acceptodds
Under review as a conference paper at ICLR 2027

FisherCBM: Concept Bottlenecks with Closed-Form Concept Transport and Fisher-Bounded Residual

15 views

Abstract

Concept bottleneck models route a prediction through a layer of annotated concepts, but in practice the layer transmits information the concepts do not encode, and nothing determines which part of the input a concept is read from. Both problems arise because each concept is predicted from the entire input. We present FisherCBM, in which each concept is computed from an explicit set of input tokens. A concept transports the token features of the input, such as the patch tokens of a Vision Transformer, toward each of two learnable prototypes through an unbalanced optimal transport problem with a closed-form solution: the plan toward each state is the concept's saliency map and its mass is the evidence for each state. The head receives a one-bit state and a residual gated by an evidential belief and bounded by a fixed constant. The residual's sensitivity to a token is proportional to that token's share of the plan, its token Jacobian is explicit, and the resulting Fisher penalty on off-support tokens costs a forward pass. The bounded residual gives every intervention a guaranteed direction and a lower bound on its magnitude, and reduces whether the prediction follows to a sign condition on the head. On CUB and AwA2, FisherCBM matches CEM in task accuracy with one fifth of its trainable parameters and rises monotonically to 100% accuracy under correct interventions, while wrong interventions drive it to chance as they do a hard CBM; it attains the highest concept accuracy on CUB and CelebA, and on CelebA, where two label attributes are not concepts, the bounded residual passes only part of the missing information.

Primary area:
interpretability and explainable AI
Keywords:
Concept Bottleneck Model, Representation Learning, Optimal Transport

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kqHMOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FisherCBM, a concept bottleneck model that uses a variant of an optimal transport formulation to derive concept values from input tokens. This hard concept representation is combined with a per-concept residual that can encode input information beyond the binary concepts, but is limited in how much and from which parts of the input it can encode information. Together, FisherCBM is a CBM variant that preserves the intervenability of hard CBMs, but maintains competitive performance with soft and logit-based CBM variants.

Critical strengths
  • This work introduces a CBM that integrates an optimal transport formulation into the concept representation to obtain most of the advantages from hard and soft CBMs: better and clearer interventions, and good downstream performance, while also localizing concept information to specific parts of the input.
Critical weaknesses
  • The readability and clarity of the paper can be improved. The method section can be difficult to follow at times, and the description of experimental results is often written in a convoluted way with unnecessary information.
Questions for authors
  • Q1: l. 149: Why should the excess information be in general less localizable than the concept?
  • Q2: Could the approach be adapted beyond binary concepts or make use of categorical concept information in a better way? Currently, the concepts are all individually treated, but the relevant tokens for all belly-color tokens should be the same ones. And is there a way to adapt the approach to continuous concepts?
  • Q3: You introduced the parameter γ\gamma in Eq. 8, but this parameter seems to be always set to one in the evaluations. Can the author discuss this parameter and under what circumstances changing it from one would be sensible?
  • Q4: The AwA token visualizations (in C.3) show that the token representation moves towards a general class representation when there are class-level concepts. As far as I understand it, the token-level localization of the concepts only works properly if there is sufficient per-sample variance in the concept annotations. Is there a way to find out how much of this variance is needed to form proper per-concept token representations instead of only class-level token representations?
  • Q5: For the CelebA experiments, why is it desirable to recover around one-third of the performance gap to the black-box model, but not all of it? Is it clear that the other models include the two hidden attributes in their concept space, and how can FisherCBM guarantee that this information is not present in its residual?
  • Q6: l. 452: What means an area of 93.9% equal to CEMs?
  • Q7: I do not understand the insights of the intervention experiment, in particular the "wrong" interventions. While the performance should drop by making more and more concepts incorrect, the speed at which the performance drops is difficult to compare between models, as, for example, the robustness of the concept predictor also plays into this.
  • Q8: What do the "wrong" interventions on the concept groups look like? Setting them to an arbitrary different value would make more sense, but having different possibilities, while flipping all concepts in the group would result in an invalid state.

For the results section in the main text and the appendix, the quality of the text could be improved. In its current version, the text is often convoluted and repeats a lot of information that can already easily be read from the respective figures or tables. Instead, the text should focus more on the actual insights that can and should be drawn from the results.

I think there are some related works that should be considered in the context of this work:

  • In the context of concept leakage, I think the works [1, 2] should be discussed, as they both also discuss the information leakage problem from different perspectives.
  • Considering localization of concept information, it might be interesting to discuss [3] as well, which also localizes information in the input, but with a different approach.
  • This work [4] further provides a good discussion and motivation for several points of this work, so it might also be valuable to include.

