acceptodds
Under review as a conference paper at ICLR 2027

From Tokens to Policy: Causal and Interpretable Heterogeneous Treatment Effects Identification

Abstract

Heterogeneous Treatment Effect (HTE) identification is essential to explain the impact of an intervention and optimize our policies accordingly. Existing approaches trade expressivity for interpretability, but, if some active heterogeneity drivers are unmeasured, methods at both ends of this spectrum allow for spurious HTE characterization with no causal reading. In this work, we focus on controlled experiments and argue that an oracle HTE causal characterization via the latent interactors is now within reach, thanks to (i) more extensive pre-treatment measurements, i.e., multi-modal and multi-view, and (ii) scalable representations with minimal human supervision. We then re-frame HTE identification as a Markov-blanket discovery problem on a sufficient and aligned pre-treatment representation, and introduce Neural EXposure Interaction Search (NEXIS), an iterative procedure with a provable selection guarantee. We deploy NEXIS on two anti-poverty programs in Africa, augmenting each with satellite imagery capturing previously unmeasured environmental effect modifiers, leading to new candidate prescriptive guidelines to optimize the programs' next iterations.

Primary area:
causal reasoning
Keywords:
Heterogeneous treatment effects, causal effect modifiers, causal representation learning, sparse autoencoders, multimodal causal inference

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_xVg4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies heterogeneous treatment effects in cases where the relevant effect modifiers are not directly observed but may be encoded in complex pre-treatment measurements such as satellite imagery. The key idea is to identify a small set of features that contains all information about treatment-effect heterogeneity. The method is validated and on two real-world anti-poverty programs.

Critical strengths
  • The problem is very relevant and the core idea is very good and ambitious.
  • The combination of learned representations with conditional effect-modifier selection is conceptually interesting.
  • The real-world applications are very strong and make the work more interesting than a purely synthetic study. They are discussed rigorously.
  • Figure 2 is useful.
  • The semi-synthetic CelebA experiment is a neat way to test if the method can recover effect modifiers.
  • Compared with approaches that treat a satellite embedding directly as an effect modifier, conditional feature selection is a more principled way to deal with correlated representation features.
Critical weaknesses

Presentation

  • The introduction does not clearly position the method relative to causal inference, causal discovery, and conditional feature selection. The paper does not explain the core idea simply enough at the start.
  • The paper moves between a general method and very specific poverty examples, which makes the intended scope unclear.
  • A running example would help. The cash-transfer example from Page 1 could be used throughout.
  • The full pipeline should be shown in one simple diagram with more detail and explanation than the given the “From Tokens to Policy” overview.
  • The main assumptions are spread across the paper. They should be summarized in one place together with their role and failure modes.

Method

  • Overall, the causal interpretation seems stronger than what can be established in practice. NEXIS selects features in a learned representation, but treating these features as the true causal effect modifiers requires strong assumptions that are hard to follow and to verify.
  • The method assumes that the pre-treatment data capture all relevant treatment-effect heterogeneity. This seems very strong in realistic policy settings.
  • The method also assumes that the learned representation exposes the relevant latent modifiers in a useful way. To me, it is unclear how realistic this is.
  • The representation step can itself create problems, for example when several concepts are mixed in one coordinate or one concept is spread across several coordinates.
  • The paper should separate the randomized and non-randomized settings more clearly.
  • It is unclear how stable the selected modifiers are across different spatial resolutions, feature definitions, or learned representations.
  • The jump from CelebA to the real-world applications is large. CelebA uses known modifiers in a controlled setting, while the policy data contain unknown, correlated, spatially aggregated, and, most importantly, potentially unmeasured modifiers.
Questions for authors
  • How general is the method beyond the satellite-imagery applications?
  • How common are randomized experiments in the policy settings NEXIS is meant to address? As many real policy interventions are not randomized.
  • What happens if intervention placement depends on geographic features (specifically if these are visible in the satellite images, such as roads or infrastructure)?
  • How should selected variables be interpreted when they are proxies for earlier socioeconomic processes rather than directly manipulable causes?
  • How sensitive are the results to the choice of spatial resolution and aggregation level?
  • The treatment may be randomized within the study, but the study population and geographic setting are not randomly selected. What does this imply for the causal interpretation?
  • Can the choice of spatial unit or resolution change the detected heterogeneity and create ecological-fallacy-type interpretations?
2 comments

