acceptodds
Under review as a conference paper at ICLR 2027

Residual Stochastic Concept Bottleneck Models

12 views

Abstract

Concept-based models (CMs) learn human-interpretable intermediate representations that enable users to understand the model’s reasoning and steer its predictions through test-time concept interventions. However, in realistic settings, their performance can be limited by concept incompleteness, where the training concepts fail to capture all task-relevant information. Adding an unconstrained residual pathway to the model’s predictive process can overcome the consequences of incompleteness. However, this introduces an inscrutable, leak-prone prediction pathway that may weaken concept interventions, as these often leave residual variables unchanged. To address this limitation, we introduce Residual Stochastic Concept Bottleneck Models (Res-SCBMs), a family of concept-based models that jointly model concept and residual variables through a multivariate Gaussian distribution. Res-SCBMs learn a covariance structure that explicitly captures dependencies between the known concepts and learnt residual variables, allowing interventions to propagate to statistically dependent residual variables through Gaussian conditioning. This extends the intervention capabilities of previous CMs without requiring concept completeness during training. Our results show that, across multiple concept-incomplete benchmarks, Res-SCBMs improve predictive performance over competing baselines while retaining effective test-time concept interventions.

Primary area:
interpretability and explainable AI
Keywords:
Explainable Artificial Intelligence, Concept-based Explainability, Concept Bottleneck Models, Interventions, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yafxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Res-SCBM, a concept bottleneck model that models concept dependencies between observed concepts and unobserved residuals via their covariance. Thereby, the method combines the direction of intervention-aware CBMs with residual CBMs. By capturing the dependencies between concepts and the residual, the model can propagate intervention on known concepts to the unobserved but potentially correlated residuals, and thus has improved intervention effectiveness in settings with incomplete data.

Critical strengths
  • The paper combines two mostly distinct directions in the area of CBMs: CBMs that handle interventions effectively and residual CBMs. How interventions can and should affect residuals is so far still underexplored in the CBM literature, and the paper takes a first step in this direction.
Critical weaknesses
  • While the paper extends the existing SCBM with a residual and combines this with an intervention-aware training objective for CEMs, this makes quite strong assumptions about the way the residual connection is modeled and how the concepts are encoded. While this enables the propagation of interventions in the case of Res-SCBM, this modeling choice is not compared against other approaches with residual connections that are potentially more flexible.
Questions for authors

Major Questions/Suggestions:

Q1: Structure of the residual:

  • As far as I understand, no part of the training particularly encourages the residual to capture different information from the concepts, is this correct? L. 254 states that the intervention-aware loss might encourage this, could the authors please explain this in more detail?

Q2: Exploration of the residual connection:

  • I think the design choice and effect of the residual connection are not sufficiently explored. There are several works that include residual connections in CBMs, e.g. [1,2], see also [3, 4] for more general overviews. These are usually introduced exactly for the incomplete concept setting that is discussed in this work, but this work does not compare itself to these methods. A discussion of how the modeling choice affects the residual connection, and an experimental evaluation, is necessary in my opinion (I understand that the setting is not directly comparable to all residual CBM methods, but at least one or two other methods should be evaluated, in combination with a broader discussion).
  • The width of the residual connection of 20 is far below the number of removed concepts in both datasets. Can the authors discuss why this number is relatively small, and also provide a bit more analysis on how the model behaves in this context? I would assume that a higher residual capacity (at least to some extent) would improve the performance of the model, but the experiment in Fig. 3 seems to indicate otherwise. In particular, in the context of these results, a more detailed analysis would be necessary (and also a comparison to how other residual methods work with different residual capacity).

Q2: Results in Tab. 1:

  • Why is ResCBM better than the incomplete linear head on AWA, but substantially worse on CUB? (or the other way round, why is the linear head in CUB so good given that it has access only to very few concepts - are the CUB concepts used in the majority-class voting fashion that is often done?)
  • On CUB incomplete, there is a substantial performance difference between Res-SCBM w/o Lint\mathcal{L}_\text{int} and Res-SCBM, where does this come from, given that there are no interventions involved yet?
  • I am a bit confused about the performance of the blackbox CBMs / baselines, in particular on CUB. Both the initial CBM paper [5] and CEM [6] report higher values for their baseline DNNs (83% and 74%, respectively). The reported DNN performance in this work of 68% is substantially lower. Where does this come from?
  • Furthermore, CEM reports a performance on CUB of about 76%, which differs substantially from the performance of 70% reported in this work. Can the authors explain these differences?

Q3: Interpretation of results of Tab. 2 (l.356)

  • Why does the high concept accuracy of Res-SCBM imply that the residual paths learn complementary information? They could just learn redundant concept information and encode the same information again (in particular in the complete tasks, where there is little other beneficial information to learn).

Q4: Intervention performance Fig.2:

  • When adding interventions on AWA, the concept ROC goes down initially after intervening up to ~20 concepts. I assume this is the case due to not properly learned covariance between concepts? This is concerning given that interventions on about 20% of the data are necessary to not make the concept accuracy worse than before, which goes a bit against the idea of interventions.
  • Also regarding the task accuracy, the interventions on AWA are less effective compared to all baselines. Can the others explain why this is the case and provide an idea of how to mitigate this?
  • How does this effect connect to the evaluations in Fig. 3? (i.e. how would the curves of Fig. 3 look for the AwA dataset?) I would assume that the effect partially comes from the residual (as the effect is less pronounced for SCBM), and seeing how the residual capacity affects this would be helpful to understand these results better.

