acceptodds
Withdrawn from ICLR 2027

Orthogonality Is Not Realizability: Signed Generative Risks and Probability-Constrained Counterfactual Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Orthogonal scores remove first-order nuisance bias, but their empirical targets need not be probability laws. We characterize this gap through a failure–recovery theory for proper generative risks. We construct smooth nuisance estimates whose \(L_4\) errors vanish while their pseudo-density ratios remain negative on rare covariate regions. For categorical log risks, we give recovery conditions in minimum cell mass, observations per stratum, and nuisance-product error, with separate counterexamples isolating sampling and nuisance effects. Within a released counterfactual diffusion architecture, growing denoising-feature moments yield an exactly integrated objective that is unbounded below even with nuisance-product error \(o(n^-1/2)\). We formulate orthogonal embedding projection (OEP) as a minimum-distance probability repair of doubly robust feature estimates, preserving accuracy in the chosen distance and the nuisance-product error bound. Exact finite-feature guarantees and a Project STAR study connect the theory to computation.

Primary area:
causal reasoning
Keywords:
causal inference, counterfactual generation, orthogonal learning, probability realizability, kernel mean embeddings, doubly robust estimation, signed generative risks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_F1L6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces orthogonal embedding projections to correct counterfactual predictions as usual corrections might yield results that are no distribution. It supports its findings by theoretical statements and some numerical results.

Critical strengths

None

Critical weaknesses

While I am not too familiar with this area of research, there is a wide range of severe issue with the generalo scientific representation of the contribution:

  • Lack of introduction / related work / research questions: the paper fails to clearly state its contribution and takeaways; there is little to no connection to the related work or the general problem setting -- related work is discussed in the appendix despite the contribution not reaching the page limit. Overall the paper rather reads as a collection of different theoretical statements and some results than to follow a coherent research question.

  • Stating multiple digits of precision in the result section instead of capturing the overall tendency:

projecting [...] reduces RFF correction from 0.005434 to 0.001227 for action 0 and from 0.007752 to 0.001856 for action 1. Its held-out CRPS is 16.953 and 17.181, respectively. CDF projection gives 15.830 and 16.184

  • Supplementary / AI guided search: the provided code strongly suggests that generative AI tools were used to automatically run experiments until some numerically favourable results were attained. I don't have any issues with using AI for code generation or running and also giving ideas for experiments but that for example the following lines are included in a supplementary markdown file along with the code suggest that the generated output was, in contrast to what is claimed in the "AI Use Statement", in fact not checked again.

(from protocol.md of the suppl.) This is an exploratory extension requested by the author after the original Gaussian and STAR comparisons showed no CDF advantage over positive part. The original results remain unchanged. Every attempted setting is retained.

Questions for authors
  • What is a "stratum" / "strata"? I am not aware of this terminology in any ML context and it is never defined

  • In Theorem 2, what does "The argument applies fold by fold to a fixed-fold cross-fit." mean?

Reviewer_aASFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper unveils the important issue of learning the distributions of potential outcomes from observational data: The debiased meta-learners rely on finite-sample objectives that don’t have a minimizer and whose minimization thus diverges. To tackle this, the authors suggest a remedy called an orthogonal embedding projection (OEP). The OEP uses a distributional distance to project the improper pseudo-distribution (= the problematic part of the finite-sample objectives) onto a proper distribution (with positive measure). By doing so, the paper promises to preserve the debiasedness while improving the finite-sample performance of the debiased meta-learners. Finally, the authors apply their method to the real-world STAR project data.

Critical strengths

The paper studies an important and often overlooked issue of debiased meta-learners.

Critical weaknesses

The paper is nearly impossible to read (for humans): It seems to be low-quality LLM-generated content. Specifically, it has multiple severe issues, due to which it does not meet basic standards for a scientific publication:

  • The structure is odd (e.g., the introduction starts by introducing highly specific notation and method details w/o providing context)
  • Missing parts (no related work, no description of the setting)
  • Undefined notation (while some terms are defined, more than 50% of the notation is never introduced; almost every new equation reads disconnected from the previous)
  • Inconsistent terminology (e.g., small p vs. capital P, etc.)

