acceptodds
Under review as a conference paper at ICLR 2027

AnyBottle: A Recipe to Only Keep the Concepts You Really Need

Abstract

Concept bottleneck models (CBMs) make predictions inspectable and intervenable by routing them through human-interpretable concepts, but originally required concept annotations. Annotation-free variants remove this requirement, but typically use large concept vocabularies, static at both training and inference, producing bottlenecks larger than any task or prediction needs and harder to inspect. We propose AnyBottle, a single recipe for building compact, task-specific CBMs. AnyBottle assumes only a frozen backbone and an unsupervised concept pool, such as a sparse autoencoder. A black-box teacher trained on the same backbone then guides selection: each round adds the concept that best explains the bottleneck’s current failures, with candidates restricted to regions of teacher/student disagreement. Trained with nested dropout over this selection order, the final bottleneck predicts accurately from any concept prefix, so inference spends fewer concepts on inputs it is confident about early and more on hard ones. Since no stage is modality-specific, a new domain and task requires swapping only the backbone and concept pool. Across six vision and two text datasets and two teacher paradigms, AnyBottle yields bottlenecks with fewer concepts and higher concept consistency than annotation-free baselines, while staying close to the black-box reference. Overall, AnyBottle shows that going annotation-free need not mean going large: a small, discovered vocabulary can be as expressive as a much larger, fixed one.

Primary area:
interpretability and explainable AI
Keywords:
Concept Bottleneck Models, Interpretability, Sparse Autoencoders, Concept Discovery, Concept Selection, Nested Dropout, Adaptive Inference, Explainable AI

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_Zww5OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to reduce the global number of concepts in concept bottleneck models (CBMs), which output final predictions by using human-interpretable concepts as a bottleneck. Existing CBMs sparsely predict concepts for input images from a large concept pool. However, as the global number of concepts increases, human interpretability is compromised, which can hinder intervention and debuggability of concepts. This paper proposes a teacher-student-based recipe for constructing task-specific CBMs called AnyBottle. The goal of AnyBottle is to identify a subset of concepts required for a given task from an SAE with a large concept pool and achieve performance comparable to that of a black-box model. To this end, the paper introduces an algorithm based on orthogonal matching pursuit (OMP) that iteratively identifies important concepts. This algorithm focuses on samples where the predictions of the black-box teacher and the student currently being trained diverge, and extracts tokens that the teacher emphasizes but the current student fails to capture as "blind spots." From the set of concepts present in these blind spots, those with high activation scores are selected as a candidate set. The student is trained using nested dropout based on the selected set of concepts, and during inference, concepts are added sequentially to minimize the concept budget. In the experiments, AnyBottle is evaluated against existing CBMs on image and text datasets. As a result, AnyBottle achieves accuracy close to that of black-box models while maintaining a lower number of global concepts than existing CBMs. Additionally, analytical experiments were conducted to verify the effectiveness of each component.

Critical strengths
  • The paper proposes an approach to concept selection based on teacher-student residuals.
  • The paper proposes an inference method that dynamically adjusts the number of concepts for each sample using a student classifier trained with nested dropout.
  • The paper confirms that these techniques allow us to maintain accuracy even with a small number of global concepts.
Critical weaknesses
  • Although this paper proposes complex techniques for constructing CBMs with a small number of global concepts based on the blackbox teacher and SAE, the rationale for adopting these techniques is weak. In particular, if the goal is to achieve a small number of global concepts, reducing the number of intermediate dimensions in the SAE is the simplest baseline. Furthermore, constraining the overall budget using group lasso during task-specific fine-tuning of the SAE is also a promising option. The paper should provide evidence demonstrating why these baselines cannot achieve the goal, both theoretically and experimentally.
  • Naming by MAIA relies heavily on external models, and its semantic interpretation does not reflect the internal mechanisms of the model. While MAIA extracts image regions from concept activation heatmaps and assigns names to the discovered concepts, the meanings of these named concepts have no causal relationship with the actual model outputs. The interpretations provided in this way are misleading, making it difficult to apply MAIA in critical fields such as healthcare, where CBMs are expected to be utilized.
  • The relationship between compactness and interpretability has not been sufficiently verified. In particular, the paper lacks human intervention and concept debugging evaluations, making it impossible to determine whether the paper's objectives were achieved. The paper's primary motivation is the reduced interpretability for humans caused by a large number of global concepts. In this sense, it is essential to verify whether AnyBottle actually improved human interpretability.
  • The proposed concept selection algorithm may overfit to a small number of samples. Although the paper introduces a mechanism to prevent overfitting by selecting concepts based on the validation score, preserving validation accuracy is unrelated to the validity of the selected concepts. In the algorithm described in the paper, residuals are calculated using weights obtained via least squares with respect to the teacher's predicted distribution, rather than the discrete ground-truth labels (Eq. (1)). This carries the risk of incorporating concepts that are inherently unnecessary for reproducing the distribution, as well as the risk of overfitting to the validation set due to hundreds of optimization iterations.
  • There is a lack of evaluation regarding training time and computational resources. Since the concept selection process involves selecting one concept per round, this can result in a very large number of iterations in some cases. In particular, ISIC-7 requires as many as 447 concepts, raising concerns about the computational time required for optimization.
  • The comparison conditions with competing methods are not standardized, making the comparison unfair. While the proposed method uses CLIP-DINOiser, other baselines utilize different backbones. The accuracy gap relative to the black-box teacher is not the result of a fully controlled experiment and is significantly influenced by differences in backbone performance. Furthermore, it would be desirable to conduct experiments with standardized concept dictionaries and concept budgets for SAE-based CBMs.
Questions for authors