Q5: Analysis of the residual covariance/redundancy (RQ4): While the qualitative analysis of the residual covariance in Fig. 4 is interesting, I think the evaluations in this paper should go further, given that this is the main distinguishing factor of Res-SCBM compared to other residual methods. Therefore, the following evaluations would be of great interest to this work and would shed more light on how the residual functions and how the covariance analysis can bring benefits:

  • How redundant is the information in the residuals given the concepts? Relating this to the learned covariance would be further interesting.
  • How much does the residual information align with the concepts not present in the dataset (i.e., the removed but necessary concepts)? Both of these should also contain quantitative, not just qualitative, analysis.

Minor Points:

  • As far as I understand, Interventions are performed on the logits, so that the conditional activations can be considered, and the actual values of the concepts are then sampled based on the conditional activations. Could this not in principle result in a concept being sampled as the opposite of the intervened value if the logits are the 5th or 95th percentile of its activations? If so, does this happen often in practice?
  • For the intervention-aware objective: Why are all concepts intervened on during training? I would consider interventions rather sparse in practice, and thus it would be good if the model can handle sparse interventions as well?
    • Related to that: What is ℓy\ell_y in Eq. 6?
  • Given that there are some works that started to incorporate specific structure into the concept space, for example in hierarchy [7] or object-centric [8] ways, how could Res-SCBM be adapted to such settings?
  • Why do you think CEM performs so well in incomplete settings, despite not having explicit residual connections?
  • It would also be cool to have an evaluation on some datasets with naturally incomplete concept sets, as opposed to datasets that were artificially reduced, but I understand that it is more difficult to find such datasets with sufficient concept annotations for the remaining concepts, so this is more a nice-to-have than a requirement.

[1] Shang, Chenming, et al. "Incremental residual concept bottleneck models." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. [2] Martínez-García, María, et al. "A probabilistic hard concept bottleneck for steerable generative models." International Conference on Learning Representations. Vol. 2026. 2026. [3] Zabounidis, Renos, et al. "Benchmarking and enhancing disentanglement in concept-residual models." arXiv preprint arXiv:2312.00192 (2023). [4] Knab, Patrick, et al. "What’s in the bottle? a survey and roadmap of concept bottleneck models." Transactions on Machine Learning Research (2026). [5] Koh, Pang Wei, et al. "Concept bottleneck models." International conference on machine learning. Pmlr, 2020. [6] Espinosa Zarlenga, Mateo, et al. "Concept embedding models: Beyond the accuracy-explainability trade-off." Advances in neural information processing systems 35 (2022): 21400-21413. [7] Hill, Oscar, Mateo Espinosa Zarlenga, and Mateja Jamnik. "Hierarchical concept-based interpretable models." International Conference on Learning Representations. Vol. 2026. 2026. [8] Steinmann, David, et al. "Object-centric concept-bottlenecks." Advances in Neural Information Processing Systems 38 (2026): 68899-68924.

6 comments

Thank you for your thoughtful feedback. We are glad you see value in exploring how concept interventions can affect residual variables, especially since this remains underexplored. We appreciate your suggestions and have addressed them below. Please let us know if you have any further concerns. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our replies below. For clarity, as the reviewer labels two sets of questions as Q2, we refer to the second Q2 as Q3 and increment the numbering of all subsequent questions accordingly.

[Q1] Structure of the Residual

We agree that our objective does not explicitly enforce non-redundancy. However, LintL_{\mathrm{int}} implicitly encourages residuals to capture complementary information by minimising task loss using corrected ground-truth concepts and conditioned residuals. Since the observed concept information is already provided, residuals are encouraged to capture additional predictive information rather than duplicate it (see [1] for an information-theoretic justification).

[Q2.1] Comparison with Existing Residual CBM Methods

We thank the reviewer for this suggestion. We agree that additional baselines would strengthen the experimental evaluation. To address this, we have added two baselines, namely Hard Autoregressive CBM with a side-channel [2] and IntCEM [3], whose evaluations are now included in Tables 1 and 2 (Section 5.1), Figure 2 (Section 5.2), and Figure 11 (Appendix B.7). The former is particularly relevant as it combines explicit concept dependencies with a residual pathway, allowing us to compare against an alternative architecture incorporating both components. While IntCEM does not have an explicit residual channel, its high-dimensional concept embeddings provide additional representational capacity that can implicitly encode information beyond the supervised concepts, effectively acting as a form of residual information. Our updated results show that Residual SCBM outperforms both models in task accuracy while maintaining competitive mean concept ROC-AUC. In terms of interventions, Res-SCBM remains responsive to concept corrections and achieves competitive performance with Hard AR with side-channel. IntCEM generally achieves higher task accuracy as more concepts are intervened on, potentially due to its learned intervention policy. Moreover, we have expanded our related work discussion to better position our approach relative to existing residual-based CBMs.

[Q2.2] Choice of Residual Width and Effect of Residual Capacity

To assess the effect of residual capacity, we added an ablation on CUB-Incomplete with R∈{20,50,100,500}R \in \{20,50,100,500\} (Appendix B.6). Res-SCBM remains highly responsive to concept interventions across all widths, even when residual capacity greatly exceeds concept capacity. Comparing against a variant without LintL_{\mathrm{int}} further shows that intervention-aware training becomes increasingly important as residual capacity grows.