We thank Reviewer @xVg4 for the detailed feedback, concerning (i) the assessment of the pipeline requirements, including the gap between CelebA and the applications and the stability of the results (shared concerns with all reviewers), (ii) presentation, and (iii) further minor questions.

i. Requirements assessment and results interpretation

As for any causal identification result, and as acknowledged in the paper's limitations, the pipeline requirements cannot be verified from data alone. NEXIS discoveries are therefore, to start with, hypotheses of causal and interpretable treatment interactions. Our belief in these requirements can still be strengthened empirically, and this revision does so in two ways that directly bridge the gap from CelebA to the applications:

  • Principal Alignment can be checked, and holds, across modalities, including satellite imagery (see new Appendix C.6). On text, image+text and satellite imagery, distinct dictionary-learning pipelines consistently satisfy Principal Alignment for well-defined annotated concepts, and the CelebA results replicate, including the power paradox of marginal testing and NEXIS mitigation. On satellite imagery, the results further replicate at 3 distinct spatial units, up to 12× smaller in area, suggesting that the requirements can be met in practice on satellite imagery.
  • The case-study results replicate across valid representations, measurements and spatial units (see new Appendices D.7 and E.8). We re-ran both case-study analyses with 9 retrained SAEs, survey-only versus satellite-only pools, and a halved spatial unit. The main results consistently replicate. In Ghana, every retrained dictionary certifies effect heterogeneity in the imagery, which the survey alone misses: the satellite measurements are necessary. Only the weakest findings vary more (3–4 of 9), ranking the hypotheses for field validation. Halving the tile shows that the two localized Ghana neurons describe a community's surroundings at the 5 km scale: the spatial unit is part of the measurement, to be chosen with domain knowledge.

ii. Presentation

  • Assumptions in one place: the new Appendix A.4 gathers them, with their role, failure modes and checks. Measurement and Representation Sufficiency are needed only for the causal reading, not for the selection guarantee.
  • Running example: the cash-transfer setting already runs through the paper: the motivation and Figure 1 (Section 1), the market/road example of the effect-modifier taxonomy (Section 2.1), and the two case studies (Section 6).
  • Scope: NEXIS is a general method, now validated on image, text, image+text and satellite imagery (Appendix C); anti-poverty programs are its motivating application.

iii. Further questions

  • Non-randomized settings: randomization is not strictly necessary. NEXIS only needs a CATE-equivalence test valid under the design, as shown for the regression discontinuity with difference-in-differences of LEAP 1000 (Appendix E.2).
  • Placement driven by visible features: this does not occur in our case studies (randomized in YOP, household poverty score in LEAP 1000). More generally, if placement depends on such features, they act as observed confounders, and the guarantee still holds as long as the CATE-equivalence test is valid under this design, e.g., by adjusting for them.
  • Mixed or split concepts: a principal proxy may mix several concepts, as long as it screens off its modifier (Principal Alignment), and when a concept is spread over two coordinates, NEXIS selects both, so the concept is still recovered (controlled violation, Appendix C.3).
  • Proxies of earlier socioeconomic processes: a selected coordinate may proxy an unmeasured factor. It then locates the heterogeneity without offering a causal handle (Appendix A.4), as we now state for the YOP language groups, which may be such proxies (Appendix D.4).
  • Ecological fallacy: the satellite features are community-level, so the paper interprets the discoveries as variation across communities, not individual exposure (Appendices D.8 and E.9).

We hope these new results address the reviewer's concerns and reinforce the assessment of the paper's contribution.

