- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper targets semantic grounding in Concept Bottleneck Models (does the model look at the right image regions for making the concept predictions). The paper tries to improve this grounding by introducing architectural changes and multiple loss terms (e.g. for disentanglement), and evaluates the approach on multiple datasets.
Critical strengths
Semantic grounding is an important problem in CBMs. The assumption CBMs implicitly make (that the concepts mean to the model what they should mean to the human) is vital and CBM works should not take it for granted.
Critical weaknesses
The paper’s central grounding claims are not sufficiently supported.
- The approach uses a soft-attention mechanism for spatial pooling (Eq 3), similar to soft masking. Thus, every image region can still contribute to the concept prediction. There are no guarantees about which pixels contribute and which don't. To me, having such guarantees is very important for a paper centered on semantic grounding. Moreover, the authors discard the only loss that explicitly tries to enforce reliance on only the most important image regions (Eq 8) because it significantly hurts performance (Table 1). The final model therefore does not enforce reliance on the highlighted regions. There do appear to be existing CBM works where this reliance is strictly enforced (e.g. [1], which is also not mentioned in the related work). [1] also seems to have a similar story (regarding semantic grounding / alignment), which weakens part of the first contribution ("we identify concept grounding as a missing learning objective").
- Concept-level IoU remains low even with the proposed regularization. Figure 3 supports this concern qualitatively: the spatial evidence maps for the displayed concepts overlap significantly, and the "leg color" map attends primarily to the bird’s upper body rather than its legs.
- The statistical significance of the results is unclear. The number of seeds is not reported, and the tables don't provide standard deviations.
The paper is mathematically imprecise and insufficiently reproducible.
- The text before Eq 5 states that entropy minimization is applied only to "present concepts", while the equation does not reflect this. Similar ambiguities occur in other equations.
- The supplementary material misses important reproducibility details. It is unclear whether each baseline backbone is frozen or finetuned, how the CUB-pretrained ResNet was obtained, and what some of the hyperparameters are (e.g. batch size, number of epochs, learning rate schedules, ...).
[1] Colamonaco et al. Prototype-Grounded Concept Models for Verifiable Concept Alignment.
Questions for authors
I refer to the stated weaknesses.
5 comments
PGCM and the Positioning of Our Contribution. We thank the reviewer for highlighting PGCM. We agree that PGCM provides an important notion of grounding through prototype-based verification. However, PGCM and WSG-CBM address different questions. PGCM assumes access to externally defined regions (segmentation masks or segmentation-derived parts) and explains concept predictions through prototype matching. In contrast, WSG-CBM investigates whether concept-specific spatial evidence can be discovered directly from image-level concept supervision without segmentation masks, predefined parts, or external localization priors. Notably, for CUB, PGCM treats the entire image as a single part and therefore does not explicitly recover concept-specific regions. Our objective is complementary: learning concept-specific spatial evidence under weak supervision.
PGCM focuses on verifiable concept alignment through prototype-based explanations and externally provided regions, whereas WSG-CBM focuses on recovering concept-specific spatial evidence directly from image-level supervision. In particular, PGCM does not explicitly recover concept-specific regions on CUB, as the entire image is treated as a single part. Consequently, the two approaches emphasize different aspects of concept grounding and should not be expected to be equivalent on localization metrics. We added corresponding discussion in related work in the revised version.
To further examine the relationship between reliance guarantees and concept grounding, we evaluated PGCM under the same experimental protocol. The full results are provided in Tables 7, 9, and 10 of the Appendix. Under the evaluated setting, WSG-CBM achieves stronger concept-level localization and faithfulness metrics than PGCM: concept-level localization metrics (IoU: 10.98 vs. 7.82, Dice: 17.48 vs. 12.87, C-IoU: 12.87 vs. 9.48) and faithfulness metrics (ROAD: 20.12 vs. 34.00, AG: 22.49 vs. 18.69), while maintaining competitive predictive performance (class accuracy: 74.46 vs. 72.72; concept accuracy: 88.38 vs. 91.22).
These results suggest that reliance guarantees and concept localization quality capture related but distinct aspects of concept grounding. In particular, stronger reliance guarantees and stronger concept localization should not be assumed to be equivalent objectives under weak supervision. Instead, the comparison indicates that meaningful concept-specific spatial evidence can be recovered without explicit region constraints, while simultaneously maintaining competitive recognition performance and prediction faithfulness.
Low Concept-Level IoU and Figure 3. We thank the reviewer for raising this concern. We agree that the absolute concept-level IoU values remain low and that the leg-color example in Figure 3 is imperfect.
However, recovering fine-grained concept regions from image-level concept supervision alone is inherently challenging, particularly on CUB, where many concepts correspond to small object parts. We therefore do not claim that WSG-CBM solves concept grounding or that every concept map perfectly matches the intended semantic region.