[Q3.1] Differences Between Ground-Truth Concept Performance

The difference reflects how much predictive information the observed concepts retain. On CUB, we use majority-voted class-level annotations [4], meaning all images of a species share concept labels. The 22 retained concepts are highly informative, explaining the strong performance of a ground-truth concept classifier, whereas image-based models introduce concept prediction errors. In contrast, the 9 retained concepts in AwA2-Incomplete are less informative, allowing Res-SCBM to outperform the ground-truth concept classifier by capturing additional information through residuals.

Importantly, ground-truth concept accuracy serves as a reference for the information available in observed concepts, not an upper bound on Res-SCBM's performance.

[Q3.2] Effect of the Intervention-Aware Loss on Initial Task Accuracy

Although LintL_{\mathrm{int}} is designed to encourage responsiveness to concept interventions, it still penalises task errors under full intervention, influencing the learned representations and classifier. This additional training signal may explain the improved initial accuracy on CUB-Incomplete, similar to effects observed with IntCEM [3]. Quantifying improvements in predictive performance independently of intervenability remains future work.

[Q3.3] Differences in DNN Performance Compared to Prior Work

Our DNN follows the architecture in [5] but uses a smaller hidden dimension of C+20C+20 to match Res-SCBM's capacity. To address this difference, we added a less constrained DNN with C+200C+200 hidden dimensions (Table 1), following [5]. Its performance is closer to their reported results, suggesting that the discrepancy largely stems from architectural differences.

[Q3.4] Differences in CEM Performance Compared to Prior Work

For fairness when evaluating against SCBMs, our CEM baseline uses the same hyperparameters and training procedure as those used in Vandenhirtz et al. [7]. We have clarified this in Appendix A.3. Our CEM results on CUB are comparable to those reported in that work.

[Q4.1] Complementary vs Redundant Information in the Residual Channel

We agree that high concept accuracy alone does not establish that residuals capture complementary rather than redundant information. Our intention was to show that improved task accuracy does not compromise concept prediction quality. On concept-incomplete datasets, Res-SCBM outperforms concept-only baselines with comparable concept accuracy and even surpasses a ground-truth concept classifier on AwA2-Incomplete, suggesting that residuals capture additional task-relevant information. We have clarified this distinction in Sections 5.1 and 5.3.

[Q5.1–Q5.3] Intervention Effectiveness and Residual Capacity

The smaller improvements in task accuracy on AwA2 may partly reflect the residual pathway, which captures additional predictive information and can reduce the influence of concept corrections. To investigate this, we evaluated an uncertainty-based intervention policy that corrects the most uncertain concepts first. This improves intervention performance on both AwA2 datasets, largely eliminating the initial concept ROC-AUC drop on AwA2 and reducing it on AwA2-Incomplete. This suggests that the initial drop partly reflects random concept selection rather than poorly learned dependencies. Results are shown in Section 5.2 (Figure 2), with original random-policy results retained in Appendix B.4.

To further investigate residual capacity, we compared R=20R=20 and R=500R=500 on AwA2-Incomplete, with and without LintL_{\mathrm{int}} (Appendix B.1). Res-SCBM remains responsive even with 500 residual dimensions, whereas the variant without LintL_{\mathrm{int}} shows little improvement under interventions. This demonstrates the importance of LintL_{\mathrm{int}} for preserving intervenability as residual capacity increases.

[Q6] Analysis of Residual Covariance and Redundancy

We thank the reviewer for these suggestions. We agree that quantifying residual redundancy and alignment with omitted concepts would provide valuable insights into the learned representations.

We explored several approaches to identifying omitted concepts in the residuals, including training linear probes to predict removed CUB attributes from the learned concept and residual representations, and comparing their performance against shuffled concept labels. While these experiments suggest that residuals contain information predictive of omitted concepts, distinguishing this from information associated with class identity remains challenging.

We therefore leave a more systematic quantitative analysis of residual redundancy, alignment with omitted concepts, and their relationship to the learned covariance structure for future work. We have added this to the Limitations & Future Work section.

Minor Comments

  1. Reliability of Logit-Based Concept Interventions: The 5th/95th percentiles define intervention logits zc,iintz_{c,i}^{\mathrm{int}} for Gaussian conditioning and propagation. Intervened concepts are fixed to ground-truth values (0 or 1) before Bernoulli sampling, while only non-intervened variables are sampled. Thus, concept accuracy under full intervention is exactly 100%.

  2. Training with Full vs Sparse Concept Interventions: In LintL_{\mathrm{int}}, we intervene on all concepts during training, following [1], encouraging residuals to capture task-relevant information beyond the observed concepts. While partial-intervention training may better reflect practical settings, our results show that Res-SCBM remains responsive to sparse corrections. In Eq. 7 of the updated manuscript, ℓy\ell_y denotes the task loss, implemented as cross-entropy.

  3. Extending Res-SCBM to Structured Concept Spaces: Extending Res-SCBM to hierarchical representations such as HiCEMs [8] is challenging because our model uses scalar concepts rather than learned embeddings, and Gaussian conditioning does not enforce hierarchical consistency. Such an extension would require architectural changes, which we leave for future work.

