li Kevin Wenliang, Alexandre Galashov, James Thornton, Valentin De Bortoli, and Arnaud Doucet. On the Wasserstein Gradient Flow Interpretation of Drifting Models. arXiv:2605.05118, 2026.

[6] Michael Arbel, Anna Korba, Adil Salim, and Arthur Gretton. Maximum Mean Discrepancy Gradient Flow. NeurIPS, 2019.

[7] Anirban Chatterjee, Sayantan Choudhury, and Rohan Hore. One-shot Conditional Sampling: MMD meets Nearest Neighbors. arXiv:2509.25507, 2025; ICML, 2026.

[8] Valentyn Melnychuk and Stefan Feuerriegel. GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes. ICLR, 2026. arXiv:2509.22953.

(moved to the correct reply section) Thank you for the detailed rebuttal and additional analysis. The new quality–computation comparisons help address my concerns about baseline evaluation. It is also helpful to clarify that the online gains reflect policy-level improvements, rather than the isolated contribution of DriftCF. However, my main concern about novelty remains only partly addressed. The theory supports useful properties of the conditional drift formulation, but I am still not fully convinced that it provides a substantial algorithmic advance over existing drifting and particle-based distribution-matching methods.

The rebuttal strengthens the paper, but I am inclined to keep my original rating.

Thank you for the thoughtful follow-up and for recognizing the strengthened evaluation. The algorithmic advance is a treatment-aware conditional correction designed to recover joint counterfactual outcomes from selectively observed factual vectors. This setting requires preserving cross-outcome dependence under each intervention: marginal-product approaches miss that dependence, while generic joint generators and particle-matching objectives do not by themselves resolve which factual observations should contribute to an intervention-specific correction or which covariate information must survive representation learning. DriftCF addresses these requirements within the update itself: the treatment kernel enters both factual attraction and model repulsion, using exact arm masks or local dose weights to construct intervention-specific interactions; the multivariate energy-score representation retains distributional effect modifiers beyond conditional means; and the separately normalized, full-vector kernel interactions produce a conditional score-difference correction that responds to dependence errors even when every marginal is correct.

These choices change the observations entering each update, the geometry of their interactions, and the resulting correction direction—not simply the conditioning input of an existing generator. Relative to conditional MMD, separate normalization generally changes the update direction rather than merely rescaling it; our analysis establishes when this treatment-aware operator recovers the original covariate-conditional causal law. Amortizing these corrections yields complete joint draws in one forward pass. Tables C1–C2 support the practical value of this construction: compared with the one-step CGMMD baseline, DriftCF achieves lower errors on all four benchmarks, Dependence 0.07 versus 0.13, and median sampling latency 2.82 versus 3.17 ms. We hope this clarifies the problem-specific algorithmic contribution underlying the theoretical and empirical results.

Reviewer_7qCtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns treatment-specific joint potential-outcome distributions with a one-step conditional particle-drifting generator. It provides a fixed-point analysis and evaluates the method on four benchmarks and an online incentive-allocation study.

Critical strengths

The paper addresses a practically relevant problem: learning joint treatment-outcome distributions when decisions depend on several correlated outcomes.

Critical weaknesses

The gains in Table 1 are not very clearly attributable to the attraction–repulsion objective, because lack of statement on whether the joint diffusion and flow-matching baselines use the same mixed-outcome adapter and frozen encoder. Table 3 shows only a small Energy gap between conditional flow matching and DriftCF (1.17 vs. 1.14), although the Dependence gap is larger (0.13 vs. 0.07).

Questions for authors
  1. How does DriftCF's context-weighted attraction–repulsion objective differ from the nearest-neighbour conditional distribution matching in CGMMD [a]?
  2. For ACIC/MIMIC W₁ and the covariance/Spearman errors, are the metrics computed at each fixed (x,t) and then averaged, or after pooling different test units? If pooled, how do they distinguish conditional-law recovery from a covariate mixture?
  3. What settings define Tables 2 and 3? Their shared variants have different values, and Table 3 reports a flow-matching field cosine although Table 2 says this field metric is inapplicable to flow matching.
  4. Proposition A.3 requires ϕ(X)\phi(X) to preserve the full conditional outcome distribution. Appendix C says this is assessed through held-out distributional diagnostics, but I could not find their results; what evidence supports this condition in the reported experiments?

References

[a] One-shot Conditional Sampling: MMD meets Nearest Neighbors. ICML 2026.

4 comments
AuthorsOpenReview ↗

Rebuttal [1 / 2]

We appreciate the reviewer's constructive comments on objective attribution, representation sufficiency, and evaluation protocols. We hope the controlled comparisons, derivations, and diagnostics below resolve these concerns. These concerns are addressed in the updated PDF. Revisions addressing the reviews are highlighted in blue.

W1. Shared components and attribution of the training objective

Table 1 shares covariate encoding, splits, tuning budgets, evaluation draws, and outcome normalization; encoder optimization, corrections, and adapters remain model-specific, including end-to-end training where used.Our objective control fixes initialization, representation, particles, factual minibatches, kernels, adapter, architecture, batches, and compute; only training direction varies. Dependence 0.07 versus 0.21 for DriftCF/conditional MMD gradient (Table 3) supports objective-level attribution.