Thank you for your response and the additional effort. I will need some more time to carefully review the revisions before deciding whether to raise my score.

Reviewer_uo31OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for identifying a causal and interpretable characterization of heterogeneous treatment effects (HTE) from complex, unstructured pre-treatment measurements such as satellite imagery. The authors build on VanderWeele’s taxonomy of effect modifiers. Variables can predict treatment effect heterogeneity because they directly interact with the treatment, or because they are proxies, common-cause companions, or upstream ancestors of such direct modifiers. The paper argues that only the direct modifiers are useful for policy design.

To find them, the authors embed raw observations with a frozen foundation model and train a sparse autoencoder (SAE) on top, so that each SAE atom becomes a candidate effect modifier. The search for direct modifiers is then recast as finding the Markov blanket of the conditional average treatment effect (CATE) over this dictionary. The proposed algorithm, NEXIS, adapts classic forward–backward Markov blanket discovery using “CATE-equivalence” tests. By default, this is a t-test on a treatment and feature interaction term, conditional on the features already selected.

The key assumptions used to prove recovery are: 1) Measurement and representation sufficiency: the measurements, and the SAE dictionary built from them, contain all of the heterogeneity information, 2) Principal alignment: each latent direct modifier has a single SAE coordinate that screens it off from the rest of the dictionary, and 3) A faithfulness condition, together with valid and consistent tests.

The authors provide a semi-synthetic experiment on CelebA, and apply the method to two anti-poverty programs in Uganda and Ghana augmented with satellite imagery.

Critical strengths

The core idea of using SAE dictionaries over foundation-model embeddings as a scalable, interpretable hypothesis space for effect modifiers is a nice one. It seems to me to be a genuinely useful bridge between interpretability tooling and applied econometrics.

The authors principal alignment assumption is an elegant way of stating what we actually need from an SAE without requiring full monosemanticity. The paper is also commendably candid in its appendices about the limitations of the real-world analyses.

Critical weaknesses

I’m struggling to see how the method distinguishes direct modifiers from proxies, which is the paper’s central motivation. NEXIS finds the Markov blanket of the CATE over the dictionary, i.e. the coordinates that carry non-redundant information about the treatment effect. That is not the same as being a direct modifier. Indeed, NEXIS finds the smallest set of features that statistically explains the variation in the treatment effect, but a proxy or a correlated common cause can belong to that set whenever the true driver is missing or poorly measured. Calling the selected features “direct modifiers” therefore rests on untestable assumptions and the foundation models labels, not on anything the algorithm itself guarantees.

Consider the paper’s own road/market example. If market presence is not separately encoded in the dictionary, the road-density atom simply is the principal proxy, and NEXIS would select it. As I read it, the separation of direct modifiers from proxies comes entirely from the measurement sufficiency and principal alignment assumptions, plus the post-hoc foundation model labelling. It does not come from the algorithm itself. Is this reading correct? If so, I think the causal and “prescriptive” claims in the abstract and Table 1 should be toned down. The paper should state clearly what the algorithm guarantees versus what the assumptions supply.

The YOP results seem to illustrate this concern directly. Two of the certified modifiers for skilled employment are ethnolinguistic-group dummies (Karamojong and Pallisa). The authors themselves interpret these as proxies for post-conflict market connectivity. Can you help me understand how these can be direct modifiers in the sense of Section 2.1? If they are not, doesn’t this suggest the method is returning exactly the kind of non-direct modifiers it is designed to exclude?

A more direct test might be to add a CelebA attribute correlated with a true modifier that has its own neuron but no treatment interaction. Even better, consider the case where the direct modifier has no clean neuron but its proxy does. Would NEXIS reject the common-cause modifier? What would it return in the latter case? I would also find it useful to see a concept-level evaluation metric. Currently, selecting a second neuron that genuinely encodes “hat” counts as a false positive.