  4. CEM Performance in Concept-Incomplete Settings: CEM [6] uses high-dimensional, sample-dependent concept embeddings that can encode task-relevant information beyond supervised concepts, effectively providing implicit residual capacity. This allows CEM to maintain strong predictive performance under concept incompleteness without an explicit residual pathway.

References

[1] Espinosa Zarlenga, "Position: In Defense of Information Leakage in Concept-Based Models." ICML, 2026. [2] Havasi et al., "Addressing Leakage in Concept Bottleneck Models." NeurIPS, 2022. [3] Espinosa Zarlenga et al., "Learning to Receive Help: Intervention-Aware Concept Embedding Models." NeurIPS, 2023. [4] Koh et al., "Concept Bottleneck Models." ICML, 2020. [5] Espinosa Zarlenga et al., "Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts." ICML, 2025. [6] Espinosa Zarlenga et al., "Concept Embedding Models: Beyond the Accuracy-Explainability Trade-Off." NeurIPS, 2022. [7] Vandenhirtz et al., "Stochastic Concept Bottleneck Models." NeurIPS, 2024. [8] Hill et al., "Hierarchical Concept-Based Interpretable Models." ICLR, 2026.

Reviewer_s9uCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper builds on SCBM by adding a residual pathway to improve its performance when concepts are incomplete. To preserve SCBM’s ability to jointly model concept relationships, it extends the covariance structure to include residual variables and their updates. The approach thus combines SCBM with mechanisms for learning and responding to concept interventions used in methods such as CEM and IntCEM.

Critical strengths

The paper provides valuable analysis of how to intervene on concept predictions when concepts are incomplete and how to extend these updates to the residual pathway. It offers a way to combine two existing lines of research.

Critical weaknesses
  1. Limited novelty and insufficient evidence for the coupling mechanism. Joint Gaussian modeling and intervention propagation through conditional distributions follow SCBM [1]. Using residual capacity to address incomplete concepts has also been studied [2], and the intervention-aware objective adapts existing training principles [3]. The main extension is to include residual variables in SCBM’s joint distribution. Moreover, the ablation removes the intervention loss while keeping the coupling between concepts and residuals. It therefore mainly shows the value of the training objective, rather than the contribution of the coupling mechanism.

  2. The comparisons do not fully support the claims. The paper compares against a DNN with a restricted feature dimension, a linear head, Hard CBM, CEM, and SCBM. For the claim of improved prediction under incomplete concepts, several relevant methods beyond CEM are missing, including probabilistic approaches such as ProbCBM [4]. The gains over Hard CBM and SCBM mainly come from access to information that these models exclude by design. For the claim that dependency modeling improves interventions, the comparisons should include ECBM, which models dependencies through a joint energy function [5], and IntCEM, which explicitly trains the model to respond to interventions [6]. The black-box comparison also restricts the DNN’s feature dimension to C + 20 by introducing an additional hidden layer. I would like to see a comparison with a standard classifier using the original backbone architecture. In particular, I find it surprising that concept-based models outperform or closely match the DNN in the complete-concept setting.

  3. The intervention setting limits what the results establish. The paper assigns the same concept template to all images within each class. The original CBM paper also used this setting, but it explicitly restricted interventions to concepts visible in each image [7]. It is unclear how visibility is handled here. For a paper focused on concept relationships, this setting alone makes it hard to assess whether the intervention gains extend to broader cases, such as concepts that are not visible or multiple concept templates within the same class. One possible explanation is that the model learns a proxy for intervention: it shifts the overall concept predictions toward a more likely class template.

Questions for authors
  1. Compare full and block-diagonal covariance models, both with and without intervention-aware training. Also test performance when residual updates are disabled during intervention.
  2. Add more closely related or recent baselines, including ProbCBM, ECBM, and IntCEM, using the same backbone, concept set, intervention budget, and tuning procedure. Please also include a standard DNN with a fixed feature dimension.
  3. Please explain the intervention setting, including how concept visibility is handled, and provide experimental evidence addressing the concerns above.

References [1] Stochastic Concept Bottleneck Models. NeurIPS, 2024. [2] Incremental Residual Concept Bottleneck Models. CVPR, 2024. [3] Position: In Defense of Information Leakage in Concept-based Models. ICML, 2026. [4] Probabilistic Concept Bottleneck Models. ICML, 2023. [5] Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations. ICLR, 2024. [6] Learning to Receive Help: Intervention-Aware Concept Embedding Models. NeurIPS, 2023. [7] Concept Bottleneck Models. ICML, 2020.

3 comments

We appreciate you taking the time to review our paper. It is encouraging that you found our analysis of interventions in concept-incomplete settings valuable. Thank you also for your helpful suggestions, which we address below. We would be happy to clarify anything further. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our replies below.

[Q1] Covariance Structure and Residual Propagation Ablations

We thank the reviewer for this suggestion. We conducted additional experiments on CUB-Incomplete to compare full and block-diagonal covariance structures, both with and without the intervention-aware loss LintL_{\mathrm{int}}. We also investigated the effect of disabling intervention propagation to the residual variables. Our results, now collected in Appendices B.2 and B.5 in our updated manuscript, suggest that:

  • Full vs block-diagonal covariance: With LintL_{\mathrm{int}}, removing the concept–residual cross-covariance Σcr\Sigma_{cr} reduces task accuracy under full intervention from 92.7% to 90.1%. Without LintL_{\mathrm{int}}, performance remains nearly unchanged.
  • Disabling residual propagation: For the full-covariance model with LintL_{\mathrm{int}}, keeping residuals fixed during interventions substantially reduces full-intervention accuracy from 92.7% to 84.3%. Without LintL_{\mathrm{int}}, the difference is negligible (86.2% vs 86.3%).

