acceptodds
Under review as a conference paper at ICLR 2027

Don't Mind the Gap: Interpretable Vision-Language Models via Independent Sparse Autoencoders

Abstract

Contrastive Vision-Language Models (VLMs) produce general-purpose, multi-modal representations, but their black-box nature hinders safe and ethical deployment. Sparse Autoencoders (SAEs) can disentangle these dense representations into interpretable features, but the application of a single SAE across modalities creates a split-dictionary phenomenon, where learned features activate for a single modality, preventing the discovery of shared features. Current approaches attempt to bridge this gap using a co-activation loss, but we find that forcing a single shared SAE inherently compromises reconstruction quality, individual modality interpretability, and zero-shot performance. Instead, we propose decoupling the SAEs by modality while retaining the co-activation guidance. This dual architecture discovers features that retain their modality-specific nature while being synchronized to enable consistent cross-modal mapping. Our approach achieves superior interpretability and downstream zero-shot performance, especially in challenging high-to-medium sparsity regimes. Our findings indicate that exploiting the modality gap, rather than collapsing it, is the key to interpreting contrastive VLMs.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, sparse autoencoders, contrastive VLMs

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

This paper studies the interpretability of contrastive vision-language models (VLMs) using sparse autoencoders (SAEs). The authors identify the split-dictionary phenomenon, where a shared SAE tends to learn modality-specific features rather than discovering shared cross-modal features. They argue that enforcing a single shared SAE can compromise reconstruction quality, interpretability, and zero-shot performance. To address this issue, they propose a dual-SAE architecture that uses separate SAEs for each modality while retaining co-activation guidance to facilitate cross-modal feature alignment.

Critical strengths
  1. The authors analyze the activation properties of shared and independent SAEs.
  2. The authors introduce a permutation matrix to encourage SAEs from different modalities to learn similar concepts.
  3. The authors conduct experiments on multiple CLIP-style models, such as CLIP and SigLIP.
Critical weaknesses
  1. The authors only consider conventional shared SAEs, which use a shared encoder and decoder. They overlook an important prior work, VL-SAE [1], which uses a modality-shared encoder to construct concepts and independent decoders to learn modality-specific information. This work appears to have already addressed the problem raised by the authors.
  2. Although the authors use a permutation matrix to improve concept alignment across modalities, they lack a corresponding evaluation of cross-modal concept consistency. For example, they could measure the CLIP similarity between images and text associated with the same activated concept, as done in VL-SAE.
  3. As shown in Figure 7, the proposed method still sacrifices model performance for interpretability, unlike works such as VL-SAE and TORINO [2], which improve model performance alongside interpretability.
  4. The proposed method relies heavily on CLIP's dual-tower alignment architecture and therefore cannot be directly applied to MLLMs like LLaVA and Qwen-VL, which are more capable vision-language models.

[1] Shen, S., Sun, J., Huang, Q., & Wang, S. (2026). Vl-sae: Interpreting and enhancing vision-language alignment with a unified concept set. Advances in Neural Information Processing Systems, 38, 45235-45265.

[2] Renzulli, R., Spadaro, G., Gowda, S., Mazouz, A. E., & Nguyen, V. T. (2026). TORINO: Token Reduction via Interpretable Concept Overlap in Vision-Language Models. arXiv preprint arXiv:2607.04593.

Questions for authors

Overall, I encourage the authors to conduct a more thorough review of existing studies on multimodal SAEs to better clarify the novelty of the problem they aim to address. From my perspective, some of the challenges discussed in this paper appear to have been addressed by prior work (e.g., VL-SAE), which also reports promising results. It would therefore be helpful to include more comprehensive comparisons with these methods to better demonstrate the advantages of the proposed approach and provide further insights into the differences in their performance.

4 comments

S2) The authors introduce a permutation matrix to encourage SAEs from different modalities to learn similar concepts.

Regarding the second strength, we would like to clarify that the permutation matrix is not used to encourage SAEs from different modalities to learn similar concepts. The permutation matrix is used once after training to align the already-learned per-modality dictionaries. As the results about the multimodal monosemanticity score show (Figures 6, 13, 22, 26), independent SAEs are able to learn shared semantics without using any inductive factor during training. This property is emergent.