Furthermore, I think it would be helpful to differentiate more clearly between the function of the per-concept residual introduced in this work and what is typically considered a residual connection in the CBM literature. While the residual in this work is an addition to the hard, 1-bit concept representation and operates on a per-concept level, it is typically considered a separate part of the CBM, e.g. a side-channel, where information is controlled to complement the concept set, more to target completeness of the concept set [5, 6].

This work would benefit from a discussion of FisherCBMs advantages and limitations. While most of the paper is used to derive the mathematical details of the method, making the advantages clear (to some extent via experimental evidence, but also with a more general discussion), and being open about potential limitations of the work would be helpful.

Minor Points:

  • l. 161: abbreviation IUCT is not introduced (is only introduced in l. 174)
  • In the final version, ensure that the reference does not break over pages (currently happens at the end of page 2)

[1] Zarlenga, Mateo Espinosa. "In Defense of Information Leakage in Concept-based Models." International Conference on Machine Learning (2026). [2] Almudévar, Antonio, José Miguel Hernández Lobato, and Alfonso Ortega. "There was never a bottleneck in concept bottleneck models." International Conference on Learning Representations. 2026. [3] Jeon, Sujin, et al. "Locality-aware concept bottleneck model." arXiv preprint arXiv:2508.14562 (2025). [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] Shang, Chenming, et al. "Incremental residual concept bottleneck models." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. [6] Yuksekgonul, Mert, Maggie Wang, and James Zou. "Post-hoc concept bottleneck models." arXiv preprint arXiv:2205.15480 (2022).

5 comments
AuthorsOpenReview ↗

Response to Reviewer kqHM

We thank the reviewer for the detailed questions and constructive suggestions. We have revised the manuscript to clarify the theoretical scope, intervention protocols, and limitations.

Q1. Why is excess information less localizable?

Excess information is not inherently less localizable; it may itself have a localized source. The issue is that concept supervision does not specify where this additional information should originate. We have clarified this distinction in Sec. 3.1. Transport exposes the token sources of concept evidence, while Fisher regularization constrains the residual's off-support sensitivity.

Q2. Categorical and continuous concepts.

Lemma 1 establishes independence of concept-wise transport optimization, not statistical independence of concept labels or disjoint spatial supports. Related concepts can therefore share tokens, as illustrated by the belly-color concepts in Fig. 1. Our current formulation treats concepts as independent binary attributes without enforcing categorical constraints. Extending it to categorical and continuous concepts would require modified readouts and corresponding generalizations of the capacity and intervention guarantees, as discussed in the Conclusion.

Q3. The role of γ\gamma.

We use γ=1\gamma=1 throughout to avoid dataset-specific tuning. Increasing γ\gamma suppresses low-belief residuals more strongly, whereas smaller values transmit more continuous information. Theorem 4 quantifies the corresponding gate contraction when correcting confidently wrong concepts. The intervention-direction guarantee holds for any γ>0\gamma>0 with η<1/2\eta<1/2. We have added this explanation after Eq. (8).

Q4. How much per-sample variation is needed?

CUB also uses class-level concept annotations, obtained through majority voting, yet its transport plans exhibit strong object-level localization: 91% of the mass lies inside the bird bounding box, with part-level pointing outperforming the evaluated baselines (Table 10; Appendix C.3.1). This shows that within-class variation in concept labels is not strictly necessary for spatially concentrated transport. We believe concept discriminability may be more relevant: distinguishing visually similar bird species can encourage reliance on localized attributes, whereas in AwA2, species-level cues may suffice to predict class-level attributes (Appendix C.3.3). However, this does not guarantee correct part-level grounding, and determining how much instance-level supervision is required would need a controlled study. We have clarified this distinction in Appendix C.3.3.

Q5. Incomplete-concept CelebA.

Recovering approximately one-third of the accuracy gap is an observed outcome, not a predefined target. We do not claim that competing models explicitly represent hidden attributes as concepts. Table 4 measures their decodability: for Wavy Hair, ROC-AUC is 0.888 from CEM embeddings, 0.791 from ProbCBM embeddings, and 0.730 from our full bottleneck, compared with 0.712 from ground-truth concepts.

We also do not claim that our residual contains no hidden information. Theorem 2 bounds only the identity channel to one bit per concept. The residual is bounded in amplitude and regularized in off-support sensitivity, but is not guaranteed leakage-free. We have clarified this trade-off in the revision.

Q6. Intervention AUC.

Intervention AUC is the mean task accuracy over all nonzero intervention levels (Appendix B.4). On CUB, FisherCBM achieves 93.9±0.193.9\pm0.1%, compared with 94.0±0.394.0\pm0.3% for CEM (Table 3). We have replaced "equal" with "comparable" and clarified the definition in Sec. 4.2.

