acceptodds
Withdrawn from ICLR 2027

Proximal Causal Concept Explanations with Unobserved Concepts

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Concept-based explanations can reproduce a model's predictions yet yield biased causal attributions when unobserved concepts confound the target concept and model output. We introduce Proximal Causal Concept Explanations (PACE), which adapts proximal causal learning to estimate conditional mean concept effects without reconstructing unobserved concepts. PACE assigns two proxy measurements distinct roles: one enters an outcome bridge, while the other supplies conditional moments for learning it from factual observations. Under explicit causal, proxy, and bridge assumptions, averaged bridge contrasts identify these mean effects. We develop finite-feature moment estimators with optional predictive stabilization and characterize, for fixed features, how approximation, moment conditioning, and regularization affect population effect contrasts. Structural simulations demonstrate correction of confounding bias that predictive adjustment can retain. Analytical factor examples show that regression can recover a particular reference mean without recovering the bridge, while exact bridges can retain noise in individual contrasts. On CEBaB human-edited reviews, we evaluate intervention predictions for review-rating models using lexical features and pretrained language-model representations, including large language model embeddings. These same-review views have unverified proxy validity, and the comparisons show no consistent advantage over feature-matched regression. Together, these results distinguish bridge recovery, mean-effect accuracy, and individual-intervention prediction, highlighting the importance of measurement design and evaluation targets in causal concept explanation.

Primary area:
causal reasoning
Keywords:
concept-based explanations, causal concept effects, proximal causal inference, unobserved confounding, interpretable machine learning, language-model interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_BwFUOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper focuses on the task of causal concept explanations. The goal of the task is to understand whether a model causally relies on a given human-interpretable concept. Existing methods for concept-effect estimation can be biased when unobserved concepts or factors affect both the target concept and the model prediction. To overcome the bias, the paper proposes PACE, which uses proxy measurements of the unobserved confounders to estimate a deconfounded causal effect of the target concept on the model output. PACE can recover the conditional mean concept effect without directly reconstructing the hidden concepts.

Critical strengths

The paper provides a proximal-causal formulation of omitted-concept confounding.

Critical weaknesses
  1. The writing is a major weakness. The paper lacks a clear logical flow, introduces key ideas before properly explaining them, and often leaves important logical connections implicit. The motivation and the key task are also not clearly introduced.

  2. PACE relies on strong assumptions about the chosen proxies and bridge function class. It assumes that the constructed W and V are valid proxies and that the chosen bridge function class contains the correct hh. These assumptions may be difficult to satisfy in realistic settings.

Questions for authors

None.

Reviewer_n8wFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is technically credible, clearly qualified, and supported by informative controls. Its separation of bridge recovery, mean effects, and individual prediction is valuable. However, the methodological contribution relative to existing proximal learning remains modest, and the experiments provide limited evidence of practical gains in mean-effect estimation under more general conditions or supervision-matched comparisons. In particular, the absence of a pooled baseline using the same intervention labels weakens the evidence for the proposed estimator's added value. These concerns motivate my recommendation of weak rejection, although a clearer methodological distinction and stronger supervision-matched evidence could change my assessment.

Critical strengths

S1. Precise identification and evaluation targets prevent misleading causal conclusions. The paper explicitly distinguishes the conditional mean identified by Theorem 1 from the pointwise bridge contrast evaluated against individual edits (Eqs. (1)–(2), p. 3). Its assumptions separate latent bridge existence from observable uniqueness, and different encoders are not treated as sufficient evidence of proxy validity. The identification argument in Appendix A.1 appears coherent under the stated assumptions. The analytical factor examples make the distinction consequential: a valid bridge can have zero population mean error and substantial individual error, while regression can recover a particular reference mean without recovering the bridge (pp. 20–21). This is a useful contribution to how concept explanations are evaluated. This provides a meaningful scientific contribution despite the limited predictive gains, and partially offsets my concerns about the method's incremental novelty. S2. Matched controls substantiate the mechanism and delimit its practical benefit. Reg-W shares the bridge features, and the duplication/permutation experiments test whether the second measurement supplies useful moment information rather than merely predictive capacity. At n = 5,000, structural-model MAE is 0.029/0.059 for PACE-Moment versus 0.063/0.082 for Reg-WV in the linear/cubic settings (Table 1). Fixed-configuration controls further show that low factual prediction error can coexist with biased effects (Table 6). These results provide a credible mechanism demonstration. The authors also retain high-noise failures and disclose the absence of consistent CEBaB gains. The text analysis uses matched features and review-cluster uncertainty, while the joint-selector comparison controls intervention-selection information. This makes the positive and negative findings more interpretable. The controls materially increase confidence in the reported evidence, although they do not establish a general advantage for proximal estimation.