Please see the weakness section and address the concerns.

6 comments
AuthorsOpenReview ↗

Individual Response 1

W1. Rationale for concept selection

We implemented both suggested alternatives. Please see “Smaller SAEs,” “Group-lasso SAEs,” and “Representation Fine-tuning and Concept Selection” in our general response for the protocols and results.

On Imagenette, smaller SAEs retain accuracy but yield substantially lower concept consistency: at comparable budgets, a 32-latent SAE achieves C²patch of 0.28–0.29, versus 0.46 for AnyBottle’s 30 selected concepts.

Group-lasso fine-tuning gives dataset-dependent trade-offs: it is more compact on Machines-15 and more accurate on Dogs-10, whereas AnyBottle has higher consistency on both and retains more accuracy with fewer concepts on Flowers-102 and ISIC.

These results explain why learning a compact representation and selecting from a richer pool are not interchangeable. Fine-tuning determines the available features, can change their decomposition and can be difficult to finetune satisfactorily due to multi-objective finetuning. Selection identifies a task-relevant subset while keeping those features fixed.

Notably, AnyBottle already combines task-aware fine-tuning with selection, with the former evaluated in Table 4. We do not claim an optimal concept pool in our work: selection applies to pools chosen for their coverage or granularity. The contribution is therefore a selection procedure with demonstrated benefits, not a claim that representation learning cannot sometimes aid or replace it.

W2. Concept naming and semantic faithfulness

We agree that semantic descriptions require verification, particularly in healthcare. However, the statement that the named features have “no causal relationship with the actual model outputs” overlooks their explicit role in prediction.

Specifically, only naming is post hoc; the bottleneck is part of the prediction pathway. AnyBottle first selects features and trains the predictor on their activations. For the linear head, each feature contributes its activation multiplied by its class weight to the corresponding logit. Naming subsequently describes these features; it does not create their relationship with the output. This follows the classic procedure of bottom-up approaches (see [1] for the distinction).

Next, MAIA tests interpretations rather than merely labeling heatmaps. It generates or edits inputs and measures feature responses to distinguish candidate interpretations. Our additional CLIP-based evaluation also finds higher mean image–description alignment for MAIA than decoder-based naming on all six datasets (App. D.13). We have updated our manuscript to make this more clear and refer to the original work for further details (Sec. 3.4).

Overall, the concepts in AnyBottle have a clear and observable connection to the output, and while AnyBottle requires proper concept naming in the same way as all bottom-up concept approaches, its compact concept space allows hypothesis-driven naming approaches such as MAIA which go beyond standard CLIP-based naming procedures often used in CBMs nowadays. Nevertheless, interactive human verification remains important (as recommended in Sec. 3.4 and discussed in Sec. 6).

[1] Knab, Patrick, et al. "What’s in the bottle? a survey and roadmap of concept bottleneck models." Transactions on Machine Learning Research (2026).