W2. Interpreting Energy and Dependence jointly

Energy score assesses the full law; Dependence targets covariance/rank correlations. (R6) relates energy score to energy distance. Corrected Table 3 gives DriftCF/flow-matching Energy/Dependence 1.14/0.07 versus 1.17/0.19: a small overall-score gain accompanies stronger dependence recovery. Common Tables C1–C2 compare full methods, including GDR-learners [6].

Q1. Difference from nearest-neighbor conditional MMD

Our separately normalized data/model score correction differs from CGMMD's MMD gradient. Treatment conditions generation/interactions; shared contexts enable causal identification at equilibrium.

CGMMD [1] minimizes expected conditional MMD,

EC∣μPc−μQθ,c∣H2,\mathbb{E}_C\\|\mu_{P_c}-\mu_{Q_{\theta,c}}\\|_{\mathcal H}^2,

estimated by nearest neighbors for one-shot vector generation. DriftCF fits detached, separately normalized mean-shift displacements.

At a common conditional law and Gaussian kernel, define

mP(z)=EZ∼Pckz(z,Z),mQ(z)=EZ∼Qckz(z,Z).m_P(z)=\mathbb{E}_{Z\sim P_c}k_z(z,Z), \qquad m_Q(z)=\mathbb{E}_{Z\sim Q_c}k_z(z,Z).

For F(Q)=12MMD⁡2(Pc,Q)F(Q)=\tfrac12\operatorname{MMD}^2(P_c,Q), the first variation is δF/δQ=mQ−mP\delta F/\delta Q=m_Q-m_P up to a constant, giving Wasserstein descent velocity [2]

vMMD(z)=∇mP(z)−∇mQ(z).v_{\mathrm{MMD}}(z)=\nabla m_P(z)-\nabla m_Q(z).

(R1)

Since ∇zkz(z,Z)=kz(z,Z)(Z−z)/hz2\nabla_zk_z(z,Z)=k_z(z,Z)(Z-z)/h_z^2, DriftCF gives

vDrift(z)=hz2[∇log⁡mP(z)−∇log⁡mQ(z)].v_{\mathrm{Drift}}(z) =h_z^2\left[\nabla\log m_P(z)-\nabla\log m_Q(z)\right].

(R2)

Normalization changes direction: for independent a=∇mPa=\nabla m_P, b=∇mQb=\nabla m_Q and mP≠mQm_P\ne m_Q, proportionality of a/mP−b/mQa/m_P-b/m_Q to a−ba-b requires one scalar to equal both 1/mP1/m_P and 1/mQ1/m_Q. (R1)–(R2) are population operators; CGMMD's empirical gradient also depends on its neighbor graph.

For context-smoothed population laws,

Vh(c,z)=hz2∇zlog⁡(ShP)(c,z)(ShQ)(c,z).V_h(c,z)=h_z^2\nabla_z\log\frac{(S_hP)(c,z)}{(S_hQ)(c,z)}.

(R3)

Gaussian drifting gives this score interpretation [3]. Treatment conditions generation and field weights: binary/unordered arms use within-arm interactions; continuous doses use local weights with leading symmetric-smoothing bias O(ht2)O(h_t^2) under Proposition A.2's interior smoothness assumptions.

ODS connects equilibrium to the original conditional law (Propositions A.3, A.5; Q4).

By the chain rule, −∇θL=∑lJl⊤(−∇zlL)-\nabla_\theta L=\sum_lJ_l^\top(-\nabla_{z_l}L). DriftCF fits (R2); CGMMD uses nearest-neighbor MMD. Table 3 fixes geometry; C1–C2 compare full methods.

At NFE = 1, C1 gives DriftCF/CGMMD Dependence 0.07/0.13, supporting full-pipeline recovery.

Q2. Conditional-law recovery and covariate mixtures

Oracle/model samples are compared separately at each query (x,t)(x,t), then discrepancies averaged—never pooled across queries. Covariance/Spearman errors use the same ordering.

For query set Q\mathcal Q, the estimand requires

EDcond=1∣Q∣∑(xi,ti)∈QD(Pxi,ti,Qxi,ti).\mathcal E_D^{\mathrm{cond}} =\frac1{|\mathcal Q|}\sum_{(x_i,t_i)\in\mathcal Q} D(P_{x_i,t_i},Q_{x_i,t_i}).

(R8)

Pooling instead gives

EDpool=D(1n∑iPxi,ti,1n∑iQxi,ti).\mathcal E_D^{\mathrm{pool}} =D\left(\frac1n\sum_iP_{x_i,t_i},\frac1n\sum_iQ_{x_i,t_i}\right).

(R9)

For equal-weight queries P1=δ−1P_1=\delta_{-1}, P2=δ1P_2=\delta_{1}, Q1=δ1Q_1=\delta_{1}, Q2=δ−1Q_2=\delta_{-1}, pooled laws coincide but conditional W1W_1 averages to 2. We use (R8), not (R9), preventing cancellation.

Q3. Settings and field metrics in Tables 2 and 3

Endpoint correction differs from flow velocity. Table 2 uses

V∗(c,z)=hz2∇zlog⁡ShP(c,z)ShQ(c,z),V^*(c,z)=h_z^2\nabla_z\log\frac{S_hP(c,z)}{S_hQ(c,z)},

at common particles. Flow matching learns interpolation-time velocity; its endpoint fields are n/a in Tables 2–3. Final energy score, Dependence, and sampling cost assess its distribution.

Shared Dependence entries are corrected. Table 2 fully retrains per query; Table 3 fixes encoders except in encoder ablations. Both average within-query metrics. Table 2 gives L2 mean/mode-mass errors; Table 3 gives energy score/effective support normalized to DriftCF. Both use Field cosine.