Q7. Interpretation of incorrect interventions.

The full incorrect-intervention accuracy is used as a behavioral indicator of reliance on intervenable concept states, rather than a direct information-theoretic measure of leakage. We agree that AUC is partly affected by clean accuracy, which is why Table 3 also reports half- and full-intervention accuracy.

On AwA2, FisherCBM and CEM have nearly identical clean task accuracy (96.1% vs. 95.9%) and concept accuracy (both 99.2%), yet retain 4.8% and 47.7% accuracy, respectively, after half the concept groups are corrupted under the original protocol. Our additional CUB evaluation using valid alternative group states further addresses the protocol concern: at 14/28 corrupted groups, FisherCBM retains 25.8% accuracy versus 62.9% for CEM (Fig. 5). These results support greater reliance on intervenable concept states without equating incorrect-intervention accuracy with a direct leakage measure.

Q8. Validity of incorrect group interventions.

The group-wise CUB evaluation follows the CEM setting, while the original incorrect-intervention protocol flips selected binary attributes as ci↦1−cic_i\mapsto1-c_i. We acknowledge that this can produce invalid categorical assignments.

We therefore conducted an additional experiment replacing each selected group with a different group state observed in the CUB training annotations, without retraining any model. This preserves valid within-group configurations, although the resulting combination across groups need not correspond to an observed class. At full intervention, FisherCBM and Hard CBM fall to 0.04% and 0.01% accuracy, respectively, whereas CEM retains 7.1% and CEM without RandInt retains 68.5% (Fig. 5). Although the original bit-flip protocol produces a sharper decline, the substantial difference between FisherCBM and CEM persists under valid group-state substitutions.

Additional Comments: Related Work, Clarity, and Limitations.

We have expanded Sec. 2 to discuss recent leakage analyses (Zarlenga, 2026; Almudévar et al., 2026), locality-aware CBMs (Jeon et al., 2025), and related concept-residual approaches. Unlike a separate residual side channel, FisherCBM attaches a bounded residual to each hard concept coordinate, preserving its identity under interventions. We have also added Algorithm 1, clarified the intervention metrics and terminology, and expanded the discussion of limitations, including part-level grounding, the Fisher penalty's effective regularization scale, incomplete concepts, and the current binary formulation.

I thank the authors for their responses to my review. Most concerns have been addressed. It would be nice to highlight changes in the paper during the rebuttal with text in a different color, which would make it easier to check the paper and identify updated and changed parts. Below are some remaining points:

Role of γ\gamma.

  • Do you have any results where γ\gamma is needed? Because if this parameter might be helpful in some theoretical circumstances, but never in practice, it could also be dropped from the method to make it a bit simpler and easier to understand.

Interpretation of incorrect interventions:

  • The predictors of Fisher-CBM and CEM still differ and also affect the outcomes. I agree that a difference between 4.8% and 47.7% is quite large and unlikely to be affected only by this, but I still think this fact should be discussed when presenting and discussing the experiment.
Reviewer_oGLsOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles two important problems of concept bottleneck models (CBMs): 1) concept leakage, where concepts encode unrelated visual information, 2) some concepts activate based on wrong image regions. The authors argue both come from predicting each concept from the whole image. To address this, they propose an optimal transport based formulation to learn concepts and a Fisher penalty term that penalizes the model when the information it passes on about a concept is sensitive to image regions that the model itself considers unrelated to that concept (this is a circular mechanism). The paper reports results on CUB, Awa and CelebA datasets.

Critical strengths

The paper tackles an important problem: concept leakage in concept bottleneck models.

Critical weaknesses

The paper has several major weaknesses. The most important one is that results do not support the main claim and proposed methods. First, the only result about the main claim of the paper (that problems come from predicting each concept from the whole image) is shown on only one image in a qualitative manner (Figure 1). Such a central claim should have been substantiated with experiments and performance measures.

The paper essentially has two proposals: 1) an optimal transport formulation for learning concepts, and 2) Fisher penalty term. Neither make a difference in performance. The ablation Tables 5–7 (Appendix C.2) show that removing the mass (which is the only thing that separates the OT formulation from plain per-concept softmax attention) causes mixed effect: it helps on CUB, does nothing on AwA2, and hurts slightly on CelebA.

When the Fisher penalty term is removed (Beta=0 in Tables 5-7 in Appendix C.2), the performance hardly changes.

So, the proposed mechanisms are not validated at all.

The method section is extremely hard to follow/understand. The method section takes up 5 pages + 10 pages of Appendix with proofs, both of which contain very-hard-to-follow mathematics.