It seems to be that the baselines can be expanded. The only comparisons are marginal screening variants, which Section 3 already argues are the wrong tool. A fairer comparison would include conditional methods, for instance: 1) lasso on the treatment and feature interactions (e.g. an L1-penalized R-learner), or 2) causal forest variable importance.

Do you agree these are applicable here? If so, I’d be very interested in seeing them.

If these points could be addressed adequately, I would raise my score.

Questions for authors

The measurement and representation sufficiency assumptions seem very strong to me, and unverifiable in practice. They appear to be doing most of the work in giving the selected coordinates a causal reading. Is there any way to probe them, even partially, in a real application? For example, the authors could check whether conclusions are stable across independently trained SAEs, different foundation-model encoders, or different dictionary sizes. Such stability checks would add a lot of credibility to the real-world findings.

One thing that would improve the readability of the paper is to state early and plainly what a practitioner can and cannot conclude from a NEXIS output. That means separating which atoms carry non-redundant heterogeneity information, which the algorithm guarantees, from which atoms correspond to intervenable direct modifiers, which rests on the assumptions and on interpretation. This would help readers avoid over-reading the policy implications.

3 comments

We thank Reviewer @uo31 for the careful feedback and suggestions, concerning (i) the identification of direct modifiers and requirements assessment, and (ii) conditional baselines (both points similarly raised by Reviewer @JZPM).

i. Direct modifier identification and requirements assessment

We agree with the reviewer's reading, which matches the two statements of Theorem 4.1. Under Principal Alignment, faithfulness and valid tests, NEXIS recovers the principal proxies: the minimal set of coordinates explaining the effect heterogeneity, free of the non-principal correlates that marginal methods retain. Reading them as direct modifiers, and hence prescriptively, additionally requires Measurement and Representation Sufficiency, which, as for any causal identification result, cannot be verified from data alone: if market presence is unmeasured, road density can stand in for it. The paper already reads the YOP language groups as indexing regional factors (Section 6.1); Appendix D.4 now states that they may be proxies. NEXIS discoveries are thus educated hypotheses, hence the candidate prescriptive guidelines of the abstract. Following the reviewer's suggestions, we add new evidence that strengthens them:

  • Principal Alignment can be checked, and holds, across modalities, including satellite imagery (see new Appendix C.6). On text, image+text and satellite imagery (the latter at 3 spatial units), distinct dictionary-learning pipelines consistently satisfy Principal Alignment for well-defined annotated concepts (e.g., AUC 0.95 for inland water on satellite imagery, against ≤ 0.75 for the runner-up), and the CelebA results replicate, including the power paradox of marginal testing and NEXIS mitigation.
  • The case-study results replicate across valid representations (see new Appendices D.7 and E.8). As suggested, we re-ran both case-study analyses with 9 retrained SAEs, a halved spatial unit, and survey-only versus satellite-only pools. The main results consistently replicate. In Ghana, every retrained dictionary certifies effect heterogeneity in the imagery, which the survey alone misses. In Uganda, 8–9 of the 9 dictionaries recover the language groups and river presence, and three satellite neurons proxying the same regional heterogeneity are dropped once the language groups enter. Only the weakest findings vary more (3–4 of 9), ranking the hypotheses for field validation.
  • The suggested controlled tests are already covered by our benchmark. NEXIS rejects correlated companions without interaction (Section 3) and recovers a concept split over two coordinates as a whole, i.e., at the concept level (Appendix C.3).

Appendix A.4 and the revised Section 6 make this explicit; Appendices C.8 and D.7 turn both checks into practitioner guidelines.

ii. Conditional baselines

We agree that both apply, as replacements of the search step on the same dictionary (not of the full pipeline), and compared NEXIS with an L1-penalized R-learner [1] (R-lasso) and with causal-forest importance [2] with a permutation threshold [3] (Table R1). Both fall short of NEXIS:

  • R-lasso tunes its penalty for prediction without controlling spurious selections, even when no interaction exists.
  • The causal forest matches NEXIS at large n in the main setting, but without selection guarantees. A simple failure mode: when a concept is split over two coordinates, it misses both halves in all 50 runs at n = 2,000. It also needs an external stopping rule and is about 35× slower at n = 10,000.