Instead, our claim is that meaningful concept-specific spatial evidence can be learned under weak supervision and can improve grounding quality relative to existing CBMs. Importantly, WSG-CBM consistently improves or remains competitive across localization, faithfulness, and predictive performance, rather than optimizing a single localization metric in isolation. We have revised the paper accordingly. Our contribution is not that concept grounding is solved, but that concept-specific spatial evidence can be learned from image-level concept supervision and can jointly improve localization quality, prediction faithfulness, and predictive performance under weak supervision.
We agree that Figure 3 alone is insufficient to support this claim. Accordingly, we added additional qualitative results (Figure 7-12 in the Appendix), including both successful and challenging examples, to provide a more representative assessment of the learned spatial evidence.
Statistical Significance. We thank the reviewer for raising this concern. All experiments were conducted using multiple random seeds. In the revised supplementary material, we now report the number of seeds together with mean ± standard deviation for the primary metrics (Tables 7, 9, and 10 in the Appendix). The resulting statistics confirm the same performance trends and do not alter the conclusions of the paper. Due to page limitations, we present the complete multi-seed results in the supplementary material while retaining the original tables in the main paper for readability. Should the reviewer prefer, we are happy to prioritize multi-seed statistics in the main manuscript and move less essential results to the appendix.
Mathematical Clarity and Reproducibility. We thank the reviewer for identifying these issues. The regularization in Eq. (5) is applied only to present concepts, as stated in the text immediately preceding the equation. The implementation and reported results are already consistent with this formulation. We agree that this condition could be made more explicit.
We also agree that the original supplementary material omitted several implementation details important for reproducibility. In the revised version, we provide the complete training protocol, including backbone initialization procedures, whether each backbone is frozen or fine-tuned, the source of the CUB-pretrained ResNet, learning rates and schedules, batch sizes, training epochs, optimizer settings, loss coefficients, and all hyperparameters required for reproduction.
Summary. Taken together, these revisions clarify the intended problem setting, improve the precision of the technical presentation, and address the reviewer's concerns regarding reproducibility and experimental transparency. Combined with the additional statistical analysis and clarification of our grounding claims, we believe the revised manuscript more clearly positions WSG-CBM as a framework for weakly supervised concept grounding rather than one providing formal reliance guarantees.
We thank the reviewer for the detailed feedback. We agree that the original manuscript did not sufficiently distinguish weakly supervised concept grounding from reliance-guaranteed grounding. Our goal is not to provide formal guarantees, but to investigate whether concept-specific spatial evidence can be learned directly from image-level concept supervision. We revised the manuscript accordingly, including additional comparisons, statistical analyses, qualitative evaluations, and reproducibility details.
Soft Attention and the Lack of Strict Reliance Guarantees. We thank the reviewer for raising this important point. We agree that methods providing formal reliance guarantees represent an important direction for grounded CBMs. However, obtaining such guarantees is substantially more challenging in the weakly supervised setting, where only image-level concept supervision is available, and no concept-level spatial annotations are provided.
Our work therefore studies a complementary problem: can concept-specific spatial evidence be discovered directly from image-level concept labels? To this end, WSG-CBM explicitly decomposes grounding into two stages. The concept-agnostic evidence branch identifies spatial regions likely to contain object-related evidence based on the learned representation, while the concept-specific evidence branch learns which regions support each individual concept. The resulting concept predictions are obtained by aggregating these learned evidence maps, allowing concept supervision to directly influence the localization process.
Accordingly, we do not claim that WSG-CBM provides formal guarantees regarding which pixels contribute to a prediction. Rather, we formulate concept grounding as a weakly supervised spatial evidence discovery problem and investigate whether meaningful concept-specific evidence can be recovered without localization supervision. Our objective is to find a practical balance between spatial alignment (whether the learned evidence corresponds to concept-relevant regions) and prediction relevance (whether the evidence contributes to the prediction), which we evaluate through both localization and faithfulness metrics. We revised the introduction to make it clear.
Exclusion of the Erase Loss (Eq. 8). We agree that Eq. (8) is the component most directly related to enforcing prediction reliance on highlighted regions. However, our experiments consistently show that incorporating Eq. (8) degrades not only concept and class prediction accuracy, but also localization and faithfulness metrics. In our setting, stronger reliance constraints derived solely from self-predicted evidence appear to reinforce incorrect localizations rather than improve concept grounding.
We therefore view Eq. (8) as an informative analysis component, but not a beneficial training objective for weakly supervised concept grounding.
More fundamentally, reliance on highlighted regions alone does not guarantee semantic correctness. Because erased regions are derived from the model's own predictions, they may themselves correspond to spurious evidence. Consequently, increased reliance can sometimes amplify incorrect localizations rather than improve semantic alignment.
Our objective is therefore not merely to increase sensitivity to highlighted regions, but to recover concept-specific spatial evidence that better aligns with concept semantics under weak supervision. We made this design rationale explicit in the revised manuscript.