The method assumes that ground-truth concept labels (done by humans) are available. This drastically limits where the method can be applied. Modern CBMs have long moved on to VLM-CBMs. The authors should have validated their methods on the modern VLM-CBMs, too.

There are other major weaknesses in the paper but I think the ones above are more than sufficient to make a decision.

Questions for authors

I do not have any questions. Authors can provide answers to major weaknesses above, if they wish to.

5 comments
AuthorsOpenReview ↗

Response to Reviewer oGLs

We thank the reviewer for the feedback. We have revised the manuscript to clarify our motivation, theoretical claims, and experimental evidence.

Circular Mechanism

The transport plan defines the support, but does not determine the residual's sensitivity outside it, which also depends on token and prototype geometry (Theorem 3). The support weights are detached in the Fisher penalty (Eq. 13), so the penalty acts on residual sensitivity rather than directly optimizing the support weights. Semantic alignment is evaluated independently using bounding boxes and part annotations (Table 10; Appendix C.3.1).

W1. Motivation and Experimental Evidence

We agree that the causal relationship between whole-image concept prediction and concept leakage was proposed as motivation rather than directly established. We have revised the Abstract and Introduction to distinguish concept leakage from spatial attribution.

Figure 1 is illustrative, while Appendix C.3 provides additional visualizations across CUB, AwA2, and CelebA. Quantitatively, 91.0% of the transport mass lies inside the bird bounding box across 1,000 CUB images and 112 concepts, compared with 59.0% for ProbCBM and 52.7% for uniform allocation (Table 10). We also acknowledge the weaker part-level alignment (Appendix C.3.1).

Regarding leakage, FisherCBM and CEM achieve nearly identical task accuracy (96.1% vs. 95.9%) and concept accuracy (99.2%) on AwA2, yet retain 4.8% and 47.7% accuracy, respectively, after half the concept groups are incorrectly intervened (Table 3). This supports stronger reliance on intervenable concepts, without establishing the proposed causal relationship.

W2. Effectiveness of Concept Transport

The effect of transport mass depends on concept granularity and the evidential readout. On CUB, where 112 fine-grained concepts require discriminative evidence, removing mass reduces task accuracy from 82.5% to 79.3%, or 78.3% after temperature matching (Table 5). On AwA2, the coarse class-level attributes and near-saturated performance leave little room for improvement. IUCT introduces a data-dependent evidence mass beyond normalized attention (Eqs. 3–5). CelebA differs because its annotated concepts are inherently incomplete: the task label also depends on two hidden attributes. An accuracy gap is therefore expected when restricting information beyond the concepts. Removing mass improves CelebA accuracy by 1.3 points at the default temperature, but reduces it by 2.0 points after temperature matching (Table 7). These results reflect the accuracy–information-control trade-off rather than a uniform benefit of transport mass across datasets.

W3. Effectiveness of Fisher Regularization

The Fisher penalty targets off-support residual sensitivity rather than classification accuracy. Theorem 3 shows that transport already suppresses this sensitivity, explaining why the default β=1\beta=1 has negligible additional effect.

To assess the Fisher penalty beyond task accuracy, we added post-hoc sensitivity measurements (Table 8) and finite token-swap experiments (Table 9). At β=105\beta=10^5, the off-support sensitivity share decreases from 4.9×10−44.9\times10^{-4} to 1.7×10−41.7\times10^{-4} on CUB (65%), and from 10.0×10−410.0\times10^{-4} to 2.2×10−42.2\times10^{-4} on CelebA (78%), with unchanged CUB task accuracy and a 0.3-point decrease on CelebA (Tables 5, 7, and 8). Finite token-swap experiments further show that increasing β\beta from 0 to 10610^6 reduces the near/on residual-response ratio from 0.37 to 0.06 and the fraction of changed predictions from 16.9% to 7.9% (Table 9).

These results support the effectiveness of Fisher regularization in controlling residual sensitivity, rather than improving accuracy or semantic localization. Indeed, stronger regularization reduces CUB in-box mass from 91% to 84%, a limitation we now explicitly acknowledge.

W4. Clarity and Complexity

We have added a three-stage overview and Algorithm 1 in Section 3, covering closed-form concept transport (Eq. 3), the evidential dual-channel bottleneck (Eq. 8), and Fisher regularization using an explicit token Jacobian (Eqs. 11–13).

Neither iterative OT optimization nor second-order automatic differentiation is required. The theoretical derivations establish the properties of these components rather than introducing additional computational stages.

