acceptodds
Under review as a conference paper at ICLR 2027

Symmetry-Informed Causal Partial Identification

Abstract

Partial identification (PI) entails estimating bounds on causal effects by encoding different assumptions on data generation as a constrained optimization problem. Such bounds can suffice to inform policy decisions even if the causal effect itself is not identifiable. Often vacuous in practice, practitioners seek to exhaustively encode domain knowledge as additional constraints to make the PI bounds more informative. We introduce known *data symmetries*—invariance of the causal effect under certain data transformations—as a new source of constraints to inform PI. We operationalize this as a *shape constraint* on the causal function, and via a *change of measure* against which PI is posed using simple data pre-processing. Both approaches are shown to sharpen bounds under two canonical PI models. This is shown both theoretically for the population case, and via experiments in the finite-sample case. More broadly, our framework establishes data symmetries as a natural, underutilized source of background knowledge for robust causal inference.

Primary area:
causal reasoning
Keywords:
Causal inference, partial identification, distribution shift, invariance, data augmentation, instrumental variables, sensitivity analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DsHwOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes methods for improving partial identification intervals using known symmetries of causal response functions. It analyzes both the direct addition of symmetry constraints and the use of data augmentation to change the observational distribution, and provides conditions governing the causal interpretation of augmentation, validity, interval size, and changes in error. For approximate symmetry, it uses interval padding and intersection with the original interval, and also discusses an extension that uses the transformation itself as an instrumental variable. Experiments cover synthetic data, an optical apparatus, cigarette demand, and do-MNIST.

Critical strengths
  1. The paper systematizes additional domain knowledge for partial identification. Symmetry is used as information that restricts the set of compatible functions, rather than only as a regularizer for point estimation. Distinguishing and comparing direct constraints with data augmentation is also useful (Sections 3–4).
  2. The paper decomposes interval narrowing into several factors. It distinguishes confounding dilution, residual risk, mean shifts, and geometric changes induced by augmentation, and gives conditions for improvements in interval size and error (Propositions 1–2, Theorem 2; pp. 7–8). This goes beyond the observation that adding constraints makes a set smaller.
  3. Approximate symmetry and potential failures are considered. The paper does not assume only exact symmetry; it addresses tolerance-dependent padding and budget recalibration. The experiments also examine symmetry misspecification and changes in budgets and sample sizes (Theorems 3.A–B, Appendix F.2).
Critical weaknesses
  1. There is a condition-dependent difference between the general IV guarantee for approximate symmetry and the implementation budget. The general recalibration in Appendix C is γ~z(ϵ)=(ϵσ~+γzρ)2\widetilde{\gamma}_z(\epsilon) = \left(\frac{\epsilon}{\widetilde{\sigma}} +\sqrt{\frac{\gamma_z}{\rho}}\right)^2 whereas the implementation in Appendix E uses ϵ2/σ~2+γz/ρ\epsilon^2/\widetilde{\sigma}^2+\gamma_z/\rho, without the cross term (pp. 18 and 20). The smaller budget corresponds to an additional condition that the relevant residual components of symmetry violation and confounding are orthogonal; the paper separately explains this condition (Appendix G.3, Remark 2; p. 29). The two expressions therefore cannot be interchanged unconditionally. To assess the robustness claim, the paper needs to show why orthogonality holds or is reasonable in the experiments, and whether interval improvements persist under the conservative general budget.

  2. The intersection of valid population sets must be distinguished from the combination of finite-sample confidence intervals. The dominance statement in Remark 1 concerns valid population intervals, whereas the implementation is described as using bootstrapped 95% Imbens–Manski confidence intervals (pp. 8 and 20). Simply intersecting two intervals that each have 95% coverage does not guarantee 95% coverage for their intersection. This does not contradict the population theorem; it identifies the need for an additional adjustment when applying the result to the intervals actually reported. The explanation of the bootstrap resampling unit, which must account for dependence among augmented rows, and the handling of joint coverage for the intersection is insufficient to assess the relationship between width reduction and statistical coverage guarantees.