Overall, these results demonstrate that propagating concept interventions to the residual representation improves intervention performance when combined with LintL_{\mathrm{int}}. They also highlight the importance of intervention-aware training for effectively leveraging the learned concept–residual dependencies.

[W2/Q2] Additional Baselines and Standard DNN Comparison

We agree that additional comparisons with closely related methods would strengthen the evaluation. To address this concern, we have added two additional baselines, namely Hard Autoregressive CBM with side-channel [1] and IntCEMs [2], and included their results in Tables 1 and 2, with intervention comparisons in Figure 2 (Hard AR w/ side-channel) and Figure 11 (IntCEM and Hard AR w/ side-channel, Appendix B.7). Specifically, we added these baselines because of the following reasons:

  • Hard Autoregressive CBM with side-channel [1]: These architectures model concept dependencies through an autoregressive architecture while incorporating a residual channel. As such, just like our method, they can operate in incompleteness while also propagating interventions to other concepts (however this propagation is done implicitly, in contrast to ours which is explicitly captured by the covariance matrix).
  • IntCEM [2]: IntCEMs use high-dimensional concept embeddings that provide additional representational capacity beyond the observed concepts and are trained with an intervention-aware pipeline. This makes it a strong baseline for evaluating concept interventions.

Our results show that Residual SCBM outperforms both models in task accuracy while maintaining competitive mean concept ROC-AUC. In terms of interventions, Res-SCBM remains responsive to concept corrections and achieves competitive performance with Hard AR with side-channel. IntCEM generally achieves higher task accuracy as more concepts are intervened on, potentially due to its learned intervention policy. Moreover, to address concerns about the capacity of our DNN baseline, we additionally evaluated a DNN with the same architecture but a larger intermediate representation of dimension C+200C+200, where CC is the number of observed concepts. This allows us to assess the effect of increasing the DNN's capacity beyond our original configuration of C+20C+20.

We note that we do not include ECBM [3] in our updated evaluation, as its reported intervention performance on CUB and AwA2 is worse than that of CBM [7] and CEM [6], which are already included as baselines. We also exclude ProbCBM [4], as it does not allow residual information to propagate when all concepts are intervened (since interventions fix embeddings to static vectors that can’t leak information). Therefore, ProbCBM is not a strong candidate for incomplete settings, our setup of interest.

[W3/Q3] Concept Visibility and Intervention Setting

Following common practice in concept-based models, including CEM [6] and SCBM [5], our original experiments did not account for concept visibility during interventions.

To address the reviewer's concern, we have added an ablation on CUB-Incomplete following the visibility-aware intervention protocol of Koh et al. [7]. Under this protocol, intervened concepts are set to 0 when their corresponding body parts are not visible in the image, regardless of their class-level labels. Compared with our standard intervention setting, this reduces task accuracy under full intervention across all models. Nevertheless, Res-SCBM remains responsive to concept corrections, demonstrating that its intervention capabilities extend to this more restrictive setting. Further details are provided in Appendix B.3.

References

[1] Havasi et al., "Addressing Leakage in Concept Bottleneck Models." NeurIPS, 2022.

[2] Espinosa Zarlenga et al., "Learning to Receive Help: Intervention-Aware Concept Embedding Models." NeurIPS, 2023.

[3] Xu et al., "Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations." ICLR, 2024.

[4] Kim et al., "Probabilistic Concept Bottleneck Models." ICML, 2023.

[5] Vandenhirtz et al., "Stochastic Concept Bottleneck Models." NeurIPS, 2024.

[6] Espinosa Zarlenga et al., "Concept Embedding Models: Beyond the Accuracy-Explainability Trade-Off." NeurIPS, 2022.

[7] Koh et al., "Concept Bottleneck Models." ICML, 2020.

AuthorsOpenReview ↗

Follow-up before the discussion period ends

Thank you again for your feedback! With the discussion period ending in the next day, we’d appreciate hearing your thoughts on our responses and revised manuscript. We’ve responded to your concerns point by point and added experiments and clarifications. We’d be very happy to discuss any remaining questions.

Reviewer_pPvKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work extends a prior architecture called Stochastic Concept Bottleneck Model (SCBM) by adding a residual pathway to increase the model's performance in concept incomplete settings. This new architecture, called Residual SCBM (Res-SCBM), jointly models the concepts and residuals using a Gaussian distribution. This Gaussian modelling allows for concept interventions to propagate to other concepts and also the residual variables.

Critical strengths

Thank you so much for submitting this work. I appreciate the time taken to write it up and submit it to ICLR. Below are what I believe are this paper’s main strengths:

  • S1: Extending the Gaussian distribution idea followed by SCBM to a residual pathway, to handle concept incompleteness and at the same time making the residual intervenable.
  • S2: Making the connection between concepts and the residuals explicit and inspectable via their covariances, can be a useful tool for debugging the non-interpretable residual part of Res-CBM.
  • S3: The text is very well written, and easy to understand. The figures and tables are also easy to read.
Critical weaknesses

In contrast, I believe these are the main weaknesses of this work:

Missing comparisons to other side-channel methods.

