- 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
-
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?
-
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.
-
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
Rebuttal [2 / 2]
Ideal velocity gives a transport view. With , smooth positive densities, and vanishing boundary flux,
(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 , half-squared MMD induces [6]; our separately normalized correction is
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 be true/learned outcome laws. Suppose , , and
The reproducing property and Cauchy–Schwarz yield
For true-utility maximizer and model-utility maximizer over the same feasible set, adding/subtracting model utilities gives
(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.
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). We establish when conditioning on (ϕ(X),T) preserves it.
For H=ϕ(X), the representation-conditional law is
P(Y∣H,T)=E[P(Y∣X,T)∣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).
(Y2)
Here Px,t=L(Y∣X=x,T=t), Ph,tϕ=L(Y∣H=h,T=t); expectations use observational contexts. Inserting p(Y∣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 Z, multivariate energy score gives
EZ∼PSES(Q,Z)−EZ∼PSES(P,Z)=21DE(P,Q),
where
DE(P,Q)=2E∣ZP−ZQ∣−E∣ZP−ZP′∣−E∣ZQ−ZQ′∣.
Score expansion proves the identity. With finite first moments, strict propriety [1] yields, for optimal population risks over all conditional laws,
Rϕ∗−RX∗=21EX,TDE(PX,TZ,PH,TZ,ϕ).
(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=t with unit variances, correlation xρt, and X∈−1,1 independent of T, 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−MQkc, with kc=kxkT 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),
with c=(ϕ(x),t) and fixed kc=kxkT. Gaussian outcome mean shift gives [4]
Vh(c,z)=hz2∇zlogShQ(c,z)ShP(c,z).
(Y4)
Under Theorem 4.4, zero field implies ShP(c,z)=a(c)ShQ(c,z). The shared context marginal normalizes it:
∫ShP(c,z)dz=Ch∫kc(c,c′)ν(dc′)=∫ShQ(c,z)dz,
Integrals cover the full outcome domain; Ch is the Gaussian kernel integral. Thus a(c)=1; injectivity gives Pc=Qc 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, clipped field u, and generator Jacobian J, detached regression [3] gives
∇θL=−N2ηJ⊤u,Δz=N2αηJJ⊤u+o(α).
(Y6)
Consequently,
⟨u,Δz⟩=N2αη∣J⊤u∣2+o(α).
This is particle motion through the generator tangent operator. Population zero semi-gradient gives J∗u=0. Tangent coverage ensures u∈range(J); orthogonality then gives u=0. Radial clipping preserves zeros, so Vh=0 and (Y5) applies.
Thus stationarity identifies the causal target under tangent coverage, ODS, decoding, and identification.