AuthorsOpenReview ↗

Individual Response 2

W3. Human interpretability and concept debugging

We agree that compactness alone does not establish improved human "interpretability". What we tackle in this work is to reduce the effort required to inspect, name, and verify a concept bottleneck, while retaining predictive performance and coherent concepts. AnyBottle addresses this through both a compact global vocabulary and sample-dependent concept budgets. Our motivation for compactness is supported by prior human studies [2] supporting the practical importance of reducing explanation size. We will clarify this distinction in the framing of the paper (e.g. introduction).

Notably, our evaluation goes beyond concept count. We jointly measure predictive performance, vocabulary size, and concept consistency through C² and AutoInterpScore. Figure 7 compares selection against the same SAE dictionary and shows comparable or higher patch-level consistency with substantially fewer concepts on most datasets. Thus, the result is not simply a smaller intermediate representation.

We have also added a direct concept-debugging experiment (App. D.15; see general response “Shortcut Debugging”). After introducing a border shortcut in Imagenette, AnyBottle selects the border-sensitive concept first in all three seeds (and its most likely meaning can be identified through MAIA). Removing it and refitting only the linear head reduces the student’s susceptibility to the shortcut from 9.5% to 0.6%, while retaining performance on samples without the shortcut. The teacher and SAE remain unchanged.

These results establish compactness, concept coherence, and a concrete debugging capability. We agree that interpretability claims beyond this aspect of compactness and consistency such as downstream user ability, require comprehensive studies. We have updated our framing on this point to be aligned with our intentions. Thank you for raising that point.

[2] Ramaswamy et al. “Overlooked Factors in Concept-Based Explanations: Dataset Choice, Concept Learnability, and Human Capability.” CVPR, 2023.

W4. Selection targets and overfitting

The focus examples guide candidate proposal, but candidates are scored against teacher residuals across the training dataset. Proposal also aggregates evidence across distinct examples and filters insufficiently supported candidates. Selection therefore does not rely only on the small focus set. Tables 1 and 3 evaluate the benefits of this guidance against simpler selection variants.

Equation (1) intentionally targets teacher outputs: a feature can help reproduce confidence differences even when unnecessary for matching discrete class decisions. Using one-hot labels would change that objective. Residual reduction is not the sole stopping criterion, however, and the final predictor is trained against ground-truth labels. We do not claim that every selected concept is indispensable.

Validation accuracy controls predictive model selection and stopping, not semantic validity. We assess concept coherence separately through C², AutoInterp, and qualitative inspection.

Repeated validation use can fit noise. Our generalization results therefore come from a separate test split, and the stopping-tolerance analysis examines sensitivity, particularly on ISIC. We clarified in the Discussion that validation-based stopping is not a guarantee against overfitting or a globally minimal vocabulary.

AuthorsOpenReview ↗

Individual Response 3

W5. Computational costs

We added a stage-by-stage runtime analysis on a single RTX 4500 Ada GPU (App. D.14; see “Computational Cost” in our general response).

Briefly, selection takes 17 seconds for 28 rounds on Imagenette and about 20 minutes for approximately 450 concepts on ISIC. The full Imagenette construction takes 21 minutes versus 2.1 minutes for DN-CBM, excluding naming and external pretraining; the additional cost mainly comes from patch-level features and SAE fine-tuning. These costs occur during construction, not inference. Importantly, the smaller vocabulary reduces estimated MAIA naming time from 269 to 6.4 GPU-hours under the same naming procedure.

W6. Controlled comparisons across methods

Please see “Same Dictionary Runs” in our general response and App. D.9. We compare the SAE-based DN-CBM and AnyBottle under the exact same settings.

Briefly, both methods use DN-CBM’s frozen backbone and released SAE, the same teacher and data splits, and no SAE fine-tuning. AnyBottle uses 2.5–20× fewer concepts, maintaings comparable accuracy and improves image-level consistency. The only exception is ISIC-7, which is most out of distribution from the pretrained SAEs.

Importantly, our original evaluations standardize datasets, splits, validation-based model selection, and seeds. Teacher-relative accuracy measures how much performance each bottleneck retains within its own representation, although it does not eliminate all representation effects. The new experiment additionally controls the backbone and dictionary, showing that the observed benefits do not depend on CLIP-DINOiser or our task-adapted SAE.

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

This paper proposes ANYBOTTLE, a method for building compact, task-specific CBMs without concept annotations. It uses a SAE to generate concepts, selects useful concepts based on teacher-student differences, and applies Nested Dropout for adaptive inference. Experiments on vision and text datasets show that ANYBOTTLE achieves accuracy close to the teacher while using significantly fewer concepts.