  3. Interval validity in the experiments depends substantially on access to ground truth and the choice of assumptions. Synthetic symmetries are constructed from the null space of the true coefficients, and the optical experiment approximates ground truth through regression using recorded confounders (p. 21). For do-MNIST, the confounding budget γ=0.062\gamma=0.062 is selected as the minimum value that covers the true function on 99.5% of a separate split (p. 22). These settings are useful as controlled demonstrations of the principle, and I am not claiming that they constitute data leakage. However, they do not establish reliability when symmetry error and budgets must be chosen without access to causal ground truth. The cigarette-demand example is also illustrative rather than a validation of validity, as the authors acknowledge. In particular, the infeasibility of direct symmetry constraints on do-MNIST is intertwined with the function class that adjusts only the final layer, making it difficult to interpret this result as evidence of the general superiority of augmentation.

Questions for authors
  1. What justifies omitting the cross term in each experiment? If the main existing experiments are recomputed using the conservative IV budget from Appendix C, how much of the validity and width improvement is retained?
  2. Does the bootstrap for the reported confidence intervals resample original observations or augmented rows? How is joint coverage adjusted when intersecting the baseline and augmented intervals?
  3. For do-MNIST, could you report both width and coverage across a range of budgets rather than only at a single budget chosen using ground truth? This would help assess how much the improvement depends on that particular calibration using ground truth.
1 comment

We thank the reviewer for their feedback, especially regarding the experimental corrections. We have provided updated results in the manuscripts and described the changes below. We would be grateful if the reviewer would consider whether these address their concerns and the overall rating. For reference; red highlights mean new additions to the manuscript, blue highlights mean re-written/edited content.

W1 + Q1. IV budget recalibration

The reviewer is correct. The submission introduced the no-cross-term budget in App. C (lines 938–940), calling orthogonality between symmetry violation and confounding "plausible" (Rem. 2, lines 1557–1565), and used it in the implementation (App. E, line 1078). This is an additional assumption that need not hold in general, however plausible, and we cannot justify it experiment by experiment. We have now removed it entirely in the revision. Rem. 2 is deleted, and every experiment is recomputed with the conservative budget γ~z(ε)=(ε/σ~+γz/ρ)2\widetilde{\gamma}_z(\varepsilon) = \big(\varepsilon/\widetilde{\sigma} + \sqrt{\gamma_z/\rho}\big)^2 of Thm. 3.B (App. B). All results and trends are retained.

W2 + Q2. Bootstrap and intersection coverage

Intersection coverage. We thank the reviewer for pointing this out; the reported 95% indeed did not apply to the intersected bounds. In the revision every individual set is built at the 97.5% level, so that the intersection of two sets, as in Rem. 1, retains ≥95%\ge 95\% asymptotic coverage by a union bound (App. D of the revision).

Bootstrap resampling. The revision no longer bootstraps intervals. We replaced the Imbens–Manski CIs with the finite-sample budget adjustment of Prop. 4 (App. D of the revision) to solve certain infeasibility artefacts at small nn in the previous version (Fig. 14, and Fig. 6 pi+iv and pi+inv).

W3 + Q3. do-MNIST budget sweep

We have provided a gammagamma sweep and aggregated results for do-MNIST in Table 4 of Ablations, with accompanying analysis.

The reviewer is right that validity in our experiments is assessed with access to ground truth. This is a limitation of methods research in observational causal inference generally: validity can only be demonstrated where the truth is known or constructed, hence the standard practice of semi-synthetic experiments, which ours follow. Budget calibration without ground truth is likewise an open problem for sensitivity analysis as a whole. Cigarette demand shows the workflow we have in mind (γ\gamma benchmarked against an observed covariate, γz\gamma_z swept up to γ\gamma; App. E.1 of the revision), and our response to Reviewer hAEC (W3) covers ε\varepsilon.

Lastly, the last-layer infeasibility in do-MNIST was not meant to claim any general superiority of DA. But rather the implementation/operational simplicity of DA. Of course this comes at the cost of DA being least sharp of the symmetry-informed PI methods (Prop. 3). We have re-written the do-MNIST passage in Experiments to avoid such a confusion, and also gave more commentary on this in Ejkh W4 .

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

The paper uses known symmetries as additional assumptions to tighten causal bounds. Basically, adding information that we often use in prediction but may not have realized could also help partial identification. It proposes both directly adding symmetry constraints and transforming the data before applying an existing solver. The idea is novel, natural, and potentially useful.