AuthorsOpenReview ↗

Rebuttal [2 / 2]

Q4. Distributional sufficiency of the learned representation

ODS is the exact-recovery condition; diagnostics assess retention empirically. Matched-capacity treatment-conditional probes (C3; PDF Table 6, Appendix E.5) give ES gaps from raw covariates 0.02–0.03 for distributional versus 0.17–0.26 for mean-trained representations, supporting retention without proving ODS.

How the representation objective supports ODS

Energy score targets ODS's full-law information. Representation and generation errors separate:

Let H=ϕ(X)H=\phi(X), Px,t=L(Y∣X=x,T=t)P_{x,t}=\mathcal L(Y\mid X=x,T=t), Ph,tϕ=L(Y∣H=h,T=t)P^\phi_{h,t}=\mathcal L(Y\mid H=h,T=t). Then

Ph,tϕ=E[PX,t∣H=h,T=t].P^\phi_{h,t}=\mathbb{E}[P_{X,t}\mid H=h,T=t].

(R4)

For generator Qh,tQ_{h,t} with finite KL quantities,

EX,TKL⁡(PX,T∣QH,T)=EX,TKL⁡(PX,T∣PH,Tϕ)+EH,TKL⁡(PH,Tϕ∣QH,T).\begin{aligned} \mathbb{E}_{X,T}\operatorname{KL}(P_{X,T}\\|Q_{H,T}) ={}&\mathbb{E}_{X,T}\operatorname{KL}(P_{X,T}\\|P^\phi_{H,T})\\\\ &+\mathbb{E}_{H,T}\operatorname{KL}(P^\phi_{H,T}\\|Q_{H,T}). \end{aligned}

(R5)

Proof. Insert p(Y∣H,T)p(Y\mid H,T) into the likelihood ratio and iterate expectation. The first term, I(Y;X∣H,T)I(Y;X\mid H,T), vanishes exactly under ODS a.e.; (R5) averages observational contexts.

Appendix C targets (R4). For finite-first-moment adapted outcomes,

EZ∼PSES(Q,Z)−EZ∼PSES(P,Z)=12DE(P,Q),\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(Q,Z) -\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(P,Z) =\tfrac12D_{\mathrm E}(P,Q),

(R6)

where

DE(P,Q)=2E∣ZP−ZQ∣−E∣ZP−ZP′∣−E∣ZQ−ZQ′∣.D_{\mathrm E}(P,Q)=2\mathbb{E}\\|Z_P-Z_Q\\| -\mathbb{E}\\|Z_P-Z'_P\\|-\mathbb{E}\\|Z_Q-Z'_Q\\|.

Score expansion proves (R6). Strict propriety [4] gives unrestricted conditional-law optima:

Rϕ∗−RX∗=12EX,TDE(PX,TZ,PH,TZ,ϕ).R_\phi^*-R_X^* =\tfrac12\mathbb{E}_{X,T} D_{\mathrm E}(P^Z_{X,T},P^{Z,\phi}_{H,T}).

(R7)

The population score gap measures lost full-law information. The distance-induced identity DE=2MMD⁡kE2D_{\mathrm E}=2\operatorname{MMD}_{k_{\mathrm E}}^2 [5] gives its RKHS view.

Full-law risk targets ODS beyond means/balance.

Why the representation objective targets the full joint law

Consider X∈−1,1X\in\\{-1,1\\}, T⊥XT\perp X, and

Y∣X=x,T=t∼N(0,(1xρt\xρt1)),0<∣ρt∣<1.Y\mid X=x,T=t\sim\mathcal N\left(0, \begin{pmatrix}1&x\rho_t\\\x\rho_t&1\end{pmatrix}\right), \qquad 0<|\rho_t|<1.

Constant means/marginals and treatment balance survive a constant representation; opposite dependence structures merge. Energy score distinguishes them, motivating distributional training and dependence alongside mean diagnostics.

From the representation to the causal fixed point

With shared context marginal ν(c)\nu(c), Theorem 4.4 gives, at zero (R3),

(ShP)(c,z)=a(c)(ShQ)(c,z).(S_hP)(c,z)=a(c)(S_hQ)(c,z).

Full-outcome integration equates mass:

∫(ShP)(c,z)dz=Ch∫kc(c,c′)ν(dc′)=∫(ShQ)(c,z)dz.\int(S_hP)(c,z)dz =C_h\int k_c(c,c')\nu(dc') =\int(S_hQ)(c,z)dz.

Thus a(c)=1a(c)=1; injectivity identifies adapted laws. Lossless decoding, ODS, and identification yield

d_\\#Q^Z_{\phi(x),t}=\mathcal L(Y(t)\mid X=x)

a.e. on supported contexts. Frozen representation/bandwidths fix each branch's fitting operator; cross-fitting estimates geometry on another fold. Tangent coverage makes zero semi-gradient imply zero field.

Approximate recovery: for finite-first-moment adapted laws, dE(P,Q)=DE(P,Q)d_{\mathrm E}(P,Q)=\sqrt{D_{\mathrm E}(P,Q)} is MMD-equivalent. Triangle inequality gives

dE(Px,tZ,Qϕ(x),tZ)≤dE(Px,tZ,Pϕ(x),tZ,ϕ)+dE(Pϕ(x),tZ,ϕ,Qϕ(x),tZ).d_{\mathrm E}(P^Z_{x,t},Q^Z_{\phi(x),t}) \le d_{\mathrm E}(P^Z_{x,t},P^{Z,\phi}_{\phi(x),t}) +d_{\mathrm E}(P^{Z,\phi}_{\phi(x),t},Q^Z_{\phi(x),t}).