W5. Concept Supervision and VLM-Based CBMs

Our work focuses on the supervised CBM setting to isolate the properties of the bottleneck mechanism, including information transmission, spatial attribution, and intervention guarantees. FisherCBM requires concept targets, as do the supervised baselines, but introduces no additional annotation requirements.

The framework could potentially use VLM-generated concept labels as supervision. However, we have not evaluated this setting and do not claim that the present results establish its effectiveness for VLM-based CBMs.

I thank the authors for their effort and time put in the rebuttal.

I have read it together with the other reviews, the discussions, and I looked at the revised PDF.

First of all, I accept the supervised-CBM scope. The lack of a VLM-CBM evaluation should not be a ground for rejection. Although, it should be put in the paper as an acknowledged limitation. And, the overview together with Algorithm 1 help. I also acknowledge that the submission already reported object-level localization (in-box mass), which I missed in my review.

However, I still think that my primary concerns are standing. And there are new issues with the revised pdf.

My primary concerns that still stand

Fisher penalty is one of the major contributions of the paper. It is in the title. However, as I pointed out in my review, it doesn't make any significant difference in task accuracy, concept accuracy, correct AUC and incorrect AUC (see tables 5-7). The authors added Table 9 for this concern, and it shows the same thing: no difference betwen Beta=0 and Beta=1 (Beta is the loss weight for Fisher penalty). I am comparing Beta=0 and Beta=1 because all main results (listed in 4 bullet points below) use Beta=1.

  • Table 1: task and concept accuracy (CUB 82.5 / 98.0, AwA2 96.1 / 99.2, CelebA 33.2 / 89.1).
  • Figure 2 and Table 3: the intervention curves (rises to 100% under correct interventions, falls to chance under wrong ones).
  • Table 4: the CelebA hidden-attribute probes.
  • Section 4.3 and Figure 1: the localization numbers (91% of mass inside the bird box).

But as the paper's own tables show there is no meaningful difference that the Fisher penalty term make.

The situation with Transport mass, which is the other contribution, is similar to above. It helps on CUB only. For CelebA, the rebuttal cites the temperature-matched drop (−2.0), but the revised Appendix C.2 still attributes that drop to the dataset's temperature preference "rather than the absence of the mass".

So, these results do not show that two proposed components (that are both in the title) are responsible for the reported performance.

I also mentioned in my review that the paper has 5-pages of hard-to-crack methods. This leaves only 1-page for experiments. I find this oddly imbalanced.

New issues

The previous (originally submitted) introduction, referring two concept leakage and poor localization problem, had this sentence: "We argue that these two failures are one: a concept predicted from the entire input has no region to be localized to, and whatever it draws from outside the concept’s own region is exactly the surplus that leaks."

The revised pdf says "These failures are related but distinct: leakage concerns information beyond the annotated concepts, whereas spatial attribution concerns the evidence used to predict each concept."

WHile this narrowing claim is welcome, the paper is still built around the old motivation. The old claim was the design rationale: one mechanism was meant to fix both problems. If they are distinct, each needs its own evidence, right? The Fisher penalty also loses its motivation, since confining the residual to the support limits leakage only if leakage comes from outside it. Yet only the Abstract and Introduction changed, so the rest of the paper still argues for the old story.

I had to search for this sentence. I think the auhors should have uploaded a track-changes style pdf (this can be done using latexdiff or manually coloring the added parts, strikethrough'ing the deleted parts), otherwise it is too difficult to track changes.

The rebuttal cites FisherCBM’s 4.8% vs. CEM’s 47.7% on AwA2 as leakage evidence, but the plain hard CBM reaches 3.1% in the same table (Table 3), so this result comes from the hard concept bit and not from the proposed transport or Fisher penalty.

AuthorsOpenReview ↗

Response to Reviewer oGLs

We thank the reviewer for carefully examining our rebuttal, the revised manuscript, and the discussions. We appreciate the acknowledgment of our supervised-CBM scope, the methodological overview and Algorithm 1, and the existing object-level localization evaluation. We have also provided a change-tracked version of the manuscript to make the revisions easier to identify.

We agree that our previous responses did not sufficiently distinguish the intended contributions of the overall architecture from those of the individual components. We clarify these points below.

1. Fisher analysis, regularization, and the main results.

We agree that the Fisher penalty at the default weight β=1\beta=1 does not produce a measurable improvement over β=0\beta=0 in the main predictive or intervention metrics. We do not regard this penalty as the primary source of the reported performance gains.

Our work has two objectives. The first is input-to-concept interpretability: making explicit which input tokens contribute to each concept prediction. IUCT provides concept-specific transport plans that directly determine concept features and evidence. The second is to recover useful task-relevant information that a finite set of annotated concepts may not capture, without losing the dominance of hard concept identities under interventions. We address this through a bounded, belief-gated residual.