Critical strengths
  1. An interesting source of additional assumptions for tightening bounds.
  2. Two ways to use the same knowledge, especially the data-transformation approach without changing the solver.
  3. Discussion of approximate symmetry, with a procedure to widen and intersect bounds when the symmetry is not exact.
Critical weaknesses
  1. The symmetry needs to hold for the causal response, not just the observed prediction relationship. This distinction needs more discussion for practical applications.
  2. Much of the analysis relies on the additive outcome model. Unclear how far the results extend when the hidden confounder interacts with the treatment.
  3. The approximate-symmetry allowance still needs to be specified. How a practitioner would choose this value is not very clear.
  4. Data transformation does not always tighten the bounds. Would like a clearer comparison of when it helps more than directly adding the symmetry constraint, and when it does not help.
Questions for authors
  1. How can practitioners justify the causal symmetry assumption, especially outside settings with clear physical symmetries?
  2. Which guarantees remain valid beyond additive confounding?
  3. Given the same symmetry knowledge, when should we prefer data transformation over adding the constraint directly?
1 comment

We thank the reviewer for their commentary, incorporating which we believe has improved the manuscript. We provide answers to their specific questions below, along with clarification on concerns raised. We hope our response addresses these to the reviewer's satisfaction. For reference; red highlights mean new additions to the manuscript, blue highlights mean re-written/edited content.

W1, Q1. Justifying a causal symmetry

We discuss this distinction in the Limitations paragraph of §6. Specifically, that unlike a symmetry in observational E[Y∣X]\mathbb{E}[Y \mid X], which is testable/learnable, our posited symmetry must hold for h⋆h_\star, which is not just unobservable but un-identifiable. Hence, like all (partial) identification assumptions in observational causal inference (e.g., IVs, back-doors, etc), it is fundamentally untestable and has to be justified by domain knowledge. And we give examples of such symmetries outside physics (optical device): Our cigarette-demand example uses a textbook economic symmetry (demand is unchanged when all prices and income scale by a common factor), and §2.4 lists others, including exchangeable units, semantics-preserving text edits, reverse complements in genomics.

Furthermore, the entire premise of sensitivity analysis is that identificaiton assumptions in observational causal inference should not be imposed strictly precisely because they are untestable. This is the motivation for why we allow an approximate (ε,T)(\varepsilon, \mathcal{T})-invariance. See W3 for how ε\varepsilon is set.

W2, Q2. Additive outcome model

The shape constraint (§3) does not depend on the outcome model, so it is compatible with any PI model. For DA, what we prove is under the additive model of Eq. (1). Beyond it, as discussed under "Broader applicability" in §6, one can posit invariance of the response function, H⋆(x)=dH⋆(τx)H_\star(x) \overset{d}{=} H_\star(\tau x), under which we expect the soft-intervention argument (Lem. 1), and hence validity, to carry over. This would cover Balke-Pearl style IVs with a latent, non-additive UU, proxy variables, and general graphical constrants. However, our sharpening results (Lem. 2, Prop. 2, Thm. 2) are specific to the partial-r2r^2 model and would have to be rederived under specific PI models that admit non-additive UU. A formal treatment is listed as future work.

W3. Choosing ε\varepsilon

ε\varepsilon is a sensitivity parameter and can be handled like any other in PI. It is in the units of the outcome (the largest change in expected outcome that a transformation may cause), so it can be reasoned about directly. Two common approaches:

  1. Tipping point. Sweep ε\varepsilon up to the point where the bounds reach a policy-relevant value (e.g. 00), then judge whether that ε\varepsilon is plausible given domain knowledge (Cinelli & Hazlett, 2020).
  2. Benchmarking. A perhaps more promising approach borrows calibration ideas from the leaky-IV literature. Just as a plausible IV leaks less than the treatment is confounded (Nevo & Rosen, 2012), one can posit a symmetry that is violated less than the treatment is confounded. Our cigarette-demand experiment already does this for the IV budget γz\gamma_z (Fig. 6). For ε\varepsilon this makes most sense under a bounded-confounding model in the same units as ε\varepsilon, which we leave for future work.

Both are now mentioned in the Limitations paragraph of §6.