Table R1: Conditional baselines on CelebA (mean and 95% CI over 50 seeds).

Methodn = 500, η = 5 (IoU)Concept splitting, n = 2,000 (IoU)No modifier, n = 2,000 (P(≥ 1 false selection))
Marginal testing (FWER)0.51 [0.45, 0.57]0.28 [0.26, 0.30]0.00 [0.00, 0.07]
R-lasso0.60 [0.50, 0.69]0.73 [0.68, 0.77]0.46 [0.33, 0.60]
Causal forest0.72 [0.64, 0.80]0.31 [0.30, 0.32]0.04 [0.01, 0.13]
NEXIS0.90 [0.84, 0.96]0.94 [0.90, 0.98]0.00 [0.00, 0.07]

Since these methods target CATE estimation, and extend to treatment-interaction identification only without guarantees and with simple failure modes, we propose not to add this comparison to the paper, and have instead clarified the distinction in the Related Work (Section 5). There we also specify that NEXIS is prescriptive because it targets this identification by design, while its causal reading relies on the assumptions.

We believe these analyses meet the conditions the reviewer set for raising the score.

[1] X. Nie and S. Wager. Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 2021.

[2] S. Athey, J. Tibshirani and S. Wager. Generalized random forests. The Annals of Statistics, 2019.

[3] P. H. Westfall and S. S. Young. Resampling-Based Multiple Testing. Wiley, 1993.

Reviewer_uo31OpenReview ↗

Reply to rebuttal

Thank you for the detailed response and the substantial additional analyses, particularly the stability checks across retrained dictionaries and the conditional baselines. I find the work interesting and believe it has real merit, so I will raise my score.

I do think the paper would be improved by carrying your concession into the main text. The abstract, Table 1 and Section 6.1 should state plainly that NEXIS guarantees recovery of the principal proxies, and that reading these as direct, intervenable modifiers rests on the sufficiency assumptions. The YOP language-group results in particular should note in the main text, not only in Appendix D.4, that these may be proxies for regional factors.

We sincerely thank reviewer @uo31 for the prompt response, and final suggestions, which we fully share. Following them, we have revised Section 6.1 to state in the main text that the YOP language groups may act directly or proxy unmeasured regional factors, with market connectivity as one hypothesis, and aligned Appendix D.4 accordingly. For the abstract and Table 1, we believe the current wording is already consistent with this reading, though we are happy to refine it further. The abstract frames the problem on a sufficient and aligned representation and claims a selection guarantee, and it presents the applications as candidate guidelines. Table 1 defines "Prescriptive" as targeting treatment-interacting factors, a property of the target rather than a guarantee, and the accompanying Related Work paragraph states that reading the output causally relies on the assumptions of Theorem 1. We will make sure this distinction stays consistent throughout the final version, together with any remaining feedback from the other reviewers.

Reviewer_JZPMOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the identification of causal and interpretable heterogeneous treatment effect (HTE) modifiers from complex pre-treatment observations. The authors formulate the problem as Markov blanket discovery over learned representations and propose Neural EXposure Interaction Search (NEXIS), which combines foundation-model features, sparse autoencoders, and forward-backward conditional effect-equivalence testing to identify a minimal set of effect modifiers. The paper provides theoretical recovery guarantees under assumptions including measurement sufficiency, representation sufficiency, principal alignment, and faithfulness. The method is evaluated through semi-synthetic experiments using CelebA, demonstrating improved recovery of relevant modifiers compared with marginal screening approaches. Two real-world case studies involving anti-poverty programs in Uganda and Ghana further illustrate its application to satellite imagery, identifying previously overlooked environmental factors associated with treatment effect heterogeneity.