W1, Q1) The authors only consider conventional shared SAEs, which use a shared encoder and decoder. They overlook an important prior work, VL-SAE [1], which uses a modality-shared encoder to construct concepts and independent decoders to learn modality-specific information. This work appears to have already addressed the problem raised by the authors.
Overall, I encourage the authors to conduct a more thorough review of existing studies on multimodal SAEs to better clarify the novelty of the problem they aim to address. From my perspective, some of the challenges discussed in this paper appear to have been addressed by prior work (e.g., VL-SAE), which also reports promising results. It would therefore be helpful to include more comprehensive comparisons with these methods to better demonstrate the advantages of the proposed approach and provide further insights into the differences in their performance.

We thank the reviewer for bringing VL-SAE [7] to our attention. We agree it is an important prior work, and we will certainly include it in the related works section. However, our work is fundamentally different in both goals and scope.

Rather than proposing a new hybrid architecture to maximize downstream task accuracy (as VL-SAE does), our work asks a foundational mechanistic question: "can dense representations of contrastive VLMs be decomposed with a single shared SAE to effectively provide high interpretability while also retaining a faithful representation of the underlying model’s workings, even for modality-specific embeddings?"

By systematically contrasting shared architectures against independent, per-modality SAEs, we demonstrate why explicitly decoupling modalities is essential for faithful interpretability. The key differences that highlight the novelty of our problem setting are:

  • Strict Information Bottleneck vs. Downstream Accuracy: to maximize downstream task accuracy, VL-SAE adds concept-space quantities back into the original activation space (see Equations 10–11 in [7]). While highly effective for boosting predictions, bypassing the SAE bottleneck makes it impossible to isolate how decisions arise purely from the learned concepts, and VL-SAE does not evaluate reconstruction faithfulness to the original activations. In contrast, our work treats the SAE strictly as a white-box information bottleneck, prioritizing faithfulness and interpretable decompositions over downstream task gains;
  • Focus on Contrastive VLMs Geometry: Our focus is on contrastive VLMs, which exhibit the modality gap in their output spaces. While there is compelling evidence that a modality gap appears in the earlier layers of LVLMs and progressively disappears towards their final layers [1], our aim is to analyze the reasons why a single s-SAE across modalities cannot faithfully interpret contrastive VLMs. As s-SAEs are structurally biased for extracting bimodal concepts living in an intermediate subspace between the modality subspaces, we inspect what are the implications of this choice for their faithfulness to the original model activations, the interpretability of the learned concepts and the downstream performance, both zero-shot and cross-modal steering;
  • Operating Regimes for Mechanistic Interpretability: Because VL-SAE is designed to optimize downstream performance, its experiments focus on a fixed expansion ratio (k=8×dk = 8 \times d) and lower sparsity regimes (K≥256K \ge 256 in our notation). However, as established in recent interpretability literature [2] and confirmed across our experiments (Figures 6, 13, 22, 26), lower sparsity substantially degrades feature monosemanticity. To rigorously study contrastive VLMs using mechanistic interpretability, we explore a broad range of models, datasets, expansion ratios, and high-sparsity regimes by using the toolkits already developed in the literature. Furthermore, rather than evaluating only top-activating multimodal samples (on the order of hundreds), we evaluate interpretability across all activating examples (simultaneously) for each neuron in both unimodal [2] and multimodal [3] settings.

Taken together, our geometric analysis, faithfulness evaluations, and high-sparsity interpretability benchmarks offer complementary insights to VL-SAE's downstream-oriented framework, addressing essential questions for multimodal interpretability.

W2) Although the authors use a permutation matrix to improve concept alignment across modalities, they lack a corresponding evaluation of cross-modal concept consistency. For example, they could measure the CLIP similarity between images and text associated with the same activated concept, as done in VL-SAE.

The paper already includes experiments for assessing the cross-modal concept consistency. The multimodal monosemanticity score (introduced in [3] and detailed in Section 3.1 and Appendix D) specifically assesses if each neuron is activated for a single cross-modal concept:

  • How it works: It is computed by collecting all image and text examples that activate a given neuron, measuring the pairwise similarity between these examples using an external model (CLIP), and weighting each similarity by the corresponding code activation strength.
  • What it proves: Because this score is maximized only when activations correspond to image-text pairs with high CLIP similarity, it directly quantifies cross-modal semantic consistency, going beyond top-kk samples (as done in VL-SAE) to evaluate the entire activating distribution.

As shown in Figures 6, 13, 22, and 26, i-SAEs with Hungarian Matching consistently outperform s-SAEs on this metric, confirming stronger cross-modal concept alignment.

W3) As shown in Figure 7, the proposed method still sacrifices model performance for interpretability, unlike works such as VL-SAE and TORINO, which improve model performance alongside interpretability.