Critical weaknesses

W1. The distinctive methodological contribution beyond existing proximal learning remains modest. Theorem 1 specializes the established outcome-bridge argument, and Eq. (7) implements a finite-feature moment criterion with predictive and ridge penalties. The fixed-feature consistency, approximation, and penalty-displacement results clarify that estimator, but largely follow from the same finite linear system (Appendices A.2–A.4). The paper correctly acknowledges its relationship to kernel proximal estimators and the effect-targeted P-learner; the concern is the magnitude of the additional contribution, rather than a missing citation or an erroneous identification claim. Feature-matched regression isolates the effect of moment fitting, but does not establish why this estimator is preferable to established proximal alternatives for the conditional mean target. Although these diagnostic insights are valuable, it remains unclear what additional estimation capability or practical guidance PACE provides beyond existing proximal methods. This uncertainty about significance is a principal reason for my weak-rejection recommendation; superior benchmark prediction is not a prerequisite for resolving it. W2. The mean-effect evidence does not yet establish benefit beyond the designed structure and added supervision. Both factor designs have constant conditional mean effects across contexts under the reference law; Shapes additionally permits exact mean recovery by regression through symmetry (pp. 20–21). The structural simulations do include heterogeneous effects, but use small spans containing the exact bridge. The real-text experiment cannot resolve this limitation because its proxy assumptions are unverified and its primary target is edit prediction. These are appropriately disclosed boundaries, but leave the practical operating regime for the identified mean effect uncertain. Table 9 raises a more specific concern about the contribution beyond intervention supervision. Mean-targeted selection uses 1,000 additional labeled reference-validation examples, whereas the direct-label baseline estimates the 18/32 context-specific means separately, despite the constant true conditional mean effects in these particular generators (p. 23). Under this known constant-effect structure, pooling the same 1,000 labels would yield theoretical repeated-sample RMSEs of approximately 0.0135/0.0150, compared with the reported unpooled errors of 0.058/0.079. These theoretical values are not experimental results or directly equivalent to the reported seed-averaged RMSE. Moreover, pooling exploits a structural property that may not be available to a general-purpose estimator. Nevertheless, a pooled-label diagnostic would help clarify how much of the reported benefit reflects proximal moment estimation rather than additional intervention supervision and cross-context information sharing.

Questions for authors

Q1. Clarification — specify the contribution beyond existing proximal methods. Which estimation capability or empirical finding most strongly distinguishes PACE from applying existing bridge or effect-targeted proximal learning to model outputs? Please relate that contribution to the kernel procedures and P-learner already discussed. A precise distinction would affect my novelty assessment in W1; an extensive new benchmark comparison is unnecessary. Q2. Focused validation — assess cross-context pooling with the same intervention labels. Could the authors report a pooled direct-label estimate using the same 1,000 reference-validation intervention labels as a diagnostic under the known constant-effect setting? Which conclusions about benefit beyond additional supervision remain? The constant conditional mean makes this control relevant to W2. The analytical calculation in the review should not substitute for the empirical comparison. Q3. Focused validation — test a varying conditional mean within an existing factor design. In one existing analytical teacher, could a single target-by-context term produce a varying conditional mean, with the same term added to the matched estimation spans? Comparing mean-effect error under this controlled change would test whether the conclusions depend on the current constant-mean symmetry. This directly affects the scope concern in W2 and keeps the validation within the existing generator framework. Q4. Clarification from existing fits — explain the role of predictive stabilization. For representative high-noise PACE-CF rows in Table 9, which predictive and ridge penalties were selected, and what moment discrepancies did those fits attain? These summaries from the existing runs would help distinguish gains associated with moment fitting from gains obtained by moving toward predictive regression, informing W1 and the practical interpretation of W2. Moment fit alone would not certify proxy validity.