Critical strengths

The paper combines SAE-based concept discovery, teacher-guided selection, and Nested Dropout in a practical framework. The adaptive concept selection makes the resulting CBMs more compact and easier to interpret.

The method is evaluated on six vision and two text datasets, with comparisons against several CBM baselines. The results show good accuracy and concept consistency with fewer concepts. The ablation studies also support the effectiveness of the main components.

Critical weaknesses

Dependence on teacher quality. The selected concepts may inherit shortcuts or spurious correlations from the teacher. The paper lacks experiments to evaluate this risk.

Unclear computational cost. The iterative selection process requires repeated training and can involve many rounds. The paper does not provide detailed runtime comparisons with other methods.

Questions for authors
  1. Can the authors evaluate whether the selected concepts inherit shortcuts or spurious correlations from the teacher?
  2. Can the authors report the runtime and computational cost of ANYBOTTLE compared with methods such as DN-CBM?
  3. Can the authors provide a quantitative comparison between adaptive inference and fixed-prefix inference under the same average concept budget?
3 comments
AuthorsOpenReview ↗

Individual Response

W1. Teacher shortcuts and concept debugging

We agree that selected concepts can inherit teacher shortcuts, as acknowledged in Section 5 (“Faithfulness, Not Correctness”). We have now added an experiment that directly evaluates this risk and the possibility of correction (App. D.15; see “Shortcut Debugging” in our general response).

We introduce a border shortcut in Imagenette, yielding a teacher which nearly always predicts the same class in the presence of the shortcut. AnyBottle also selects a border-sensitive concept first (which can be confirmed by the automated naming), and the student also falls for the shortcut in about 10% of cases. However, the named concepts allow to easily identify and remove this cocnept from AnyBottle's concept pool, and refitting only the linear head reduces shortcut dependency to <1%, while retaining the same accuracy on clean samples without the shortcut, without having to change the teacher of SAE at all.

While this demonstrates that AnyBottle can inherit shortcuts from its teacher model, the resuls also show that targeted correction is possible. However, detecting shortcuts remains a difficult challenge, which typically requires task knowledge, targeted tests, or continued expert interactions. AnyBottle provides a compact vocabulary that makes such checks easier, but does not replace them.

W2. Computational cost

We have added a stage-by-stage runtime comparison in Appendix D.14 and we refer to “Computational Cost” in our general response.

Briefly, selection takes 17 seconds for 28 rounds on Imagenette and about 20 minutes for approximately 450 concepts on ISIC. The full Imagenette construction takes 21 minutes versus 2.1 minutes for DN-CBM, excluding naming and external pretraining; the additional cost mainly comes from patch-level features and SAE fine-tuning. These costs occur during construction, not inference. Importantly, the smaller vocabulary reduces estimated MAIA naming time from 269 to 6.4 GPU-hours under the same naming procedure. All timings use the evaluated one-concept-per-round algorithm.

Q1. Shortcut concepts See our detailed response to W1 and the general response.

Q2. Computational Cost See our detailed general response.

Q3. Adaptive inference at matched budgets

We kindly refer to our detailed general response "Quantitative Adaptive Inference".

In summary, on the accuracy–budget curves, adaptive inference improves balanced accuracy on every dataset at the same average budget. At 50% of the available concepts, mean accuracy is 84.5% versus 79.2% for fixed prefixes. Adaptive inference also requires fewer concepts to reach accuracy thresholds compared to the full model, for example using ~32% fewer concepts to reach 98% of the full model's performance. We added the detailed results in the main text under RQ3 and App. D.12.

Thank you for the response. I will maintain my score of 3.

Thank you for your follow-up and for taking the time to consider our responses and additional experiments.

Reviewer_LJJ7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces AnyBottle, a method for constructing task-specific CBMs from a large pool of concepts extracted using sparse autoencoders (SAEs). The method iteratively selects features based on how well they can explain a black-box teacher. It then trains a classifier that can use a different number of selected features depending on the confidence of each prediction. Semantic labels are assigned post-hoc to features using MAIA, an automated interpretability agent. The method is evaluated on both vision and text datasets and, more often than not, shows comparable accuracy and higher concept consistency than simple SAE probes and other CBM baselines, while having fewer concepts in the bottleneck.