The distinction between Fisher analysis and Fisher regularization is important here. Proposition 2 and Theorem 3 characterize the residual score's off-support sensitivity through a closed-form token Jacobian, with the gate and selected concept state held fixed. In particular, this Jacobian is scaled by the transport allocation, and the allocation-dependent factor in each off-support Fisher summand satisfies (1−Nπn,i)+πn,i2≤4/(27N2)(1-N\pi_{n,i})+\pi{n,i}^2\leq 4/(27N^2). These analytical properties hold even when β=0\beta=0. They do not establish an unconditional bound on total Fisher information or on the mutual information carried by the residual. The penalty is an additional mechanism for controlling this sensitivity. At β=1\beta=1, its effect is negligible because the transport-based architecture already exhibits low off-support sensitivity. At β=105\beta=10^5, however, the off-support sensitivity share decreases by approximately 65% on CUB and 78% on CelebA, with unchanged or nearly unchanged task accuracy (Table 9). The finite token-swap experiment confirms a corresponding reduction in the near-/on-support residual response ratio from 0.37 to 0.18 on CUB (Table 10). Stronger regularization also makes the plans more concentrated, but does not necessarily improve their semantic localization.

We therefore distinguish the evidence supporting the main results. Table 1 evaluates the predictive performance of the complete architecture; Figure 2 and Table 3 characterize reliance on intervenable concept states; Table 5 measures the decodability of hidden attributes from bottleneck representations; and Figure 1 and Section 4.3 evaluate the grounding of the transport plans. None of these results establishes a separate performance improvement caused by the default Fisher penalty. The independent contribution of the Fisher analysis and regularization is instead the analytical characterization and adjustable control of off-support residual sensitivity.

We recognize that the current title may obscure this distinction by placing substantial emphasis on the Fisher penalty. We will revise the title for the camera-ready version, if permitted, so that it no longer presents the Fisher penalty as a core mechanism. The intended emphasis is on closed-form concept transport and the bounded, belief-gated residual.

2. The role of transport mass.

We agree that transport mass does not uniformly improve classification accuracy across datasets. Its primary role in IUCT is to distinguish the amount of evidence supporting a concept state from the normalized token allocation. In Eq. (5), the evidence margin depends on both the transported direction and its mass; the latter consequently affects the evidential readout and belief gate.

The clearest predictive benefit appears on CUB, where removing mass reduces task accuracy from 82.5% to 79.3% at the original evidence temperature (Table 6). This supports its usefulness for fine-grained concept evidence, but we do not claim the same benefit across all datasets. On AwA2 and CelebA, the effects are small or mixed. In particular, we agree that the temperature-matched CelebA experiment does not isolate the effect of mass because it simultaneously changes the evidence temperature. We will clarify this distinction in Appendix C.2 and avoid presenting the temperature-matched result as an independent performance gain from mass.

Reviewer_k9YDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FisherCBM, which formulates concept prediction as an unbalanced optimal transport problem with a closed-form solution. This formulation decomposes the transport into a transport plan and a transport mass, capturing, respectively, where concept evidence is obtained and the strength of that evidence. Furthermore, FisherCBM combines a hard one-bit concept state with a gated residual, improving interpretability while limiting the influence of additional continuous information on downstream predictions. This design allows FisherCBM to achieve task performance competitive with CEM while offering stronger interpretability. In particular, the intervention results demonstrate that the model's predictions respond consistently to human-guided concept corrections, supporting the effectiveness of its interpretable bottleneck.

Critical strengths
  1. The formulation of unbalanced optimal transport problem naturally provides concept-level saliency maps and evidence strength. Compared with previous conventional CBMs that predict concepts from the whole input, the explicit saliency maps offer an additional message to examine whether the model relies on visual evidence consistent with human expectations, thereby enhancing interpretability.

  2. The gated residual leverages the evidence obtained from concept prediction. By combining a hard one-bit concept state with a gated residual (η<0.5\eta < 0.5), the method preserves semantic identity of each concept while allowing additional continuous information to support downstream task. Together with RandInt, this design provides a principled way to improve intervention reliability without sacrificing much predictive performance.

  3. The paper writting is well motivated and clearly organized.

Critical weaknesses
  1. Although the transport plan identifies which regions contribute to each concept, it is determined by learned prototypes and does not necessarily correspond to semantically relevant regions. Or it only supports theoretical soundness, but not necessarily semantic grounding. And the Fisher penalty also use the transport plan, it cannot directly correct such misalignments when irrelevant patches are assigned high weights. The localization results also suggest that the method is more effective at identifying objects than specific concept-related parts, which limits its interpretability claims.