Reviewer_XHmeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses causal concept explanations when relevant concepts are not observed and may confound the effect of the target concept on a model's output. It proposes PACE, which adapts proximal causal inference: a bridge function is fitted using a first proxy W of the omitted concepts, with conditional moments indexed by a second proxy V, and averaging its contrasts identifies the conditional mean concept effect. The authors evaluate it on structural simulations, factor simulations with designed proxies (dSprites, 3D Shapes), and the CEBaB text benchmark.

Critical strengths
  • The paper is well grounded in the proximal causal inference and conditional-moment estimation literature, and its assumptions are stated explicitly.
  • The problem is interesting: concepts left out of an explanation can bias causal concept effects.
Critical weaknesses
  • W1. I find the presentation poor, which makes the paper very hard to follow. Almost every sentence is heavily qualified, with a large amount of ad-hoc terminology introduced without taking the time to guide the reader through the paper. Many estimator variants are quickly introduced (PACE-Moment, PACE-Stabilized, PACE-CF, Joint-CF, Joint-CF-noPACE, Reg-W, Reg-WV full/equal, observed-only, mean/median calibration), and I find it hard to extract a clear takeaway from the results.

  • W2. The novelty seems incremental. PACE applies the standard outcome-bridge identification of proximal causal inference (which seems standard to me) to the case where the outcome is a model's prediction, as the authors themselves state (Sec. 2). Please correct me if I have missed something here.

  • W3. The related work misses important papers. The paper omits closely related prior work. Most notably, Bahadori & Heckerman (ICLR 2021), after a quick search, seems related as it seems addressing the same problem of concept-based explanations confounded by unobserved variables. I think the paper should clarify what PACE adds beyond that work. Causal concept-based models, such as Causal Concept Graph Models [2] and Causally reliable CBMs [3], are also not discussed. Although these are interpretable-by-design rather than post-hoc, they address the same interventional question in a causal setting, and would improve positioning.

  • W4. The assumptions of the method are never met in a real setting. PACE requires two proxies satisfying specific exclusion conditions (Eq. 3), which, as the authors note, two transformations of the same input do not guarantee (Sec. 4.1). Yet, if my understanding is correct (and please correct me if it is not) in the only real-data experiment, W and V are exactly that: two summaries of the same review, with W even a compression of the explained model's own input. Indeed, none of the assumptions is verified on CEBaB (Tab. 4), while in the simulations the proxies are constructed to satisfy them. The paper thus shows that the method works when valid proxies are given, but not how to obtain them in a real explanation problem, which is arguably the hard part.

  • W5. Where it is tested, PACE does not clearly improve over 'feature-matched regression', as partly acknowledged in the abstract. A clear advantage appears only partly in the structural simulations, where the bridge features contain the true bridge and only its coefficients are estimated (Sec. 5). In the more realistic settings, it seems to me the experiments do not clearly separate the two approaches.

Ref.
[1] Bahadori and Heckerman. "Debiasing concept-based explanations with causal analysis." ICLR, 2021.
[2] Dominici et al., "Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning". ICLR 2025.
[3] De Felice et al., "Causally Reliable Concept Bottleneck Models". NeurIPS 2025.

Questions for authors
  • Q1 (related to W5). Could the authors explain clearly what methodological claims the more realistic settings (dspiretes, shapes and CEBaB) are expected to support?
  • Q2 (related to W4-W5). In which practical setting would the authors expect PACE to outperform feature-matched regression?

While I remain open to a constructive discussion, and I acknowledge that I may have misunderstood substantial parts of the paper, I believe it requires substantial revision in both presentation and real impact.

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

Related papers

Discussion (0)

Sign in to comment.