acceptodds
Withdrawn from ICLR 2027

Matryoshka Concept Bottleneck Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Concept bottleneck models predict through named attributes that can be inspected and replaced. Several readouts can give a frozen encoder different input budgets. We study joint training that makes those budgets part of learning the encoder itself. Matryoshka Concept Bottleneck Models (MCBMs) optimize one concept encoder with task losses on nested prefixes of a fixed concept order. Each prefix has a trained head and a stable set of concept identities. At deployment, a query policy operates within the selected prefix. On CUB, the same model reaches accuracy with 64 concepts and with all 112, reducing visible inputs by . Single-seed studies on CelebA and the animal domain of LAD show similar retention at moderate budgets. Readout studies on the same CUB encoders examine a separate property: response to concept replacements. Active queries improve probability-head curves, while intervention-aware training improves raw-head responses. These results support joint multi-budget training as a practical design for named concept interfaces. They also separate task accuracy at reduced budgets from the ability of a particular head to use replacement values.

Primary area:
interpretability and explainable AI
Keywords:
Concept Bottleneck Models, Matryoshka Representation Learning, Explainable AI.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7o44OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work introduces Matryoshka Concept Bottleneck Models (MCBMs). MCBMs decrease the cardinality of the concept set involved in predictions. The authors use feature-selection based on minimum redundancy maximum relevance (mRMR) so as to order the concepts in terms of informativeness, and then train multiple classification heads. Each of these heads uses only a subset of the sorted concept set, thus leading to increased interpretability due to including only a subset of the whole available concept set, i.e. a sparser concept set.

Critical strengths

The motivation of the work is sound and novel to the CBM literature. I believe that CBMs would benefit from the idea of using concept subsets. More specifically:

  • Using only a subset of concepts per prediction, is, to the best of my knowledge, a new idea on CBMs. These subsets can help with interpretability and annotation efficiency of CBMs, especially in datasets with high concept set cardinality, since the end-user will have to read (or annotate) a fewer number of concepts.
  • The mRMR-based feature selection idea is used before training, thus being possible to integrate to other CBM (or CEM) architectures.
Critical weaknesses

Clarity: The main weakness of this work is the clarity of the text. In most cases, terminology outside of the machine learning field and specifically the CBM literature is used, making the text very hard to understand. I strongly suggest that the authors polish their text.

Experimental results: I also struggle with the experimental section. The authors selected to run single-seed studies on two out of three datasets. Also, for the dataset that has been run for multiple seeds, in Fig. 2b, the authors don't report multi-seed error bars, and in Table 2 they do not include standard deviations. Furthermore, the experiments include "unrun comparisons" (Line 279) in Table 1.

I think it would have also been useful to test MCBMs on other CBM architectures such as SCBM.

Finally, the authors claim to have used the test set during development time as seen in: "The test sets have been evaluated during development, so the results describe the reported configurations rather than a fresh confirmation study." (Lines 469-470) and "The test set has been observed in earlier project experiments; the present comparisons are not a fresh holdout study." (Lines 603-604)

On top of that, the authors did not include any supplementary material and there isn't a reproducibility statement included.

All of the above, make me skeptical about the validity of the presented experimental results.

Typos

  • Lines 097-098: "... while stochastic CBMs model concept dependencies...", please capitalize Stochastic, as seen in the rest of the text.
  • Figure 1: unless I am mistaken, in the equation for LtotalL_{total} it should write Lconcept(c^,c)L_{concept}(\hat{c},c) instead of Lconcept(1,c)L_{concept}(1,c). Also, shouldn't the index on the heads, go "1, 2, 3, N" instead of "1, 2, n, N"?
  • Lines 306-307: "On CUB, moving from 112 to 64 concepts changes accuracy by −0.092 ± 0.132 percentage points across paired seeds." I think it is not percentage points.
Questions for authors
  • Maybe I missed it, but given that one task head is trained per concept subset, wouldn't this lead to training KK different task heads? Can you provide results on the computational efficiency of MCBMs, including the hardware used for them?
  • How would the concept subset selection of mRMR and multiple task heads look like in a Concept Embedding Model [1], since embeddings of concepts are used instead of the concepts themselves?

[1] Espinosa Zarlenga, Mateo, et al. "Concept embedding models: Beyond the accuracy-explainability trade-off." Advances in neural information processing systems 35 (2022): 21400-21413.