  2. FisherCBM achieves competitive task accuracy and favorable intervention behavior, but intervention results do not demonstrate a consistent advantage over existing CBMs. In Figure 2, its correct-intervention performance is comparable to CEM, while its incorrect-intervention sensitivity generally falls between hard CBMs and CEM. These results support an accuracy–interpretability trade-off. Consequently, the main practical advantage appears to lie in the explicit transport-based evidence representation, while the additional benefits of the full architecture remain less clear.

Questions for authors

The method introduces several strategies, including evidential gating, Fisher regularization, and RandInt training. However, the ablation results suggest that not all of them provide substantial additional benefits. In particular, removing the Fisher penalty or RandInt produces nearly identical results under the default settings. Moreover, the current ablations do not fully isolate the contribution of the belief gate. Further clarification of the necessity of these components and comparisons with simpler alternatives would help explain which parts of the proposed architecture are essential.

3 comments
AuthorsOpenReview ↗

Response to Reviewer k9YD

We thank the reviewer for the constructive comments. We have revised the manuscript to clarify the interpretability claims, intervention guarantees, and contributions of individual components.

W1. Semantic Grounding of Transport Plans

We agree that the transport plans are not guaranteed to align with semantically correct regions, and Fisher regularization cannot correct a misplaced support. However, the plans are faithful to the model's token-level readout by construction: concept features are directly aggregated using the transport weights (Eq. 4), rather than explained through post-hoc attribution. This does not guarantee pixel-level grounding, as ViT tokens may contain information from other regions.

On 1,000 CUB test images, 91.0% of the transport mass lies inside the bird bounding box, compared with 59.0% for ProbCBM and 52.7% for uniform allocation (Table 10). Part-level alignment remains more limited, with 26% of argmax tokens falling within two tokens of the annotated part (Appendix C.3.1). We have clarified that our interpretability claim concerns faithful and inspectable token-level evidence, rather than guaranteed part-level localization.

W2. Intervention Performance and Trade-off

We agree that FisherCBM does not outperform CEM on every metric. Its advantage is combining competitive task accuracy with stronger dependence on intervenable concept states.

On CUB, FisherCBM achieves 82.5% task accuracy (CEM: 83.6%) and comparable correct-intervention AUC (93.9% vs. 94.0%), with 4.7 times fewer trainable parameters. Under half incorrect group interventions, FisherCBM retains only 0.1% accuracy, compared with 6.3% for CEM (Tables 1–3). On AwA2, where both models achieve nearly identical task accuracy (96.1% vs. 95.9%) and concept accuracy (99.2%), the corresponding accuracies are 4.8% and 47.7%, respectively.

The RandInt ablation provides further evidence. Without RandInt, CEM's half-incorrect accuracy on CUB increases from 6.3% to 58.6%, while its correct-intervention endpoint decreases from 99.6% to 92.0%. In contrast, FisherCBM's intervention curves remain nearly unchanged across all three datasets (Fig. 4). This is consistent with Theorem 4, which guarantees the direction and minimum magnitude of a corrected bottleneck coordinate independently of RandInt. Prediction-level improvements additionally depend on the head condition in Theorem 5.

On incomplete-concept CelebA, the trade-off remains: FisherCBM recovers less task accuracy than CEM while retaining less decodable hidden-attribute information (Table 4). We have clarified these distinctions rather than claiming a uniform advantage.

Q. Necessity of Individual Components

Transport mass and two-sided evidence. Replacing IUCT with per-concept softmax attention (mi≡1m_i\equiv1) reduces CUB task accuracy by 3.2 points, or 4.2 points after temperature matching (Table 5). Using a single transport also reduces CUB concept accuracy by 0.9 points and CelebA task accuracy by 1.6 points (Tables 5 and 7). We expanded the component ablations across all three datasets (Tables 5–7) and introduced additional sensitivity analyses to distinguish the effect of Fisher regularization from that of the transport mechanism (Tables 8–9).

RandInt. Its removal has little effect on FisherCBM's intervention curves, unlike CEM (W2). RandInt encourages the classifier-head condition of Theorem 5 (Proposition 4), whereas the coordinate-level guarantee of Theorem 4 follows from the bottleneck design.

Belief gate. At fixed η=1\eta=1, retaining the gate improves CUB task accuracy from 81.4% to 82.2% and concept accuracy from 97.6% to 98.0% (Table 5). Its accuracy effects on AwA2 and CelebA are small. Its primary purpose is to regulate residual transmission according to concept belief.