Critical strengths

The paper addresses the important distinction between direct causal effect modifiers and correlated proxies, providing a principled formulation of interpretable HTE identification.

The proposed method offers theoretical recovery guarantees under explicit assumptions, supported by semi-synthetic experiments and two real-world case studies demonstrating its practical potential.

Critical weaknesses

The theoretical guarantees rely on Measurement Sufficiency, Representation Sufficiency, Principal Alignment, and faithfulness (Section 2.2 and Theorem 4.1). In particular, Principal Alignment assumes that each latent direct effect modifier has a distinct principal proxy coordinate in the learned representation. It is unclear whether these assumptions hold for the satellite-derived SAE features in the real-world applications. This substantially limits the strength of the claimed causal identification.

NEXIS largely adapts existing Markov blanket discovery techniques. Section 4 explicitly states that its forward selection follows IAMB and its backward certification uses the MMPC/HITON-PC subset rule. Although applying these techniques to conditional CATE-equivalence testing is meaningful, the paper does not sufficiently establish the methodological advantages of this adaptation over existing conditional feature selection approaches.

Sections 6.1 and 6.2 interpret satellite-derived features, such as vegetation heterogeneity, NDVI, and ephemeral waterways, as potential drivers of treatment effect heterogeneity. However, Section 6 explicitly acknowledges that these findings lack ground-truth validation and remain hypotheses for future field studies. Consequently, the experiments demonstrate the discovery of interpretable associations but provide limited evidence that the selected features are genuinely direct causal modifiers or can reliably guide policy interventions.

Questions for authors

Please see my concerns in weakness.

1 comment

We thank Reviewer @JZPM for the constructive feedback, concerning (i) the assessment of the pipeline requirements and corresponding results interpretation, and (ii) our methodological contribution and positioning.

i. Requirements assessment and results interpretation

As for any causal identification result, and as acknowledged in the paper's limitations, the pipeline requirements cannot be verified from data alone. NEXIS discoveries are therefore, to start with, hypotheses of causal and interpretable treatment interactions. The new Appendix A.4 clarifies what NEXIS guarantees under each assumption, so that practitioners can match their conclusions to the assumptions they are willing to accept. Our belief in these requirements can still be strengthened empirically, and this revision does so in two ways:

  • Principal Alignment can be checked, and holds, across modalities, including satellite imagery (see new Appendix C.6). We replicated the controlled benchmark on text, image+text and satellite imagery, the latter also at 3 distinct spatial units. Distinct dictionary-learning pipelines consistently satisfy Principal Alignment for well-defined annotated concepts (e.g., AUC 0.95 for inland water bodies on satellite imagery, against at most 0.75 for the runner-up), and NEXIS consistently reproduces the CelebA results: marginal testing falls into the power paradox, while NEXIS recovers the true modifiers.
  • The case-study results replicate across valid representations (see new Appendices D.7 and E.8). We re-ran both case-study analyses with 9 retrained SAEs, a halved spatial unit, and survey-only versus satellite-only pools. The main results consistently replicate. In Ghana, every retrained dictionary still certifies effect heterogeneity in the imagery, which the survey alone does not detect. In Uganda, 8–9 of the 9 dictionaries recover the language groups and river presence. As expected, stability tracks the strength of evidence: only the weakest findings, very localized or certified at the margin, vary more (3–4 of 9), which gives a principled ranking of the hypotheses for field validation.

We then propose both checks as new guidelines for practitioners: verify Principal Alignment whenever labels exist, and report stability across retrained dictionaries.

ii. Methodological contribution

The forward–backward steps are classical on purpose, since they come with known guarantees. Our contribution is to cast causal HTE identification as Markov-blanket discovery of an unobserved target (the CATE) over a learned dictionary, with the CATE-equivalence tests and the guarantee this requires. Existing conditional selection methods do not target treatment-interaction identification. We nevertheless compared NEXIS on the same dictionary with an L1-penalized R-learner (R-lasso) and with causal-forest importance, as suggested by Reviewer @uo31 (Table R1). Both fall short of NEXIS:

  • R-lasso tunes its penalty for prediction without controlling spurious selections, even when no interaction exists.
  • The causal forest matches NEXIS at large n in the main setting, but without selection guarantees. A simple failure mode: when a concept is split over two coordinates, it misses both halves in all 50 runs at n = 2,000. It also needs an external stopping rule and is about 35× slower at n = 10,000.