The terms are adapted-space representation/generation errors: (R7) controls mean squared representation discrepancy; the generator addresses the second.

References

[1] Anirban Chatterjee, Sayantan Choudhury, and Rohan Hore. One-shot Conditional Sampling: MMD meets Nearest Neighbors. arXiv:2509.25507, 2025; ICML, 2026.

[2] Michael Arbel, Anna Korba, Adil Salim, and Arthur Gretton. Maximum Mean Discrepancy Gradient Flow. NeurIPS, 2019.

[3] Erkan Turan and Maks Ovsjanikov. Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective. arXiv:2603.09936, 2026.

[4] Tilmann Gneiting and Adrian E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477):359–378, 2007.

[5] Dino Sejdinovic, Bharath Sriperumbudur, Arthur Gretton, and Kenji Fukumizu. Equivalence of Distance-Based and RKHS-Based Statistics in Hypothesis Testing. Annals of Statistics, 41(5):2263–2291, 2013.

[6] Valentyn Melnychuk and Stefan Feuerriegel. GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes. ICLR, 2026. arXiv:2509.22953.

Thank you for the effort during the rebuttal and the additional experiments. My concern about novelty remains, as I am not yet convinced that the drifting objective offers a substantial advance over other generative models in this setting. I will keep my original score.

Thank you for your continued engagement. Our contribution addresses a specific gap in counterfactual joint-outcome estimation: recovering the treatment-specific conditional joint law from selectively observed factual vectors while preserving outcome dependence. The causal generative baselines evaluated here—including DiffPO’s marginal-product adaptation, DIME, and PO-Flow—leave substantial dependence-recovery errors in our comparisons; the marginal-product construction additionally excludes cross-outcome dependence by design. For generic joint diffusion, flow-matching, and conditional MMD generators such as CGMMD, expressive joint generation alone does not establish how factual observations should contribute to each intervention-specific correction or how conditioning should preserve distributional effect modifiers. Without these properties, matching observational or representation-induced mixtures need not recover the requested causal law. DriftCF fills this gap through a treatment-aware generative framework: treatment kernels govern factual attraction and model repulsion, while distribution-preserving covariate geometry retains the information needed for conditional joint-law recovery. Our score-based characterization and comparison with conditional MMD explain the resulting correction, and our theory connects representation fidelity, field equilibrium, and generator stationarity to recovery of the original covariate-conditional causal target. This complements GDR-learners’ focus on nuisance robustness through orthogonal generative risks. The novelty lies in this problem-specific algorithmic construction and its causal recovery analysis, whose practical effectiveness is supported by the joint-outcome experiments. Thank you again for your careful consideration.

Reviewer_SMy5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work addresses the problem of complete multivariate outcome generation in causal inference settings. The authors introduce DriftCF for estimating conditional joint distributions of potential outcomes. This method is based on the drifting models framework, and applies its attraction–repulsion mechanism to causal inference with treatment-aware kernels, producing multivariate outcome vectors in a single forward pass and thus avoiding the iterative sampling procedures of alternative generative models such as diffusion and flow-matching. The authors develop a theoretical analysis for the conditional drifting field and its fixed points, and evaluate the method on several semi-syntehtic and real-world benchmarks.

Critical strengths

The main strengths of this paper are the relevance of the addressed problem, specially for medical settings, and some promising empirical results. Also, the presentation is overall quite good, and the inclusion of an industrial application with an online randomized experiment provides evidence on potential practical applicability.

Critical weaknesses
  • The methodological contribution seems quite incremental: the paper is basically a direct application of drifting models to causal inference. The fundamental aspects of DriftCF (the attraction–repulsion field, detached regression, and one-step generation), are inherited from the drifting models framework. The authors adapt the drifting mechanism to conditional potential-outcome distributions by using treatment-aware kernels and a learned covariate representation. This adaptation can be useful, but I am not convinced this adaptation has enough substance for ICLR.
  • The main motivation of this paper is the importance of modeling dependencies between multiple outcomes. However, any expressive generative model trained to match the full joint distribution is able to capture dependencies between outcomes, and this capability is not specific to drifting. I would like to understand if there is some particular advantage in drifting to capture these dependencies: the authors should better position this work in relation to previous paper addressing the same distributional setting. Related to that, I noticed that GDR-learners (Melnychuk and Feuerriegel, ICLR 2026) is not mentioned or discussed in the related work surprising, despite the fact that it addresses the same setting and, to my understanding, GDR-learners can capture dependencies between outcomes.
  • The experiments comparisons focus primarily on diffusion, flow matching, and other approaches relying on iterative or autoregressive sampling. Since the work presents one-step generation as a fundamental advantage, it would be interesting to see a more direct comparison against one-step conditional generators.
Questions for authors

See the weaknesses.

4 comments
AuthorsOpenReview ↗

Rebuttal [1 / 2]

We thank the reviewer for the constructive comments on novelty beyond drifting, dependence recovery, and comparisons with GDR-learners and one-step conditional generators. We hope the technical explanations and evidence below address these concerns. These concerns are addressed in the updated PDF. Revisions addressing the reviews are highlighted in blue.