As noted in (W1, Q1) regarding VL-SAE, their approach aggregates SAE outputs with the original dense activations (see Equations. 10 and 11 in [7]). While this effectively boosts task accuracy, it bypasses the SAE bottleneck, allowing the model to still rely on un-decomposed dense features. Similarly, TORINO [8] addresses the different problem of decreasing the number of visual tokens in input to VLMs, thus increasing their efficiency while minimizing performance loss.

In contrast, to rigorously test the faithfulness of the learned concepts, we enforce a strict information bottleneck where contrastive VLM predictions are computed entirely from the reconstructed SAE features. Under pure sparse compression, an accuracy trade-off relative to the uncompressed model is expected.
Crucially, Equation 6 (Section 3.1) factorizes exactly why preserving accuracy under this strict bottleneck is challenging. Each cosine similarity prediction factorizes into a cross-modal co-activation and alignment term, divided by intra-modal interference. To achieve good downstream performance without any shortcut, the SAE decompositions must satisfy three strict geometric conditions:

  • Both modalities must preserve fine-grained features within their top-kk sparse selections;
  • Textual and visual top-kk selections must remain perfectly aligned (losing an important feature in either modality degrades the final logits);
  • Intra-modal interference must be minimized to prevent logit distortion.

Viewed through this lens, Figure 7 demonstrates that Independent SAEs retain substantially better downstream performance than Shared SAEs under a true white-box bottleneck. They achieve this precisely because respecting the modality gap better satisfies these three geometric constraints.

W4) The proposed method relies heavily on CLIP's dual-tower alignment architecture and therefore cannot be directly applied to MLLMs like LLaVA and Qwen-VL, which are more capable vision-language models.

The paper specifically tackles the problem of the split-dictionary phenomenon observed in contrastive VLMs output spaces and proposes a solution which exploits the geometry of those spaces by highlighting the limitations of current ones. We acknowledge that expanding the results to multimodal LLMs is a promising research direction, but it is an entirely different scope from what we consider in this work, and we leave it for future developments. Additionally, we stress that contrastive VLMs are still extensively used as the primary visual encoders or text-image alignment modules for multiple applications, including image generation [4], composed retrieval [5], image captioning [6]. Therefore, solving the interpretability challenges specific to dual-tower architectures is a critical contribution in its own right. We believe our findings will directly benefit the interpretability, safety, and reliability of the downstream applications that continue to rely heavily on contrastive VLM backbones.

References

[1] Hyeon-Woo, N., Ye-Bin, M., Lim, S., Byung-Ki, K., & Oh, T. H. Zero-Shot Rankability: Revealing Latent Ordinal Structure in Multimodal Large Language Models via Language. In Forty-third International Conference on Machine Learning, 2026.
[2] Mateusz Pach, Shyamgopal Karthik, Quentin Bouniot, Serge Belongie, and Zeynep Akata. Sparse autoencoders learn monosemantic features in vision-language models. In The Thirty- ninth Annual Conference on Neural Information Processing Systems, 2025.
[3] Chiraag Kaushik, Davis Barch, and Andrea Fanelli. Decomposing multimodal embedding spaces with group-sparse autoencoders. In The Fourteenth International Conference on Learning Representations, 2026.
[4] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022.
[5] Bill Psomas, George Retsinas, Nikos Efthymiadis, Panagiotis Filntisis, Yannis Avrithis, Petros Maragos, Ondrej Chum, and Giorgos Tolias. Instance-level composed image retrieval. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025.
[6] Kumar Ashutosh, Yossi Gandelsman, Xinlei Chen, Ishan Misra, and Rohit Girdhar. LLMs can see and hear without any training. In Forty-second International Conference on Machine Learning, 2025.
[7] Shen, S., Sun, J., Huang, Q., & Wang, S. (2026). Vl-sae: Interpreting and enhancing vision-language alignment with a unified concept set. Advances in Neural Information Processing Systems, 38, 45235-45265.
[8] Renzulli, R., Spadaro, G., Gowda, S., Mazouz, A. E., & Nguyen, V. T. (2026). TORINO: Token Reduction via Interpretable Concept Overlap in Vision-Language Models. arXiv preprint arXiv:2607.04593.

Thank you to the authors for their response. First, I encourage the authors to incorporate the rebuttal clarifications and promised improvements into the revised manuscript. One advantage of the ICLR review process is that authors can update their submissions, allowing all reviewers to assess the actual improvements to the paper. Second, I understand that the author focus on metrics like faithfulness rather than the performance improvement in downstream tasks. However, it is also necessary to conduct experimental comparisons between the proposed method and existing methods in the indicators that the author focused on.