W4, Q3. DA vs. direct constraint

The reviewer is correct that DA does not always sharpen bounds, and it should not be expected to beat the shape constraint. Under the conditions of Prop. 3 the identified sets are nested, shape constraint ⊆\subseteq TT-as-IV ⊆\subseteq DA, and TT-as-IV matches the shape constraint only when ρ=1\rho = 1.

When DA helps. Prop. 2 gives a simple observable sufficient condition under which DA sharpens, namely Ω≤1\Omega \le 1. However, this is merely a sufficient condition and not necessary which is why bounds may sharpen even when Ω≥1\Omega \geq 1. But of course since there is no guarantee in this case, we resort to the intersection trick to capture sharpening gains with guaranteed no-regression.

When to prefer DA. DA is meant as a fallback or a first pass. (i) the solver is black-box or its objective is fixed; (ii) the invariance constraint is non-convex and unstable to optimize on high-dimensional data; (iii) the solver's hypothesis class has no invariant members, so the constraint is infeasible (do\rm do-MNIST); (iv) as a cheap first check of whether the bounds already settle the policy question before committing to implement the constraint. Otherwise, use the shape constraint. A short version

Reviewer_EjkhOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper focuses on partial identification of causal effect using extra "shape constraints", which are assumptions that further limit the hypothesis class through symmetries that the practitioner can provide based on their domain knowledge. Authors suggest incorporating the invariances in three ways:

  1. Additional constraints to the solver, i.e., a soft invariance regularization;
  2. Data-augmentation, i.e., generating labeled data by applying transformations that respect the symmetries to the existing data;
  3. Transformations as instruments, i.e., taking a random transformation from the equivalence class of transformations that respect the specified symmetry and passing that as an instrument to solvers that admit one.

Authors favor Data-augmentation (DA) in their analysis, motivated by its capacity for being used in black-box solvers. Nonetheless, the supporting evidence is not convincing that DA is intrinsically better or worse than all other ways one could encode the symmetries, as I will elaborate.

Critical strengths
  1. Lemma 2 is interesting, and to me, the most important contribution of the paper: By assuming a bounded strength of confounding and either realizability or a finite-dimensional linear hypothesis class, they show that the partially-identified set is precisely the covariance ball of radius σγ\sigma\sqrt{\gamma} around the observational least-squares fit hermh_{\mathrm{erm}}, i.e., the functions whose MSE for predicting YY is within a factor 1+γ1+\gamma of the best (its scalar case is the omitted-variable-bias bound of Cinelli & Hazlett, 2020).

  2. The paper reads clearly and presents its case coherently.

  3. The visualization and conceptual figures are incredibly appreciated and helpful for the reader.

Critical weaknesses
  1. The paper claims improved bounds using the shape constraints, but the sharpening results (Lemma 2, Prop 2, Thm 2) are only valid for the linear sensitivity model (partial-r2); for leaky instruments only validity is shown (Thm 3.B). The theoretical claims presented in the paper are nonetheless sound, but I believe the paper benefits from a more clear scope/contextualization.

  2. Authors formulate the problem by assuming that the practitioner has access to an encoding/parametrization of the transformations τ:X→X\tau:\mathcal{X}\to\mathcal{X} under which the true causal effect remains invariant, i.e., h⋆(x)=h⋆(τx)h^\star(x)=h^\star(\tau x). In this case, it is sensible to assume that the same practitioner can also specify a representation ϕ:X→R\phi:\mathcal{X}\to\mathcal{R} that collapses each orbit {τx0:τ∈T}⊂X\{\tau x_0 : \tau \in \mathcal{T}\}\subset\mathcal{X} to a single value r0=ϕ(x0)∈Rr_0=\phi(x_0)\in\mathcal{R}, making R=ϕ(X)R=\phi(X) a τ\tau-invariant yet maximally-informative coarsening of XX. It is unclear whether the transformation-based formulation has any advantage: As soon as such representation is specified, any black-box PI/learning method can be used with the features RR and outcome YY, and the guarantees that follow match those of DA or class-restriction (up to restating the confounding budgets relative to RR), without the drawback of white-box access to the learning algorithm.