Table R1: Conditional baselines on the CelebA benchmark (mean and 95% CI over 50 seeds).

Methodn = 500, η = 5 (IoU)Concept splitting, n = 2,000 (IoU)No modifier, n = 2,000 (P(≥ 1 false selection))
Marginal testing (FWER)0.51 [0.45, 0.57]0.28 [0.26, 0.30]0.00 [0.00, 0.07]
R-lasso0.60 [0.50, 0.69]0.73 [0.68, 0.77]0.46 [0.33, 0.60]
Causal forest0.72 [0.64, 0.80]0.31 [0.30, 0.32]0.04 [0.01, 0.13]
NEXIS0.90 [0.84, 0.96]0.94 [0.90, 0.98]0.00 [0.00, 0.07]

Since these methods target CATE estimation, and extend to treatment-interaction identification without guarantees and with simple failure modes, we propose not to include this comparison in the paper, and have instead clarified the distinction in the Related Work (Section 5).

We hope these additions address the reviewer's concerns.

Comments on the paper

1 comment

We thank all reviewers for the very constructive feedback. The paper seems well received, with reviewers recognising the core idea as "very good and ambitious" (xVg4) and "a genuinely useful bridge between interpretability tooling and applied econometrics" (uo31), with a "principled formulation" (JZPM) and "very strong" real-world applications (xVg4). The main shared concern is whether the pipeline requirements hold in real-world applications, and how stable the results are. As for any causal identification result, these requirements cannot be verified from data alone, but they can be made more plausible by controlled experiments on the same data modality and by robustness checks on the case studies. The revision adds both, together with a clarified interpretation (see blue changes in the new manuscript upload).

1. Principal Alignment holds beyond face images (see new Appendix C.6)

We replicated the controlled benchmark on three new modalities: text (arXiv [1]), image+text, and satellite imagery (BigEarthNet [2]), as considered in our case studies, additionally varying the spatial unit. Distinct dictionary-learning pipelines consistently satisfy Principal Alignment for well-defined annotated concepts across domains, and the main results replicate, including the power paradox of marginal testing and our mitigation.

2. The case-study results are robust (see new Appendices D.7 and E.8)

We re-ran both case studies with 9 retrained SAEs, a halved spatial unit, and survey-only versus satellite-only pools. The main results consistently replicate. In Ghana, every retrained dictionary certifies heterogeneity in the imagery, which the survey alone misses. In Uganda, 8–9 of the 9 dictionaries recover the language groups and river presence. Only the weakest findings vary more (3–4 of 9), ranking the hypotheses for field validation. We thank Reviewer @uo31 for suggesting this analysis, which we now recommend as a practical guideline for new deployments.

3. Clarified assumption interpretation (see new Appendix A.4)

NEXIS discoveries remain hypotheses of causal and interpretable treatment interactions unless the requirements are accepted. The new Appendix A.4 clarifies what NEXIS guarantees under each assumption: Principal Alignment, faithfulness and valid tests give the selection guarantee, while Measurement and Representation Sufficiency are needed only for the causal reading.

In summary, NEXIS is now validated on four modalities with known ground truth and on two real-world programmes with robustness checks, and comes with a simple recipe for new deployments: check Principal Alignment whenever labels exist, and report stability across retrained dictionaries. Further points, including the requested conditional baselines, are answered individually.

[1] C. B. Clement et al. On the use of arXiv as a dataset. 2019.

[2] K. N. Clasen et al. reBEN: Refined BigEarthNet dataset for remote sensing image analysis. IGARSS, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.