1. Technical contribution beyond the drifting framework

Our treatment-specific conditional-drift formula links equilibrium to the original covariate-conditional causal law. The input eT(t)e_T(t) selects P(Y(t)∣X=x)P(Y(t)\mid X=x); kT(t,t′)k_T(t,t') selects the factual/model interactions estimating its correction. Exact arm masks handle discrete treatments; symmetric dose kernels have interior bias O(ht2)O(h_t^2) under Proposition A.2's smoothness assumptions. The formula uses attraction–repulsion, detached regression, and one-step amortization from [1].

Let wc(c′)=kx(h,h′)kT(t,t′)w_c(c')=k_x(h,h')k_T(t,t'), c=(h,t)c=(h,t), and ν\nu be the supplied context distribution. Our factual and leave-one-out model corrections have the population form (omitting finite-sample stabilization)

Vθ(c,z)=∫wc(c′)kz(z,z′)(z′−z),ν(dc′)Pc′Z(dz′)∫wc(c′)kz(z,z′),ν(dc′)Pc′Z(dz′)−∫wc(c′)kz(z,z′)(z′−z),ν(dc′)Qθ,c′Z(dz′)∫wc(c′)kz(z,z′),ν(dc′)Qθ,c′Z(dz′).\boxed{ V_\theta(c,z)= \frac{\int w_c(c')k_z(z,z')(z'-z)\\,\nu(dc')P^Z_{c'}(dz')} {\int w_c(c')k_z(z,z')\\,\nu(dc')P^Z_{c'}(dz')} - \frac{\int w_c(c')k_z(z,z')(z'-z)\\,\nu(dc')Q^Z_{\theta,c'}(dz')} {\int w_c(c')k_z(z,z')\\,\nu(dc')Q^Z_{\theta,c'}(dz')}. }

Removing the treatment kernel raises Dependence 0.07 → 0.23 (Table 3), supporting treatment-specific correction.

Let H=ϕ(X)H=\phi(X), Px,t=L(Y∣X=x,T=t)P_{x,t}=\mathcal L(Y\mid X=x,T=t), and Ph,tϕ=L(Y∣H=h,T=t)P^\phi_{h,t}=\mathcal L(Y\mid H=h,T=t). Proposition 3.1's identification assumptions give Px,t=L(Y(t)∣X=x)P_{x,t}=\mathcal L(Y(t)\mid X=x). The generator's representation-conditional law is

Ph,tϕ=E[PX,t∣H=h,T=t].P^\phi_{h,t} =\mathbb{E}[P_{X,t}\mid H=h,T=t].

(S1)

Merging covariates thus merges outcome laws: the representation must retain variance, dependence, and joint-event probabilities alongside means. Representation-induced bias is studied for treatment effects in [3]; here it determines a joint generator's equilibrium.

The exact decomposition separates representation loss from generation error. For any generator Qh,tQ_{h,t} with well-defined, finite conditional KL divergences,

EX,TKL⁡(PX,T∣QH,T)=EX,TKL⁡(PX,T∣PH,Tϕ)⏟I(Y;X∣H,T)+EH,TKL⁡(PH,Tϕ∣QH,T)⏟conditional generation error.\mathbb{E}_{X,T}\operatorname{KL}(P_{X,T}\\|Q_{H,T}) ={}\underbrace{\mathbb{E}_{X,T}\operatorname{KL}(P_{X,T}\\|P^\phi_{H,T})}_{I(Y;X\mid H,T)}\\\\ +\underbrace{\mathbb{E}_{H,T}\operatorname{KL}(P^\phi_{H,T}\\|Q_{H,T})}_{\text{conditional generation error}}.

(S2)

Proof. Split the log-density ratio as

log⁡p(Y∣X,T)q(Y∣H,T)=log⁡p(Y∣X,T)p(Y∣H,T)+log⁡p(Y∣H,T)q(Y∣H,T).\log\frac{p(Y\mid X,T)}{q(Y\mid H,T)} =\log\frac{p(Y\mid X,T)}{p(Y\mid H,T)} +\log\frac{p(Y\mid H,T)}{q(Y\mid H,T)}.

Expectations give conditional mutual information in the first term; conditioning the second on (H,T)(H,T) gives its KL term. (S2) averages over observational (X,T)(X,T).

The encoder retains conditional outcome information; conditional drift matches the retained law. ODS (Proposition A.3) zeros the first term of (S2); the conditional fixed-point analysis characterizes zero generation error.

Why the energy-score encoder serves this objective

Appendix C trains the encoder by multivariate energy score on adapted outcomes ZZ. Define

DE(P,Q)=2E∣ZP−ZQ∣−E∣ZP−ZP′∣−E∣ZQ−ZQ′∣.D_{\mathrm E}(P,Q) =2\mathbb{E}\\|Z_P-Z_Q\\| -\mathbb{E}\\|Z_P-Z'_P\\| -\mathbb{E}\\|Z_Q-Z'_Q\\|.

With finite first moments, expanding the energy score gives

EZ∼PSES(Q,Z)−EZ∼PSES(P,Z)=12DE(P,Q).\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(Q,Z) -\mathbb{E}_{Z\sim P}S_{\mathrm{ES}}(P,Z) =\tfrac12D_{\mathrm E}(P,Q).