  3. It is hard to single-out the contributions made in this paper. On one hand, the manuscript reads coherently, but on the other hand, I can not identify if DA-for-PI is the main message, or whether Lemma 2 is central. In the latter case, it is even harder to understand the scope of the results, because Lemma 2 is valid in a specific linear sensitivity model, whereas the conversation on symmetries and data-augmentation is presented far more broadly. Attribution adds to the confusion: using the random transformation as an instrument (lines 423-427) is Observation 2 of Akbar et al. (2025), and the optical-device experiment also appears there, but neither is credited as such.

  4. In the do-MNIST experiment, the invariance-restricted set is reported infeasible, which the paper takes as "making the case stronger" for DA (§F, lines 1151-1156). I must note that this is because the solver freezes all but the last layer of the ERM network, so none of the members in the restricted class is invariant to begin with; this is a property of the solver's class, not evidence against restriction. As the paper itself states, restriction of the model class or data-augmentation are both coherent solutions to partial identification with knowledge of symmetries, and I think a sanity check for this claim is having experiments that support it, not contradict it. I suspect that the choice to commit to DA in this paper was a choice made in hindsight, and I believe it hurts the rest of the contribution and confuses the reader.

Questions for authors
  1. Why do the authors promote DA as their favored choice of encoding symmetries? I even noticed that Prop 3 in appendix D of this very paper shows that explicit restriction of the function class (a.k.a., invariant learning) is the tightest of the considered strategies.

  2. What would be a one-liner message of the paper? This does not need to be a standalone specific claim, rather a general message/theme that guides a reader who is just exposed to the manuscript.

2 comments

We are very grateful for the reviewer's feedback, reflecting on which was genuinely insightful. We provide clarifications and answers to their questions below. We hope that these address the reviewer's concerns, most importantly their informativeness score. For reference; red highlights mean new additions to the manuscript, blue highlights mean re-written/edited content.

W1. Scope

The reviewer's reading is correct. Sharpening is proved under the partial-r2r^2 model (Lem. 2, Prop. 2, Thm. 2) and for incomplete IVs (Prop. 3). For leaky IVs we prove validity (Thm. 3.B), show sharpening empirically (Fig. 6), and motivate it conceptually in §4 (variance inflation). We considered spelling out this three-way hedging on page 1, but it hurt readability. We have therefore kept page 1 as is in good faith. Should the reviewer or AC prefer otherwise, we commit to removing any mention of leaky IVs from page 1 altogether in the camera-ready, which might resolve the scope concern but then at the expense of an underclaim.

W2. X~\widetilde{X} vs. φ(X)\varphi(X)

The reviewer is right that one can, in principle, compute a "maximally informative" invariant representation φ(X)\varphi(X) and feed (φ(X),Y)(\varphi(X), Y) to the solver. Under the conditions of Prop. 3 this should coincide exactly with our TT-as-IV bounds. However, it does not come for free, and DA remains the more general choice in three ways that matter for PI:

  1. Vacuity. DA applies a single transformation to each observed XX, so its ε\varepsilon-padding only has to cover the change in h⋆h_\star over one step (Thm. 3.A). φ(X)\varphi(X) collapses entire orbits, so its padding has to cover the worst-case variation of h⋆h_\star across an orbit. For example, under ±10%\pm 10\% translations in do\rm do-MNIST the former is 00. The latter is the full range of h⋆h_\star, because φ\varphi must assign the same value to xx, τx\tau x, ττx\tau\tau x, …\ldots until the digit rolls off the image; the padding then has to account for a blank image, which makes the bounds vacuous.

  2. Validity. The only way around this is to restrict orbits to supp(pX){\rm supp}(p_X). Then φ\varphi depends on pXp_X and must be learned from data, e.g. contrastively with φ^(x)≈φ^(τx)\widehat{\varphi}(x)\approx\widehat{\varphi}(\tau x). Committing to an φ^\widehat{\varphi} outside the PI solver is risky as its approximation error propagates and can compromise validity even when the downstream solver is correctly specified. DA, by contrast, only needs a sampler on T\mathcal{T}.