Reviewer_va9cOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses a limitation of training a single SAE jointly on both modalities of CLIP-like models, which learns both unimodal and cross-modal features inefficiently. The authors instead propose to train a separate SAE for each modality and match their cross-modal features through a mapping function. This improves interpretability and reconstruction metrics on both unimodal and cross-modal tasks.

Critical strengths

(S1) The motivation rests on a sound observation, supported by prior work [1] and Figure 3. An SAE trained jointly on both CLIP modalities may be suboptimal, because some features are specific to one modality and unused by the other. This justifies training a separate SAE for each modality and then matching their cross-modal features. The approach is also thoroughly grounded in prior work.

(S2) The paper contains extensive quantitative evaluations on reconstruction, monosemanticity, and zero-shot classification (Figures 4–6). The proposed method improves over the baselines, particularly at higher sparsity levels and on cross-modal metrics. The appendix extends this with a detailed evaluation across additional datasets and SAE expansion factors.

(S3) Qualitative evidence and downstream applications. Cross-modal steering (L. 420, Table 1) and Modality-specific features activating for the same concept.(L. 452, Figure 9) provide visual evidence and a downstream application of the SAEs. They also add causal evidence that complements the quantitative results.

Critical weaknesses

(W1) The method shows limited novelty, as both loss functions are taken from prior work, and the method itself amounts to training a single SAE per modality and matching features across SAEs. A more interesting topic concerns whether we can train a single SAE on multiple modalities/representations, for which prior work [1,2,3] has already presented evidence that this can be done.

(W2) One-to-one matching is questionable. Hungarian matching enforces a one-to-one correspondence between features. SAE atoms often suffer from feature splitting [4], so a concept in one modality may be spread across several features in the other, which suggests Hungarian matching may be suboptimal.

(W3) Results for SAEs with the same number of neurons as the single SAE appear only in Appendix F.2.2, and only for some experiments. The main paper results compare models with s-SAE having double the number of neurons. They are therefore not directly comparable, and the main-paper presentation is potentially misleading.

(W4) The equations in Section 2 are mostly inline, and the SAE notation does not follow prior work. As a result, it is unclear whether the single SAE uses a per-modality bias, as in [1].

(W5) The paper contains several formatting issues:

  • Citations should use \citep{} or \citet{} accordingly (Lines: L. 102 and L. 185 show examples of incorrect formatting).
  • Appendices A, B, C, E, and G and Table 1 are never referenced in the main text.
  • Related work (App. A) and limitations (App. B) are placed entirely in the appendix. At least parts of them belong in the main paper.

Minor Weaknesses

  • (W6) Most results are presented only as figures. In Figure 4 the s-SAE is barely visible, and Figure 6 is hard to read. Both should be accompanied by tables. No results include standard deviations.
  • (W7) The authors state in Appendix G that The independent SAEs, particularly those with the additional co-activation loss, consistently demonstrate superior autointerpretability performance compared to the shared SAE variants. (L. 2558). The results don’t support this: (a) Results for the group-sparse (co-activation) loss are not reported. (b) On the simulation score, independent SAEs outperform the shared SAE for only 3 of 6 values of KK on CC3M and 2 of 6 on YFCC15M.
  • (W8) Possibly incorrect attribution, as on L. 205 it cites SAE-Bench [5] in the context of limitations of shared cross-modal decomposition. As far as I can tell, [5] does not discuss this. The authors should verify the citation.
  • (W9) KK denotes the number of active neurons in the SAE and kk the number of neurons. Different letters should be used for one of them, as this notation is confusing.

References

[1] Kaushik, Chiraag, Davis Barch, and Andrea Fanelli. "Learning multimodal dictionary decompositions with group-sparse autoencoders." The Fourteenth International Conference on Learning Representations. 2026.

[2] Dhimoïla, Grégoire, et al. "Cross-modal redundancy and the geometry of vision-language embeddings." arXiv preprint arXiv:2602.06218 (2026).

[3] Thasarathan, Harrish, et al. "Universal sparse autoencoders: Interpretable cross-model concept alignment." arXiv preprint arXiv:2502.03714 (2025). (ICML 2025)

[4] Chanin, David, et al. "A is for absorption: Studying feature splitting and absorption in sparse autoencoders." Advances in Neural Information Processing Systems 38 (2026): 82318-82355.