Reviewer_pbN5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The work trains image classifiers in two steps, by first classifying attributes and then using those attribute logits to predict final categories such as bird species. This is the existing concept bottleneck approach. The paper does this using an end to end objective, and by training multiple final category classifiers at once that rely on numbers of nested attributes (8, 16, 24, 32, 64 and 112). The attributes used are ordered with mRMR (minimum redundancy maximum relevance, Peng et al 2005), a preexisting method using mutual information so that the attributes containing more new information are prioritized. The method is evaluated on three datasets, CUB, CelebA and an animal subset of LAD. Some ablations are done, e.g. with the order of attributes (with a random order absent) and number of trained downstream classifiers (varying the number of trained classifier heads, 6 vs 2). An “intervention” setup is also considered where some attributes are swapped out with the ground truth using different policies for picking which ones are swapped out. The classification head needs to be trained to handle the interventions well on the logits. A key result is that CUB accuracy matches at 64 vs 112 attributes.

Critical strengths

None

Critical weaknesses

The paper is not well written and hard to follow, this makes everything else hard to audit. It took me a long time just to get through the abstract. E.g., ”Several readouts can give a frozen encoder different input budgets.”---there are too many such examples to list.

Besides the writing, the novelty is not clear since existing approaches are being combined directly without clearly supported gains. These are the two-step concept bottleneck approach, mRMR for ordering concepts, and training multiple classification heads. Finally, the lack of statistical tests means it's unclear what the benefit of the method is, and the native CUB model with 80.05% is not the one that was intervened on.

Questions for authors

None

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

The paper trains one CBM concept encoder with task losses on nested prefixes of a fixed mRMR concept order. Each prefix has its own linear head, so a single model yields predictions across several concept budgets. The paper also tests existing query policies (entropy, CooP) inside a fixed prefix and trains readouts to receive concept corrections. On CUB, the 64-concept head reaches almost the same accuracy as the 112-concept head of the same model. CelebA and LAD are smaller single-seed studies.

Critical strengths

Interesting idea:
Fixing a global, redundancy-aware concept order and training nested heads on it is easy to implement and adds almost no training cost. Most CBM work treats the bottleneck as a flat set, and the paper clearly argues why a globally consistent prefix differs from per-sample sparse selection, which cannot be predictably truncated.

Clear budget retention on cub:
In one model, the reduced-concept head reaches 80.05%, and the full-concept head reaches 80.14%. MCBM is also better than Joint64 and FullJoint over three seeds.

Transparent reporting of results:
The deployment budget b is separated from the query budget b to show differences. Intervention curves with harm and no first-success stopping are shown. Selection of CooP weights on validation data only. Makes results more reliable. Discussion of fixed-order queries being worse against FullJoin.

Critical weaknesses

Unclear practical benefit:
The reduction applies only to the inputs that the task head sees. The encoder still predicts all 112 concepts, and the paper measures no savings in compute, annotation, or human time.

Concept quality and leakage are not evaluated:
Accuracy drops when the heads read concept values rather than raw logits. Thus, the logits possibly carry class information beyond the concepts. The main paper gives no concept accuracy or calibration for the current encoders.

The evidence is fragmented, and the paper does not explain its choice of measures:
The results come from several encoder cohorts. CelebA and LAD use one seed. On CelebA, d = 20 is better than d = 39. The test set was seen during development. The paper uses nAUC, repair, and harm, but it does not explain why it chose these measures and not others. nAUC starts at the natural accuracy of each head, so a higher start point increases the score. At d = 64, the MCBM start point is 2.6 points above FullJoint. nAUC also gives the same weight to every query, but experts answer only a few concepts.

The paper also does not cite two recent, directly related works:
Tapli et al. (2026), “Rethinking Concept Bottleneck Models: From Pitfalls to Solutions,” report a “linearity problem”: recent CBMs circumvent the concept bottleneck. MCBM uses only linear heads on raw logits, so this risk applies directly. The work also provides an entropy-based metric for concept-set suitability that is close to the mRMR ranking.

Knab et al. (2026), “What’s in the bottle? A survey and roadmap of concept bottleneck models” (TMLR). The related-work section could be improved by using this survey to show where MCBM fits, particularly in leakage, intervention, and concept selection.

Knab, P., Steinmann, D., Bartelt, C., Kersting, K., Schiele, B., Seidl, T., Schlegel, U., & Stammer, W. (2026). What’s in the bottle? A survey and roadmap of concept bottleneck models. Transactions on Machine Learning Research.

Tapli, M., Bouniot, Q., Stammer, W., Akata, Z., & Akbas, E. (2026). Rethinking concept bottleneck models: From pitfalls to solutions (arXiv:2603.05629). arXiv.

Questions for authors

Do MCBM, FullJoint, and Joint64 use the same learning rate and schedule?

What are the concept accuracy and calibration of the current encoders at each prefix?

How many of the 200 CUB class prototypes are unique with only the first 8, 32, or 64 concepts?

Why is the untrained MCBM raw head so poor at using replacements?

Which real deployment needs a fixed, smaller concept interface?

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

Related papers

Discussion (0)

Sign in to comment.