  3. Compatibility. Constraints that hold for h⋆h_\star in X\mathcal{X}-coordinates need not hold for gg in R\mathcal{R}-coordinates. In particular, shape constraints need not survive. For instance, under a sign-flip symmetry take h⋆(x)=∣x∣h_\star(x) = |x| and r=φ(x)=x2r=\varphi(x) = x^2, so that g(r)=rg(r) = \sqrt{r}. Then h⋆=g∘φh_\star = g \circ \varphi is convex in xx, but gg is concave in rr. Conversely, DA is solver-agnostic because all constraints stay in the original X\mathcal{X}-coordinates, and the three main families of PI constraints transfer verbatim (Lem. 1).

That being said, we are not favouring one approach over another. The main point of the paper is "symmetry-for-PI", and one can always make additional assumptions to buy sharpness. DA is simply an interesting case where the assumptions are surprisingly mild and the claim to black-box compatibility is good (via Lem. 1), which we believe is unusual in the PI literature.

W3. Contributions and attribution

The main message is the one in Q2 below: symmetry-for-PI. Lemma 2 is indeed a meaningful contribution in its own right, beyond its role as a conduit for the DA-for-PI analysis. Committing to two sensitivity models for the analysis does not conflict with the broader conceptual claims: symmetry-for-PI applies beyond those two models, by Lem. 1 (soft intervention) and the conceptual argument of variance-inflation in §4. The discussion is therefore intentionally broad, and some commitment to specific PI models is unavoidable for analysis.

On attribution, we have updated the manuscript. §4.2 now credits Akbar et al. (2025) for the motivation behind TT-as-IV. The difference, however small, is in the characterization. They describe the DA parameters as "IV-like" (T⊥ ⁣ ⁣ ⁣⊥YT \perp\!\!\!\perp Y), and use it for regularized point estimation. We show instead that TT-invariance makes TT an incomplete instrument, a notion with precedent in the IV literature, and this is what lets TT plug into incomplete / leaky-IV PI solvers and be compared with the shape constraint (Prop. 3). The optical-device experiment now also points to Akbar et al. (2025), though our goal there (PI) is orthogonal to theirs (robust prediction).

W4. do-MNIST

We agree, and have rewritten this passage. That the shape constraint is infeasible on do\rm do-MNIST is a property of the solver's hypothesis class. CopSens, like several neural PI approaches, freezes all but the last layer of the network to prevent bounds from being vacuous. We did not mean for this to be evidence against the shape constraint. The sentence that this "makes the case stronger" for DA (lines 1151–1156 of the original submission) is dropped to prevent readers from similar misunderstandings. In the revision we also clarify that unfreezing earlier layers would (should) indeed restore feasibility (§5 and App. E.1 of the revision), though at the cost of vacuity.

Regarding a sanity check for prop. 3, the other three experiments suffice. On simulation, optical device and cigarette demand the shape constraint is feasible, and gives the sharpest bounds of all symmetry-informed methods, followed by TT-as-IV and then DA (Figs. 5, 6 and the ablations in Figs. 10–15 of the revision). This is the ordering of Prop. 3, so theory and experiments agree that restriction is the tighter route whenever it can be implemented.

do\rm do-MNIST has a different job. It is the one experiment with an off-the-shelf solver that we do not modify, which is the use case of §4, and its only claim is that DA remains operationally usable where the shape constraint may not be (claim (iii) in §5).

Q1. Why DA?

We would like to clarify that DA is not favoured over any other route to symmetry-informed PI, including the shape constraint. The Limitations paragraph clearly positions it as a fallback for when shape constraint is difficult to implement, whatever the reason may be (e.g., black-box solvers). Prop. 3 is there precisely to state the trade-off between them. We suspect the impression comes from most of the analysis being about DA, but that is only because it is the less obvious of the two. We also find it interesting that it works at all: to our knowledge there is no comparable change-of-measure approach in the PI literature, except perhaps in experiment design (Maringgele, T. and Etesami, J. (2026)), where one can actually intervene on the system.

Q2. One-line message

Known symmetries sharpen causal bounds. They are ubiquitous, underused, and nearly effortless to integrate into PI pipelines via DA pre-processing.

References

  • Akbar, Kilbertus, Shen, Muandet, Dai. An analysis of causal effect estimation using outcome invariant data augmentation. NeurIPS, 2025.
  • Tobias Maringgele and Jalal Etesami. Optimal Experiments for Partial Causal Effect Identification. NeurIPS, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.