(S3)

Let RX∗R_X^*, Rϕ∗R_\phi^* be optimal population energy-score risks over all conditional laws based on (X,T)(X,T), (H,T)(H,T), respectively. Strict propriety [4] identifies the minimizers; (S3) gives

Rϕ∗−RX∗=12EX,TDE(PX,TZ,PH,TZ,ϕ).R_\phi^*-R_X^* =\tfrac12\mathbb{E}_{X,T} D_{\mathrm E}(P^Z_{X,T},P^{Z,\phi}_{H,T}).

(S4)

The unrestricted population score gap measures full-law information lost through compression. Lossless decoding transfers adapted-law equality to original outcomes. The distance-induced RKHS identity DE=2MMD⁡kE2D_{\mathrm E}=2\operatorname{MMD}_{k_{\mathrm E}}^2 gives the equivalent embedding view [5].

Why the conditional drift equilibrium recovers the causal target

The input eT(t)e_T(t) conditions generation; kT(t,t′)k_T(t,t') selects factual/model correction particles. With the frozen representation, they define

C=(H,T),kc(c,c′)=kx(h,h′)kT(t,t′).C=(H,T),\qquad k_c(c,c')=k_x(h,h')k_T(t,t').

Write P(dc,dz)=ν(dc)Pc(dz)P(dc,dz)=\nu(dc)P_c(dz), Q(dc,dz)=ν(dc)Qc(dz)Q(dc,dz)=\nu(dc)Q_c(dz). Generating at supplied contexts ensures their shared marginal. For the fixed smoothing operator

(ShR)(c,z)=∬kc(c,c′)kz(z,z′)R(dc′,dz′),(S_hR)(c,z)=\iint k_c(c,c')k_z(z,z')R(dc',dz'),

Gaussian mean shift gives [2]

Vh(c,z)=hz2∇zlog⁡ShP(c,z)ShQ(c,z).V_h(c,z)=h_z^2\nabla_z\log\frac{S_hP(c,z)}{S_hQ(c,z)}.

(S5)

AuthorsOpenReview ↗

Rebuttal [2/ 2]

Shared-context normalization identifies the conditional law. Under Theorem 4.4's regularity, positivity, connected-domain, and injectivity assumptions, zero field gives ShP(c,z)=a(c)ShQ(c,z)S_hP(c,z)=a(c)S_hQ(c,z). Integrating over the full outcome domain with the shared context marginal yields

∫ShP(c,z),dz=Ch∫kc(c,c′)ν(dc′)=∫ShQ(c,z),dz,\int S_hP(c,z)\\,dz =C_h\int k_c(c,c')\nu(dc') =\int S_hQ(c,z)\\,dz,

where Ch=∫kz(z,z′)dzC_h=\int k_z(z,z')dz is independent of z′z'. Thus a(c)=1a(c)=1; injectivity gives P=QP=Q. ODS, decoding, and causal identification then yield, almost everywhere on supported contexts,

V_h=0 \Longrightarrow Q^Z_{\phi(x),t}=P^Z_{\phi(x),t} \Longrightarrow d_\\#Q^Z_{\phi(x),t} =\mathcal L(Y(t)\mid X=x).

(S6)

Amortization connects this result to training. For clipped field uu and generator Jacobian JJ, detached regression gives

Δz=2αηNJJ⊤u+o(α),⟨u,Δz⟩=2αηN∣J⊤u∣2+o(α).\Delta\mathbf z=\frac{2\alpha\eta}{N}JJ^\top\mathbf u+o(\alpha), \qquad \langle\mathbf u,\Delta\mathbf z\rangle =\frac{2\alpha\eta}{N}\\|J^\top\mathbf u\\|^2+o(\alpha).

Population tangent coverage makes zero semi-gradient imply zero field. Proposition A.3, Theorem 4.4, Corollary 4.6, and Proposition A.5 link representation, particle correction, generator stationarity, and the original causal estimand.

2. The mechanism for recovering outcome dependence

DriftCF corrects joint-law discrepancies even with unchanged marginals. The calculation below identifies the signal; Table 3 tests dependence recovery under matched objectives.

The outcome kernel acts on complete vectors. At a fixed context with identity adapter, consider

P=N(0,Σρ),Q=N(0,Σ−ρ),Σρ=(1ρρ1),0<∣ρ∣<1.P=\mathcal N(0,\Sigma_\rho),\quad Q=\mathcal N(0,\Sigma_{-\rho}),\quad \Sigma_\rho=\begin{pmatrix}1&\rho\\\\\rho&1\end{pmatrix}, \quad 0<|\rho|<1.

The univariate marginals coincide. Gaussian smoothing changes each covariance to Σ+hz2I\Sigma+h_z^2I, giving

Vh(z)=hz2[(Σ−ρ+hz2I)−1−(Σρ+hz2I)−1]z.V_h(z)=h_z^2\left[ (\Sigma_{-\rho}+h_z^2I)^{-1} -(\Sigma_\rho+h_z^2I)^{-1} \right]z.

(S7)

Nonzero zz gives a nonzero field correcting pure dependence error, which also changes joint-event probability:

Pr⁡P(Y1>0,Y2>0)−Pr⁡Q(Y1>0,Y2>0)=arcsin⁡ρπ.\Pr_P(Y_1>0,Y_2>0)-\Pr_Q(Y_1>0,Y_2>0) =\frac{\arcsin\rho}{\pi}.

The Gaussian orthant identity 1/4+arcsin⁡(ρ)/(2π)1/4+\arcsin(\rho)/(2\pi) proves this decision-relevant difference, invisible to marginal comparisons.