Critical strengths
  1. The experimental evaluation in the paper is substantial and includes several ablation studies investigating the contribution of the different components of the proposed method.
  2. The method can be applied to both vision and text classification tasks, which is something I have not often seen in CBM papers.
Critical weaknesses
  1. The paper proposes selecting task-specific concepts from a large SAE trained on a general-purpose vision backbone, but it is not clear why this approach should be preferred over training an SAE directly on the representations of a task-specific backbone. Related work that does this (e.g., [1]) is not discussed in the paper despite addressing roughly the same problem and having very close motivation. I think this may be very relevant for a dataset whose inputs are out-of-distribution compared to the dataset used to pretrain the general backbone. Since no comparison with this alternative is provided, it remains difficult to position the paper on this aspect.

  2. The method uses SAE features directly as a concept bottleneck, and assigns semantic labels only post-hoc using MAIA. However, these labels are only hypotheses for what a feature represents. This creates an information leakage (i.e., when predictions depend on information that is not explained by concept semantics) risk since the model would achieve exactly the same accuracy even if these labels were random. In this regard, the method is more similar to a black-box sparse model that requires post-hoc interpretability than an interpretable-by-design CBM, where a bottleneck is added to constrain predictions to depend on human-interpretable concepts.

[1] De Santis et al., Learning Concept Bottleneck Models from Mechanistic Explanations, ICLR 2026.

Questions for authors

Why use a CLIP-based backbone if, as I understand it, the paper never leverages language alignment, for instance to interpret the SAE features? Maybe this alignment could be useful for evaluating the semantic labels assigned by MAIA.

5 comments
AuthorsOpenReview ↗

Individual Response

W1. Task-specific backbones and concept pools

We thank the reviewer for highlighting De Santis et al. [1]. We added this work and evaluated both M-CBM’s task-specific SAE pool and its VLM-processed pool across all six vision datasets (see “M-CBM Comparison” in our general response).

AnyBottle requires the backbone and pool to remain fixed during selection, but either can previously have been adapted to the task. The new results demonstrate the benefit: using M-CBM’s task-specific SAE pool improves AnyBottle’s accuracy from 58.9% to 76.72% on ISIC and from 79.7% to 83.73% on Dogs.

Selection remains useful within these pools. Compared with a probe on the same pruned SAE features, AnyBottle uses 1.2–5.3× fewer concepts and improves consistency on five datasets, with dataset-dependent accuracy costs. Thus, task-specific representation learning and selection can be combined. These results show that task-specific training and selection work together: adapting the backbone can improve the available concepts, and AnyBottle can then select a compact subset from that improved pool.

W2. Semantic faithfulness and information leakage

We agree that accuracy alone cannot validate a concept’s name or exclude semantic leakage. However, unchanged accuracy under random renaming in bottom-up CBMs does not demonstrate leakage: the computation is unchanged because names describe the features rather than determine them. Specifically, in bottom-up CBM approaches (see [2]) features are discovered and selected before their semantics are described. Predictions pass through these features by construction, while the correspondence between their activations and human-readable descriptions requires verification. We agree making this distinction more clear in the paper is important and have updated the text accordingly in the related works.

To clarify, MAIA contributes to this verification through hypothesis-driven experiments. It generates or edits inputs and measures feature responses to test candidate interpretations, rather than simply assigning plausible names to attribution maps. C² and AutoInterpScore provide further evidence of coherence and descriptive predictiveness. Our new shortcut experiment gives a concrete example (see general response, "Shortcut Debugging"): the border-sensitive concept is selected first, and removing it followed by head refitting reduces the student’s reliance on the shortcut while retaining performance on unaffected samples. Further, the discrete-bottleneck ablation (Tab 4 (left)) also removes activation magnitude as a possible leakage channel, without excluding unintended information in binary activations.

Overall, and as stated in Section 3.4, names should remain open to testing and refinement through expert interaction, whether initially proposed by MAIA, CLIP, or a human. Thus, we acknowledge leakage as a risk, but distinguish that risk from evidence that the selected features are uninterpretable.

[2] Knab, Patrick, et al. "What’s in the bottle? a survey and roadmap of concept bottleneck models." Transactions on Machine Learning Research (2026).

Q1. Backbone choice and CLIP naming

We chose CLIP-DINOiser because pretrained backbone and SAE checkpoints were available [3], and its patch-level features support localized selection. Correct, AnyBottle does not require language alignment: selection uses concept activations and teacher behavior without text embeddings, as also demonstrated by our task-specific encoder analysis.