  • W1: The related works section is missing a few key models. There are multiple architectures that both model concept dependencies and include residual connections just like Res-SCBM [1-4]. For example, CGM and C2BM's exogenous variables also feed directly into the task, not just the concepts, thus being a residual pathway. While CGM is mentioned, I think it would be nice to also mention in the residual capacity paragraph.

  • W2: This missingness also extends to the experimental section. Specifically, no other considered baseline includes residuals. I understand that CEM can work on incomplete settings due to leakage, but I think the bare minimum comparison would have been a hard CBM with a residual pathway (e.g. [3,4]). Also, other models (mentioned in W1) that propagate interventions, model concept dependencies, and use a residual pathway were not used as baselines.

Concept vs Residual importance to the classifier

  • W3: Furthermore, there is only a limited discussion on the importance of the concepts vs the residuals. Previous works [3-5] (and a bit less in [6]) experimentally show that the residuals can contribute more information to the final task, thus rendering interpretability meaningless, since the task is not predicted by them. Also, as shown in [4, 5], if the residual is too important, then there can be a gap when fully intervening (explaining AwA2 in Fig. 2) or the slope of the intervention curve becomes lower (see W4). Concept vs residual importance is probably also the cause of the unresponsiveness to interventions when increasing the residual dimension (Sec 5.3). Again, I feel that is an important missing evaluation and discussion.

Interventions

  • W4: I am very surprised that both Res-SCBMs and SCBM (and even hard CBM in AwA2) seem to have the same slope in the task accuracy vs interventions curve. Given that interventions in Res-SCBMs, affects way more variables than in SCBM (or hard CBM), I would expect a more rapid growth in task accuracy. Furthermore the regularized Res-SCBM variant, seems to have the exact same slope as the non-regularized variant, even though LintL_{int}, can "encourage responsiveness to concept interventions".

  • W5: I also have a conceptual problem with the interventions propagating to the side-channel. In general, when intervening on the concepts, we correct the predicted values, and in the case of propagating interventions, it is still possible to verify the correctness of the propagation. However, this is not guaranteed and cannot be verified for the residuals. For example, if a spurious correlation exists, then a performance-harming intervention may happen on the side-channel, thus deleting the effort that the user spent on intervening (and potentially leading to W4).

Minor comments

  • Lines 320-322: I noticed that in CUB, every model outperforms the DNN baseline, which is surprising. Either way, I think the authors should not use this case as an example of Res-SCBM outperforming the black box baseline.
  • I understand that the DNN is modified for each dataset to match the bottleneck size of Res-CBM (App A.3). I would argue that architecturally constraining the DNN baseline is slightly unfair. But, either way, including a small reference to this in the main text would help the reader understand the difference in performance of DNN across different variants of the datasets.
  • I think in line 332 the linear head numbers would have been nicer to be center formatted (for the 100% accuracy cases).
  • I struggle with the charactirazation that the LintL_{int} case is a different variant of the Res-SCBM model, since it is a regularizer that can be used on other CBMs proposed in Zarlenga 2026, but adapted for SCBM.

[1] Dominici, Gabriele, et al. "Causal concept graph models: Beyond causal opacity in deep learning." International conference on learning representations. Vol. 2025. 2025.

[2] De Felice, Giovanni, et al. "Causally reliable concept bottleneck models." Advances in Neural Information Processing Systems 38 (2026): 149099-149139.

[3] Havasi, Marton, Sonali Parbhoo, and Finale Doshi-Velez. "Addressing leakage in concept bottleneck models." Advances in Neural Information Processing Systems 35 (2022): 23386-23397.

[4] Kalampalikis, Nektarios, et al. "Towards Reasonable Concept Bottleneck Models." Transactions on Machine Learning Research (2026).

[5] Debot, David, and Giuseppe Marra. "Quantifying the Accuracy-Interpretability Trade-Off in Concept-Based Sidechannel Models." arXiv preprint arXiv:2510.05670 (2025).

[6] Yuksekgonul, Mert, Maggie Wang, and James Zou. "Post-hoc concept bottleneck models." arXiv preprint arXiv:2205.15480 (2022).

Questions for authors
  • From what I understand, the value of the sparsity regularizer is the same for between concept and residuals, and within concepts and within residuals. I wonder if the sparsity between should be higher for one type of relationship compared to the other type.
  • Would it be possible to provide a formula for one intervention in Eq. 4? If not, could you maybe provide an intuition-based explanation? I think it would help the clarity of that part of the manuscript.
  • I noticed in Table 1 that Res-SCBM without LintL_{int}, seems to be worse than the other variant. More specifically, it is always either as good as CEM or SCBM. Why would Res-SCBM need this regularizer so much, when the original SCBM paper did not?

My main drivers for the score is the missing baseline comparisons, metrics concerning the impact of the residual pathway, and the intervention procedure. I think the possible interpretability via the covariance might be a very strong idea, and I highly recommend the authors to further investigate in a future work. I would be willing to drastically increase my grade given my concerns are addressed.

6 comments

Thank you for your detailed review and valuable feedback. We are glad you found the paper well written and appreciated the potential of using learned covariances to understand the residual pathway better. Below, we address your concerns and questions and have updated our manuscript accordingly, showing changes in a different colour. Please let us know if anything remains unclear. Otherwise, we would sincerely appreciate it if you would consider updating your review in light of our replies below.