With fixed geometry, Table 3 gives Dependence 0.07 for DriftCF versus 0.19/0.21 for flow matching/conditional MMD gradient, supporting this mechanism.

3. Positioning and comparison with GDR-learners

DriftCF characterizes when representation-conditioned particle equilibrium preserves the original causal target; GDR-learners [6] use orthogonal risks for conditional potential-outcome generation. Section 5 details the distinction.

Table C1 gives DriftCF Dependence 0.07 versus 0.10/0.12 for GDR-CDMs/GDR-CNFs; the common comment contains the full comparison.

4. One-step conditional generators and computation

The normalized correction distinguishes DriftCF within one-step conditional generation.

CGMMD [7] uses a nearest-neighbor estimator of expected conditional MMD. For a fixed conditional law and Gaussian-kernel means mPm_P, mQm_Q, half-squared MMD induces ∇mP−∇mQ\nabla m_P-\nabla m_Q [8]; DriftCF uses

hz2(∇mPmP−∇mQmQ).h_z^2\left(\frac{\nabla m_P}{m_P}-\frac{\nabla m_Q}{m_Q}\right).

Separate data/model normalizations generally change direction. Table 3 isolates the objective under fixed geometry; Table C1 compares complete methods.

At NFE = 1 for both, DriftCF/CGMMD give Dependence 0.07/0.13 and median latency 2.82/3.17 ms (Tables C1–C2), supporting joint-law recovery and sampling efficiency in this configuration.

References

[1] Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative Modeling via Drifting. arXiv:2602.04770, 2026.

[2] Erkan Turan and Maks Ovsjanikov. Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective. arXiv:2603.09936, 2026.

[3] Valentyn Melnychuk, Dennis Frauen, and Stefan Feuerriegel. Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation. ICLR, 2024.

[4] Tilmann Gneiting and Adrian E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477):359–378, 2007.

[5] Dino Sejdinovic, Bharath Sriperumbudur, Arthur Gretton, and Kenji Fukumizu. Equivalence of Distance-Based and RKHS-Based Statistics in Hypothesis Testing. Annals of Statistics, 41(5):2263–2291, 2013.

[6] Valentyn Melnychuk and Stefan Feuerriegel. GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes. ICLR, 2026. arXiv:2509.22953.

[7] Anirban Chatterjee, Sayantan Choudhury, and Rohan Hore. One-shot Conditional Sampling: MMD meets Nearest Neighbors. arXiv:2509.25507, 2025; ICML, 2026.

[8] Michael Arbel, Anna Korba, Adil Salim, and Arthur Gretton. Maximum Mean Discrepancy Gradient Flow. NeurIPS, 2019.

Thanks for the careful rebuttal. The empirical additions strengthen the paper. Overall, I don't think this is a bad paper, but my novelty concern remains largely unaddressed: I still see this paper mostly as drifting models for a causal setting. Also, while I am convinced that the proposed approach can effectively capture correlations, I am not sure that drifting models have a particular advantage for that in comparison to other expressive generative models. Thus, I am inclined to maintaining my score.

Thank you for recognizing the paper’s clear presentation, strengthened experiments, and effectiveness in recovering outcome dependence. We appreciate the opportunity to clarify the remaining novelty concern.

DriftCF constructs an intervention-specific particle update for recovering the original covariate-conditional joint outcome law from selectively observed factual vectors. Its treatment-aware kernels determine which factual and model particles enter both correction terms, while its distributional covariate geometry preserves the effect modifiers needed for joint-law recovery. These algorithmic choices are accompanied by an analysis of discrete-treatment masking, continuous-dose borrowing and its smoothing bias, and the conditions connecting the resulting equilibrium and generator stationarity to the identified causal target. The contribution thus lies in jointly designing factual borrowing, interaction geometry, and conditional correction for this estimation problem. Within this construction, complete-vector corrections expose dependence discrepancies, and the learned geometry retains the covariate information needed to distinguish them. Table 3 supports these roles: removing the treatment kernel increases Dependence from 0.07 to 0.23, while replacing the correction under fixed architecture and geometry with conditional flow matching or conditional MMD gradient gives 0.19/0.21, compared with 0.07 for DriftCF.

Thanks again for your careful assessment and continued engagement; we hope this clarifies the algorithmic contribution and its connection to the causal estimation problem.

Comments on the paper

2 comments

We thank the reviewers for their constructive feedback. We have updated the PDF with the additional experiments, technical analysis, and clarified evaluation protocols. Revisions addressing the reviews are highlighted in blue.

Core contribution

DriftCF develops a treatment-aware conditional correction whose equilibrium is linked to the original covariate-conditional joint potential-outcome law. Treatment enters both the generator condition and the factual/model interaction geometry, so the correction is estimated from observations supporting the requested intervention. Distributional representation learning retains covariate information about outcome dependence, and the fixed-geometry analysis connects the resulting conditional field equilibrium to the identified causal target; tangent coverage further connects generator stationarity to that equilibrium under the stated assumptions. Detached regression, building on the drifting framework, amortizes this correction into one complete-vector forward pass. The novelty is this causal conditioning–representation–equilibrium construction, evaluated through treatment and representation controls, objective comparisons.

Additional experimental results

Table C1. Joint potential-outcome distribution recovery.* denotes an adaptation.