Backbone choice and naming procedure are separate decisions. We use MAIA because it tests candidate descriptions through generated or edited inputs and measured feature responses, rather than relying solely on similarity to a fixed vocabulary.

Following the reviewer’s suggestion, we added a CLIP-based comparison of MAIA descriptions and CFM’s decoder-based names [3] on the same selected features (App. D.13). MAIA achieves higher mean image–description alignment on all six datasets:

DatasetMAIADecoder-based naming
Imagenette0.05760.0298
Dogs-100.04450.0010
Machines-150.04250.0017
Mixed-200.05830.0180
Flowers-1020.07070.0000
ISIC-70.0986−0.0270

This provides additional evidence for the visual relevance of MAIA’s descriptions and shows how CLIP alignment can complement its feature-response experiments.

[3] Wittenmayer et al. “CFM: Language-Aligned Concept Foundation Model for Vision.” ECCV, 2026.

Dear Reviewer,

Thank you again for your thoughtful feedback. We hope our responses and additional experiments have addressed your concerns. If any points remain unclear or unresolved, we would appreciate the opportunity to discuss them further.

I thank the authors for the response and the additional experiments. Furthermore, regarding Q1, I thank you for the clarification and the additional CLIP experiment. I consider this question addressed.

Regarding W1, the substantial accuracy improvement when using task-specific SAE features supports my original concern about the limitations of general-purpose bottlenecks. I am also not fully convinced that the additional selection procedure is necessary when starting from a task-specific concept pool. As I understand, one of the main motivations of AnyBottle is to reduce a large concept vocabulary, but task-specific methods can already produce much smaller vocabularies. Furthermore, I think the reported reduction of 1.2-5.3× would benefit from being put in the context of a baseline that prunes concepts with zero or close-to-zero classifier weights.

Regarding W2, my concern relates to the fact that an SAE feature may activate on patterns that are not captured by its description, and the classifier may exploit these activations for prediction. For instance, a feature labeled "cat" may also activate on certain species of dogs in some contexts, providing information useful for predicting the dog class despite this not being reflected in its description. The information leakage in this case has little to do with whether activations are binary or continuous. One way to measure the effect of this leakage would be to suppress feature activations whenever their assigned semantic concepts are absent, activate features whenever their assigned semantic concepts are present but the features do not activate by themselves, and see how much accuracy is retained.

Overall, my two critical weaknesses still remain, therefore I am maintaining my score

Comments on the paper

5 comments
AuthorsOpenReview ↗

General Response 1

We thank the reviewers for their time and feedback. We appreciate their collective recognition of the effectiveness of our approach (RLJJ7, Rjyrz, RZww5), the breadth of evaluations across domains (RLJJ7), the relevant ablations (Rjyrz, RLJJ7), and the clarity of the presentation (RLJJ7).

We have added experiments on runtime, shortcut debugging, task-specific backbones and SAEs, smaller and group-lasso-regularized SAEs, selection on a shared dictionary, adaptive inference at matched budgets, and concept naming. These results strengthen the evaluation and clarify the contribution. We summarize them below and address specific concerns in each reviewer response.

Representation Fine-tuning and Concept Selection (RLJJ7, RZww5)

Fine-tuning a representation and selecting concepts address different choices: what features to learn, and which of the available features to retain for a task. AnyBottle already combines task-aware SAE fine-tuning with selection. The question is therefore whether adapting or compressing the representation makes the subsequent selection step unnecessary.

Our new experiments show that learning a compact representation does not make selection redundant: AnyBottle further compresses GLR-trained pools by 1.3–3.5× and M-CBM’s pruned SAE pools by 1.2–5.3×, often improving concept consistency. It also benefits from stronger task-specific representations on ISIC and Dogs. The smaller-SAE results further show that a small dictionary is not equivalent to a small subset of a richer decomposition. Since accuracy, consistency, coverage, and granularity can favor different pools, AnyBottle allows practitioners to choose or adapt a compatible pool, then select a compact task-relevant subset without changing its features. Representation improvements can therefore be incorporated into AnyBottle rather than treated as alternatives to selection (App. D.8–D.12). Keeping the pool fixed also makes feature inclusion traceable. In our shortcut experiment, the border-sensitive feature enters first in every seed and can subsequently be inspected and removed without retraining the backbone or SAE.