[W1] Related works

We agree that methods combining concept dependencies and residual capacity deserve clearer coverage. We added IntCEM [4] and Hard Autoregressive CBM with side-channel [3] to Section 2, alongside CGM [1] and C²BM [2]. Unlike these approaches, Res-SCBM jointly models concepts and residuals with an input-dependent Gaussian, propagating interventions through learned dependencies without a predefined causal graph.

[W2] Experimental baselines

To address this concern, we have added two additional baselines, namely Hard Autoregressive CBM with side-channel [3] and IntCEMs [4], and included their results in Tables 1 and 2 (Section 5.1), Figure 2 (Section 5.2), and Figure 11 (Appendix B.7). Our results show that Residual SCBM outperforms both models in task accuracy while maintaining competitive mean concept ROC-AUC. In terms of interventions, Res-SCBM remains responsive to concept corrections and achieves competitive performance with Hard AR with side-channel. IntCEM generally achieves higher task accuracy as more concepts are intervened on, potentially due to its learned intervention policy.

We acknowledge that CGM [1], C²BM [2], and CREAM [5] are relevant related works. However, we do not include them as experimental baselines because CGM and C²BM both require a causal graph constructed before model training, something we do not assume. More precisely, (1) CGM can be trained to learn the causal graph end-to-end, but the paper acknowledges that the unsupervised version does not scale in practice, and (2) C²BM incorporates external background knowledge (and foundation models) to refine its concepts and graph, something we do not assume in our setting. Similarly, CREAM explicitly encodes concept-concept and concept-task relationships specified by human experts. These approaches therefore introduce different structural assumptions or additional knowledge requirements compared with Res-SCBM, which learns dependencies directly from data without a predefined graph or external knowledge.

[W3] Concept vs Residual importance to the classifier

Our aim is to recover missing predictive information without losing meaningful concept interventions, rather than forcing predictions to rely only on observed concepts [6]. While our experiments do not directly quantify the relative importance of concepts and residuals, the intervention curves show that correcting concepts still improves task accuracy despite the residual channel (if the opposite were true, we would see flat or noisy task-accuracy-vs-intervention curves). Moreover, Section 5.3 shows that the intervention-aware loss LintL_{\mathrm{int}} preserves responsiveness even with 500 residual dimensions, counteracting the risk that the classifier ignores observed concepts. These results demonstrate that concepts remain useful for improving predictions through interventions, although they do not establish how much the classifier relies on each pathway.

[W4] Intervention Curve Slopes and the Effect of Propagation

Every model receives the same corrected ground-truth concepts. Gaussian conditioning updates dependent variables but adds no further ground-truth information; thus, more propagation need not yield steeper curves. The role of LintL_{\mathrm{int}} is clearer at (R=500): responsiveness deteriorates without it but is preserved with it (Section 5.3; Appendix B.1). The loss safeguards intervenability rather than guaranteeing steeper curves.

[W5] Importance of propagating interventions to residuals

We added a CUB-Incomplete ablation comparing interventions with and without residual propagation (Appendix B.5). For Res-SCBM, disabling residual propagation lowers full-intervention task accuracy from 92.7% to 84.3%, indicating that learned concept–residual dependencies help. LintL_{\mathrm{int}} encourages useful propagation by penalising task errors after conditioning on corrected concepts. However, spurious correlations may still cause harmful updates; we now acknowledge this in the Limitations section.

[Q1] Sparsity Regularization Across Concept and Residual Relationships

We apply the same sparsity coefficient across all off-diagonal entries of the precision matrix, similar to SCBMs, without introducing additional hyperparameters. In Appendix B.2, we present an ablation study showing that preserving concept–residual dependencies can significantly improve intervention performance. While this highlights the importance of these dependencies, exploring different sparsity regularization strengths for each block could provide further insights and remains an interesting direction for future work.

[Q2] Clarification of the Intervention Formula in Eq. 4

We expanded the explanation of Eq. 4 by providing an explicit formula for a single-concept intervention to clarify how interventions are performed and propagated through the model (Eq. 5 of the revised manuscript).

[Q3] Role and Necessity of the Intervention-Aware Loss

Unlike SCBM, Res-SCBM introduces a residual channel that allows task-relevant information to bypass the observed concepts. As argued in [6], without proper conditioning or regularisation, this creates a risk that the classifier relies on residual information rather than the observed concepts, reducing its responsiveness to concept interventions. We therefore incorporate the intervention-aware loss LintL_{\mathrm{int}} from [6] to avoid leakage from damaging our model’s intervenability. Our experiments in Section 5.3 strongly suggest this mechanism is important: when increasing the residual capacity to 500 dimensions, Res-SCBM without LintL_{\mathrm{int}} exhibits reduced intervention responsiveness, whereas including LintL_{\mathrm{int}} preserves intervenability despite the substantially larger residual channel.