Fisher regularization. At the default β=1\beta=1, its effect is negligible because transport already suppresses off-support sensitivity. At β=105\beta=10^5, the off-support sensitivity share decreases from 4.9×10−44.9\times10^{-4} to 1.7×10−41.7\times10^{-4} on CUB (65%), and from 10.0×10−410.0\times10^{-4} to 2.2×10−42.2\times10^{-4} on CelebA (78%), with unchanged CUB task accuracy and a 0.3-point decrease on CelebA (Tables 5, 7, and 8). However, stronger regularization does not improve semantic localization: CUB in-box mass decreases from 91% to 84%. We therefore position the Fisher penalty as a mechanism for controlling residual sensitivity, not as a source of classification accuracy or guaranteed grounding.

These results clarify the distinct roles of the components, and we have updated the manuscript accordingly.

Thanks for your clarification and for running additional experiments. The clarifications and additional experiments address part of my concerns, in particular the scope of the interpretability claim and the positioning of the intervention results. The Fisher penalty does not seem to be a main contributor to the reported results. I suggest either identifying settings where the Fisher penalty provides a clear benefit, or placing more emphasis on the contributions of the other components. Given the competitive task accuracy and improved intervention performance. I will retain my original score.

We thank the reviewer for acknowledging the clarifications and additional experiments. We agree that the Fisher penalty is not the primary contributor to the default model's accuracy and intervention performance. These results arise mainly from IUCT and the bounded dual-channel bottleneck, whereas Fisher regularization provides an explicit measure and additional control of off-support residual sensitivity. As shown in Table 8, stronger regularization reduces this sensitivity, while the default β=1\beta=1 has negligible effect.

We will revise the Abstract and reorder the contributions in the Introduction to emphasize closed-form concept transport, the bounded dual-channel bottleneck, and intervention guarantees. We will also clarify the role of Fisher regularization in Sec. 3.4, distinguishing the structural confinement provided by the architecture from the additional sensitivity control offered by the penalty.

In particular, the term "Fisher-bounded" in the title refers to the structural bound derived from Theorem 3 and Appendix A.1.7, where the allocation-dependent factor in each off-support Fisher contribution is bounded by 4/(27N2)4/(27N^2). The penalty provides additional control when appropriately weighted, rather than being the source of the default model's predictive improvements. We will make this distinction explicit in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback. We have revised the manuscript to improve clarity, strengthen the experimental analysis, and clarify the scope of our theoretical claims.

Clarity and Presentation

We added a three-stage overview and Algorithm 1 in Sec. 3 to make the training and inference procedures easier to follow. We also clarified that concept leakage and spatial attribution are related but distinct problems, and distinguished the coordinate-level intervention guarantee (Theorem 4) from the additional condition required for prediction-level improvements (Theorem 5).

Component Analysis

We expanded the ablations and added diagnostic experiments to distinguish the contributions of individual components:

ComponentMain Finding
Transport massRemoving it reduces CUB task accuracy by 3.2 points (4.2 after temperature matching; Table 5).
Two-sided evidenceUsing a single transport reduces CUB concept accuracy by 0.9 points and CelebA task accuracy by 1.6 points (Tables 5 and 7).
Belief gateAt fixed η=1\eta=1, removing it reduces CUB task accuracy by 0.8 points (Table 5).
Fisher penaltyAt β=1\beta=1, its effect is negligible. At β=105\beta=10^5, off-support sensitivity share drops from 4.94.9 to 1.71.7 on CUB and 10.010.0 to 2.22.2 on CelebA (all ×10−4\times10^{-4}; Table 8). Additional finite token-swap experiments validate sensitivity control (Table 9).

Fisher Regularization and Localization

Transport already suppresses off-support sensitivity to the 10−410^{-4}–10−310^{-3} range, explaining why the default Fisher penalty has little additional effect (Theorem 3; Table 8). Stronger regularization provides measurable control, further supported by finite token-swap experiments (Table 9). However, sharper plans do not necessarily improve semantic localization: on CUB, in-box mass decreases from 91% to 84% under stronger regularization. We have clarified that Fisher regularization controls residual sensitivity rather than guaranteeing semantic grounding.

Intervention Reliability

FisherCBM's intervention curves remain nearly unchanged without RandInt, whereas removing RandInt from CEM increases half-incorrect accuracy on CUB from 6.3% to 58.6% and reduces its correct-intervention endpoint from 99.6% to 92.0% (Table 3; Fig. 4). This supports the importance of the bottleneck design, although prediction-level guarantees remain conditional on the classifier head.

We also conducted an additional CUB evaluation using valid alternative group states to address the concern about incorrect categorical interventions (Q8).

We discuss the remaining limitations, including weaker part-level grounding and the restriction to binary concepts, in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.