Our focus in this work is on selecting from a supplied concept pool, rather than learning an optimal pool. As stated in the paper and demonstrated by the new experiments, AnyBottle is not tied to a particular concept source. The same-pool comparisons and selection ablations show what it contributes beyond representation learning alone.

Computational Cost (Rjyrz, RZww5)

We added a breakdown of computational cost in App. D.14 (summary below), measured on Imagenette using one RTX 4500 Ada GPU. Both methods use released SAE checkpoints, so external pretraining is excluded.

StageAnyBottleDN-CBM
Feature extraction and caching8.8 min48 s
Black-box teacher training70 s30 s
Task-aware SAE fine-tuning, including patch teacher10.5 min–
Concept selection / sparse-probe training17 s31 s
Total elapsed time, excluding naming21 min2.1 min
GPU time, excluding naming0.28 GPU-h0.03 GPU-h

Selection takes 17 seconds for 28 rounds, approximately 0.6 seconds per round. The additional construction cost comes mainly from patch-level feature extraction and task-aware SAE fine-tuning. The smaller vocabulary notably reduces the cost of detailed naming. At our measured average MAIA cost of 13.0 minutes per concept, naming approximately 30 AnyBottle concepts takes 6.4 GPU-hours. Applying the same procedure to DN-CBM’s 1242 active concepts would take about 40× longer (269 GPU-h).

The timings show that selection is practical even for the larger bottlenecks, while reducing the vocabulary makes detailed agentic naming substantially more affordable.

AuthorsOpenReview ↗

General Response 2

M-CBM Comparison (RLJJ7)

We thank the reviewer for highlighting De Santis et al. [1]. We added this work and evaluated two comparisons across all six vision datasets: AnyBottle versus a probe on M-CBM’s finetuned SAE pool, and AnyBottle on its VLM-merged features versus M-CBM’s annotation-trained bottleneck. Imagenette illustrates both:

Concept pool + procedureKAccuracy gap (pp)C²img
CLIP-DINOiser MSAE + AnyBottle29.7 ± 1.2+0.5 ± 0.30.330
M-CBM pruned SAE + probe18−0.64 ± 0.330.356
M-CBM pruned SAE + AnyBottle15.0 ± 0.0−0.13 ± 0.230.391
M-CBM, annotation-trained18−1.170.355
M-CBM merged SAE + AnyBottle15.0 ± 0.0−0.13 ± 0.230.391

Gaps use each backbone’s corresponding teacher. Probe and AnyBottle results use three seeds; M-CBM follows its NCC-averaged protocol. Full results and pool definitions are in App. D.8.

As the reviewer suggests, task-specific representations can substantially benefit AnyBottle: accuracy increases from 58.9% to 76.72% on ISIC and from 79.7% to 83.73% on Dogs relative to the CLIP-DINOiser configuration. Within the pruned SAE pools, selection reduces vocabulary size by 1.2–5.3× and improves consistency on five datasets, with dataset-dependent accuracy costs.

However, task-specific construction is not uniformly better: CLIP-DINOiser MSAE + AnyBottle achieves higher absolute accuracy than the evaluated M-CBM pipeline on four of six datasets. On Flowers, it reaches 93.3% versus 86.91%, using 125 rather than 202 concepts with higher image-level consistency (0.15 versus 0.088).

AnyBottle can also operate on the VLM-merged pools without requiring concept annotations to train a new bottleneck. E.g., compared with full M-CBM, it improves accuracy and consistency on Imagenette and Machines.

Overall, these results show that task-specific representation learning and selection can be combined. They also distinguish VLM-assisted pool construction from naming: AnyBottle can use either SAE pool, while naming in our standard pipeline occurs after training and does not affect predictions.

[1] De Santis et al. “Learning Concept Bottleneck Models from Mechanistic Explanations.” ICLR, 2026.

Smaller SAEs (RZww5)

We expanded this comparison to all six vision datasets, including 32-latent SAEs pretrained on full CC12M and 32-/256-latent SAEs pretrained on a 200k-image CC12M subset. All receive task-aware fine-tuning. We also apply AnyBottle to the subset-pretrained pools. Imagenette illustrates the two comparisons:

Method / poolAccuracy gap (pp)KC²patch / C²img
MSAE-8192-F + AnyBottle+0.5 ± 0.3300.46 / 0.33
MSAE-32-F + probe−0.7 ± 0.1320.28 / 0.21
MSAE-32-S + probe−0.5 ± 0.3320.32 / 0.23
MSAE-32-S + AnyBottle−0.8 ± 0.426.7 ± 2.50.33 / 0.24
MSAE-256-S + probe−0.6 ± 0.22550.24 / 0.17
MSAE-256-S + AnyBottle+0.4 ± 0.294.7 ± 7.20.27 / 0.19

Accuracy gap is balanced test accuracy minus teacher accuracy. Results use three seeds; the full six-dataset table is in App. D.10.