Minor comments

  1. Comparison with the DNN Baseline on CUB: We agree and have removed this statement from the paper in Section 5.1.
  2. Fairness of the DNN Baseline Architecture: To consider the black-box model's capacity, we have added an additional DNN baseline with a larger intermediate representation of (C+200) dimensions, following [7]. We retain the original DNN baseline with a representation matched to the capacity of Res-SCBM to provide a controlled comparison. Both DNN variants are now reported in Table 1, and our trends still hold: Res-SCBMs can still recover the black-box accuracy even in incompleteness.
  3. Table Formatting of Linear Head Accuracy: table entries are now all centered.
  4. Characterization of LintL_\text{int}: We agree that the intervention-aware loss LintL_{\mathrm{int}} is a general regularization. However, we distinguish between Res-SCBM with and without LintL_{\mathrm{int}} to evaluate the effect of incorporating this loss into the training objective.

References

[1] Dominici et al., "Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning." ICLR, 2025. [2] De Felice et al., "Causally Reliable Concept Bottleneck Models." NeurIPS, 2026. [3] Havasi et al., "Addressing Leakage in Concept Bottleneck Models." NeurIPS, 2022. [4] Espinosa Zarlenga et al., "Learning to Receive Help: Intervention-Aware Concept Embedding Models." NeurIPS, 2023. [5] Kalampalikis et al., "Towards Reasonable Concept Bottleneck Models." TMLR, 2026. [6] Espinosa Zarlenga, "Position: In Defense of Information Leakage in Concept-Based Models." ICML, 2026. [7] Espinosa Zarlenga et al., "Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts." ICML, 2025.

Thank you for your detailed answers. However, I still have some concerns:

W1: Related Works

I have a slight issue with the last paragraph of the related works. Please avoid overclaiming that Res-SCBM is the only model that handles both directions that are mentioned there, since previous works also do [1-5]. Also be helpful to include a qualitative comparison between them (e.g., your answers about the additional knowledge needed).

W2: Experimental Baselines

While I understand that [1,2,4] need background knowledge, you could have used their provided graphs, or use structure learning. Also, you should mention CGM’s scalability as the reason for excluding it as a baseline. Lastly, I believe that CGM would work on the incomplete Awa dataset.

I forgot to mention it in my initial review, but why isn't [5] included in the experiments? It is probably the closest work to Res-SCBM.

Minor: I noticed that you removed Hard CBM from Fig. 2, and the newly added IntCEM is not included in it, even though both seem to be strong in terms of intervenability (Fig. 11).

W3: Concept vs Residual importance to the classifier

Even if the concepts are partially "ignored", interventions can improve task accuracy [3,4,6]. This is clear in Fig. 10 where the intervention-accuracy curve becomes flat as the number of residuals increases. This points again to the concept vs residual importance discussed in [3,4,6].
Note that the goal of this work is to build an interpretable model. Quantifying the importance of concepts vs residuals is mandatory, especially in Res-SCBM where interventions also affect the residual. It seems that the LintL_{int} regularizer is improving this problem, but we cannot be sure to which extend as it is not quantified.

W4 and W5: Intervention Curve Slopes and the Effect of Propagation

Fig. 9 and Fig. 6 show that under LintL_{int} propagating the interventions to the residual increases the slope of the intervention curve (while the non LintL_{int} model is completely unaffected). Isn't this completely antithetical to your answer in W4? On the other hand, if your answer to W4 is true, then why use the Gaussian conditioning in Res-SCBM instead of just SCBM with a residual?

Q1: Sparsity Regularization Across Concept and Residual Relationships

I slightly disagree with the conclusion that preserving concept–residual dependencies can "significantly" improve intervention performance, since the performance gap already exists at 0 interventions, albeit smaller.

Q3: Intervention-Aware Loss

Given the added experiments, and the already existing ones, I keep noticing that LintL_{int} is improving results:

  1. The regularized model is almost dominating the unregularized in test task and concept performance,
  2. The unregularized model is outperformed in task accuracy by the non-residual models,
  3. The regularized model is almost dominating the unregularized in both intervention metrics (task and concept performance), and the unregularized seems to be outperformed by Hard AR.
  4. Without the regularizer, the model seems to be prone to bypassing the concepts, as seen in Figures 3 and 10.

Given that most positive results for Res-SCBM happen only under the regularizer, and no other baseline is trained with the regularizer, I am starting to feel that the comparisons are unfair.

Minor: Black-box

While I understand that you followed [7] for the black-box model's architecture, why not just use the standard resnet architectures just like most previous works? I don't see this as a major negative point, but mostly a point of confusion.

[1] Dominici, Gabriele, et al. "Causal concept graph models: Beyond causal opacity in deep learning." International conference on learning representations. Vol. 2025. 2025.

[2] De Felice, Giovanni, et al. "Causally reliable concept bottleneck models." Advances in Neural Information Processing Systems 38 (2026): 149099-149139.

[3] Havasi, Marton, Sonali Parbhoo, and Finale Doshi-Velez. "Addressing leakage in concept bottleneck models." Advances in Neural Information Processing Systems 35 (2022): 23386-23397.

[4] Kalampalikis, Nektarios, et al. "Towards Reasonable Concept Bottleneck Models." Transactions on Machine Learning Research (2026).

[5] Xu, Xinyue, et al. "Energy-based concept bottleneck models: Unifying prediction, concept intervention, and probabilistic interpretations." International Conference on Learning Representations. Vol. 2024. 2024.

[6] Debot, David, and Giuseppe Marra. "Quantifying the Accuracy-Interpretability Trade-Off in Concept-Based Sidechannel Models." arXiv preprint arXiv:2510.05670 (2025).

[7] Espinosa Zarlenga et al., "Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts." ICML, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.