MethodACIC W₁ ↓MIMIC W₁ ↓IHDP Energy score ↓News Energy score ↓Dependence ↓
DIME*1.02(0.09)1.64(0.11)0.82(0.06)3.12(0.23)0.12
DiffPO marginal product*1.26(0.14)1.92(0.17)0.89(0.12)2.71(0.36)0.16
PO-Flow*0.95(0.15)1.77(0.23)0.44(0.04)2.65(0.25)0.11
Joint conditional flow matching1.45(0.17)2.08(0.22)0.59(0.13)2.77(0.08)0.19
Joint conditional diffusion2.25(0.32)2.34(0.31)0.77(0.09)2.54(0.14)0.18
CGMMD1.12(0.10)1.48(0.12)0.69(0.08)2.92(0.18)0.13
GDR-CDMs1.07(0.12)1.35(0.08)0.37(0.10)2.44(0.13)0.10
GDR-CNFs0.96(0.11)1.41(0.16)0.48(0.15)2.59(0.17)0.12
DriftCF0.87(0.12)1.29(0.14)0.29(0.05)2.36(0.18)0.07

Table C2. Benchmark dependence recovery and sampling computation

Method / sampling configurationDependence ↓NFE ↓Median latency (ms) ↓p95 latency (ms) ↓Peak memory (× DriftCF) ↓
DIME*0.12404.715.221.26
DiffPO marginal product*0.16403.736.071.42
PO-Flow*0.11323.644.891.08
Joint conditional flow matching0.19403.224.481.23
Joint conditional diffusion0.181154.815.291.18
CGMMD0.1313.174.721.12
GDR-CDMs0.10253.255.471.15
GDR-CNFs0.12102.984.941.05
DriftCF0.0712.824.631.00

Table C3. Held-out distributional diagnostics of the learned representation

X uses raw covariates; D uses the frozen representation trained with the multivariate energy-score objective; M uses the frozen representation trained with a mean-only regression objective. All probes condition on treatment and share the same architecture and parameter capacity.

DatasetProbeHeld-out ES, mean (SD) ↓ΔES vs. X95% CI for ΔESDependence, mean ↓
ACICX0.74 (0.08)0 (ref.)—0.04
ACICD0.76 (0.10)+0.02[-0.02, +0.06]0.04
ACICM0.91 (0.07)+0.17[+0.12, +0.22]0.12
IHDPX0.27 (0.03)0 (ref.)—0.07
IHDPD0.29 (0.05)+0.02[-0.01, +0.05]0.08
IHDPM0.44 (0.04)+0.17[+0.11, +0.23]0.15
NewsX2.33 (0.15)0 (ref.)—0.06
NewsD2.36 (0.18)+0.03[0.00, +0.06]0.06
NewsM2.59 (0.21)+0.26[+0.19, +0.33]0.13
MIMIC-IIIX1.12 (0.15)0 (ref.)—0.09
MIMIC-IIID1.15 (0.14)+0.03[-0.03, +0.09]0.10
MIMIC-IIIM1.31 (0.09)+0.19[+0.13, +0.25]0.17

D's observed ES increase relative to X is 0.02–0.03, compared with 0.17–0.26 for M. D achieves lower ES and dependence error than M on every benchmark. These diagnostics support the distributional objective for retaining conditional joint-outcome information.

The individual replies give focused comparisons for each concern. The complete experimental tables and protocols are included in the updated PDF. We will publicly release the code and checkpoints for the additional experiments alongside the code included in the already-submitted Supplementary Material.

We are confident that the feedback has greatly improved the quality of our paper, and we believe that the paper is a good fit for the ICLR 27 conference. Thanks again for your time and insightful comments.

AuthorsOpenReview ↗

Closing Summary and Acknowledgments

We thank the reviewers for the discussion. The scientific problem is to recover the covariate-conditional joint distribution of multiple outcomes under a specified intervention from selectively observed factual vectors. The challenge is to preserve outcome dependence while learning from observations relevant to each treatment and retaining covariate information that modifies the full distribution. DiffPO’s marginal-product adaptation excludes dependence by construction, while DIME and PO-Flow leave dependence-recovery errors in our evaluation. Expressive joint generators—including conditional diffusion, flow matching, and CGMMD—can represent dependence, but joint generation alone does not establish that conditioning and distribution matching recover the requested causal law rather than an observational or representation-induced mixture.

DriftCF addresses this gap by jointly constructing treatment-specific factual/model interactions and distribution-preserving covariate geometry, then amortizing the resulting complete-vector correction into one-pass generation. Our theoretical contribution characterizes discrete-treatment masking and continuous-dose smoothing, separates representation loss from generation error, and connects the conditional field’s equilibrium and generator stationarity to the original causal target under the stated assumptions. The score-based characterization and conditional-MMD comparison further explain the correction’s behavior. Together, these contributions provide a treatment-aware algorithmic framework and recovery analysis for joint counterfactual estimation.

We sincerely thank the reviewers for recognizing the problem's relevance and the strengthened empirical evidence, and for clarifying that novelty remains the only shared concern. The feedback has helped us better articulate the technical gap addressed by DriftCF and distinguish our algorithmic contribution from its foundations. The reviewers' questions on novelty sharpened the distinction between generative expressiveness and the construction required for intervention-specific joint-law recovery, allowing us to position our contribution more precisely. We are grateful for the opportunity to refine the paper's positioning and presentation, which strengthened both its evidence and its explanation. Thank you again for your time and constructive engagement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.