All of my efforts to go deeper into the work have led to confusion and more frustration; hence, I stopped reading carefully starting from Sec. 4. Therefore, I did not fully understand the theoretical results, how the proposed method is defined, and how it helps to resolve the stated issue of the debiased learners.

Questions for authors
  • Why couldn’t regularization be used to counteract the issue of finite-sample debiased objectives? It would be a much simpler solution.
  • Why aren’t GDR-learners/DoubleGen part of the benchmarking in the experiments?
  • Is the method restricted to one-dimensional potential outcomes? Judging from Eq. 15, it requires making a grid of points (which is only feasible for low-dimensional outcomes).
Reviewer_QArsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies doubly robust estimate of conditional counterfactual distributions, where the corrected empirical target can be a signed measure rather than a probability distribution. It shows that negative masses can make generative training objectives unbounded below. The paper derives sufficient conditions for recovering positive cell masses in terms of minimum true cell mass, observations per stratum, and nuisance-estimation errors. It also analyzes orthogonal embedding projection, which projects an estimated counterfactual mean embedding onto the set of embeddings of probability distributions.

Critical strengths
  • The paper analyzes signed measures arising in estimated conditional counterfactual distributions, which is an important issue in doubly robust distributional treatment effect estimation. The analysis provides sufficient conditions, involving a product-error bound and the number of cells, under which positive masses can be recovered.
Critical weaknesses

I think paper can benefit from clearer explanations of several concepts and better positioning of their contributions relative to the literature. My specific concerns are as follows:

  • Line 43 in the introduction provides an example illustrating how a signed measure can arise in a doubly robust estimate of a counterfactual distribution. However, it is not sufficiently clear to the reader what qq and P^\hat P represent or where the doubly robust formula comes from. The example would benefit from a clearer introduction to these objects.
  • Section 2.2 overloads the notation. The previous section uses the notation pvjp_{vj} for the candidate model's probability whereas this section introduces pvjp_{vj} as the true probability Pv(Bj)P_v{(B_j)}. This makes the argument difficult to follow. Another example is that Equation 9 defines ψi\psi_i in terms of kernel mean embeddings whereas the proof of theorem 2 uses ψij\psi_{ij} in terms of probability laws.
  • The proof of Theorem 2 compresses several steps: e.g. lines 625-626 write "conditional independence, Hoeffding’s inequality, and a union bound over the K cells give equation 5." these intermediate details would be helpful to include. Furthermore, it is not explained how Equation 6 implies that maximum error is less than p∗/2p_* / 2 with probability tending to 1.
  • The paper writes "released GDR objective" which is not clear. Please specify exactly which objective and implementation this refers to.
  • Until Section 4, the paper was considering the estimate of conditional counterfactual distributions. Section 4 then switches to kernel mean embedding of this object without providing sufficient motivation for this transition.
  • The meaning of the operator ΠM\Pi_{\mathcal{M}} should be defined when it is introduced in line 215.
  • The experimental section does not provide enough detail despite having the available space within the nine-page limit. In particular, more details about the data-generating processes (DGPs) and more intuition about the results would be helpful.

Small things

  • Define GDR in the text when it first appears.
  • Define ΔJ∘\Delta_J^\circ in line 104.
Questions for authors
  • How does Equation 6 implies the maximum error is less than p∗/2p_*/2 with probability tending to one? It is not immediately clear to me from the proof in the appendix.
  • Which methodological and theoretical aspects of OEP go beyond combining the doubly robust conditional counterfactual mean embedding estimator of Anancharoenkij & Ponnoprat (2026) with projection onto the set of probability embeddings? I think the paper would benefit from clearer explanation of OEP's additional contributions

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

Related papers

Discussion (0)

Sign in to comment.