At comparable deployed budgets, the full-CC12M 32-latent SAE does not recover the accuracy or consistency of AnyBottle’s 30 selected concepts. Across all six datasets, our original configuration has higher patch-level consistency than either 32-latent variant. This is consistent with the motivation for larger dictionaries: separating patterns into more specific features before selecting a compact subset.

AnyBottle also benefits smaller pools. On Imagenette’s 256-latent pool, it selects 95 of 255 active features while improving accuracy and both consistency scores. Across datasets, it compresses these pools by approximately 1.2–3.6×, with dataset-dependent accuracy and consistency trade-offs.

These results support both the value of a rich candidate pool and AnyBottle’s applicability to alternative pools. Selection can improve how an existing pool is used, but cannot introduce distinctions absent from its features.

AuthorsOpenReview ↗

General Response 3

Group-lasso SAEs (RZww5)

We evaluated group-lasso regularization (GLR) during task-aware SAE fine-tuning and applied AnyBottle to the resulting pools. From 13 regularization strengths, we report the configuration closest to original AnyBottle’s vocabulary size. Retained features are fixed using training data and excluded features are masked during probe training and evaluation. Full results are in App. D.11; Imagenette illustrates the comparison:

MethodK ↓Accuracy gap (pp) ↑C²patch ↑C²img ↑
Task-aware MSAE + AnyBottle30+0.50.460.33
GLR-MSAE + probe36+0.10.420.33
GLR-MSAE + AnyBottle14.3+0.10.450.36

Accuracy gap is balanced test accuracy minus teacher accuracy. Means across three seeds; full protocol and uncertainty are reported in the appendix.

GLR can in principle produce useful compact pools, but not a guaranteed. It improves accuracy on Dogs-10 with lower consistency, while its higher consistency on Flowers accompanies lower accuracy.

Importantly, we observe that selection remains useful also on a concept pool resulting from such regularization. AnyBottle further compresses every GLR pool by approximately 1.3–3.5×. On Imagenette it improves consistency without losing accuracy; on Machines-15 and Mixed-20 it improves both consistency scores at accuracy costs of 0.4 and 0.5 points. On ISIC, it reduces 500 features to 142 and improves accuracy by 3.3 points. Further compression costs more accuracy on Dogs and Flowers.

These results support combining adaptation and selection. A pool may be preferable for its accuracy, semantic consistency, coverage, or granularity; a smaller vocabulary alone does not settle that choice. Our contribution is not a competing recipe for the best concept pool, but a procedure for selecting a compact, task-relevant subset from a supplied, compatible pool. The GLR + AnyBottle results demonstrate this directly: improvements in pool construction can be incorporated into AnyBottle rather than treated as alternatives that make selection unnecessary.

Quantitative Adaptive Inference (Rjyrz)

We added a comparison of adaptive and fixed-prefix inference across all six vision datasets and three seeds (RQ3 and App. D.12). Both approaches use the same model and concept order. Fixed-prefix inference gives every input the same number of concepts; adaptive inference stops early on confident inputs and continues on others.

At half the available concepts per input on average, adaptive inference improves balanced accuracy on every dataset:

DatasetAdaptive accuracy (%)Fixed-prefix accuracy (%)
Imagenette98.090.4
Dogs-1074.567.1
Machines-1594.391.9
Mixed-2097.293.9
Flowers-10285.575.3
ISIC-757.956.6
Mean84.579.2

Adaptive inference also requires fewer concepts to reach a target accuracy: average relative savings are 29%, 29%, and 32% at 90%, 95%, and 98% of full-budget accuracy, respectively. A separate analysis shows that harder inputs consume more active concepts, measured against the teacher’s true-class probability (Spearman ρ from −0.33 to −0.65).

The appendix reports the full curves, validation-selected operating points, and difficulty analysis. Together, these results show higher accuracy at a given average budget and fewer concepts needed for a given accuracy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.