[5] Karvonen, Adam, et al. "SAEBench: A comprehensive benchmark for sparse autoencoders in language model interpretability." arXiv preprint arXiv:2503.09532 (2025).

Questions for authors

(Q1) Following up on (W2): Hungarian matching computes a one-to-one mapping between the features of the two SAEs. Prior work on feature splitting [4] and the bridge score [2], however, shows that SAE features often have many-to-many relationships. How do the authors justify one-to-one matching in this setting? Have they considered many-to-many alternatives?

(Q2) The paper compares two auxiliary losses, but the experiments don’t show a clear winner, and the paper doesn’t discuss the comparison. Which loss do the authors recommend, and under what conditions?

(Q3) Appendix G reports an accuracy metric alongside the simulation score. Is this metric taken from an existing prior work on autointerpretability? If so, which one?

9 comments

W1) The method shows limited novelty, as both loss functions are taken from prior work, and the method itself amounts to training a single SAE per modality and matching features across SAEs. A more interesting topic concerns whether we can train a single SAE on multiple modalities/representations, for which prior work [1,2,3] has already presented evidence that this can be done.

We agree that our architectural approach is technically straightforward, and we acknowledge that prior work [1,2,3,4] has investigated using a single s-SAE across multiple modalities. However, our primary novelty is conceptual and empirical rather than architectural: we demonstrate that the use of a single s-SAE is fundamentally misaligned with the geometric structure of contrastive VLMs.

While training a single SAE might seem appealing, it encounters severe structural limitations due to the modality gap. Efforts to bypass this gap using a single s-SAE require significant compromises, such as restricting the dictionary to only bimodal atoms [2] or using a split-dictionary [1]. Because a single s-SAE is forced to represent a bimodal concept using a single atom in an intermediate subspace, it struggles to capture the distinct geometry of each modality, sacrificing both faithfulness to the original model activations and interpretability.

In light of these limitations, our core contribution lies in a fundamental change of perspective. Rather than attempting to engineer away the modality gap, we propose embracing the inherent geometric structure of the VLM. By explicitly introducing a geometric prior instantiated through per-modality SAEs combined with Hungarian Matching, we allow bimodal atoms to live natively in their respective subspaces. In doing so, we provide the following clear scientific contributions:

  • Identifying the failure of shared architectures: we demonstrate that s-SAEs fail to provide faithful interpretations, highlighting the severe implications of architectural biases that ignore the modality gap;
  • Establishing the necessity of a geometric prior: we show that respecting the geometric structure of representation spaces is a fundamental prerequisite for faithful interpretability;
  • Conducting a focused study on interpretability and faithfulness metrics: unlike previous works that largely neglect these aspects, we provide a rigorous and comprehensive evaluation over multiple VLMs, datasets, sparsity levels and expansion ratios. Through our experiments, we demonstrate that a conceptually simple approach, such as i-SAEs without any co-activation loss, can significantly outperform existing, more complex solutions.

While our resulting architecture is simple, we view this as a significant strength. An easily implementable solution that rigorously respects the model's geometry is strictly more practical and effective than an overly engineered one. The fact that this conceptually simple approach significantly outperforms existing paradigms is exactly the core message of our paper. We believe our work establishes a solid foundation for representation-level interpretability in this setting and serves as a useful basis for the community to build upon.

W2, Q1) One-to-one matching is questionable. Hungarian matching enforces a one-to-one correspondence between features. SAE atoms often suffer from feature splitting [4], so a concept in one modality may be spread across several features in the other, which suggests Hungarian matching may be suboptimal. Hungarian matching computes a one-to-one mapping between the features of the two SAEs. Prior work on feature splitting [4] and the bridge score [2], however, shows that SAE features often have many-to-many relationships. How do the authors justify one-to-one matching in this setting? Have they considered many-to-many alternatives?

We thank the reviewer for this insightful observation. We completely agree that SAEs can suffer from feature splitting and that exploring many-to-many mapping alternatives, for example by applying the Sinkhorn algorithm for solving the Optimal Transport between dictionaries, is a highly interesting and valuable next step. However, before introducing that level of complexity, our primary goal in this work is to establish a solid, rigorous foundation for representation-level interpretability in contrastive VLMs. In this setting, one-to-one matching was a deliberate methodological choice rather than a limitation, primarily to ensure a strictly fair comparison. By design, s-SAEs inherently enforce a one-to-one mapping across modalities because a bimodal concept is forced into a single shared atom. If we allowed many-to-many relationships for our independent SAEs, we would grant them significantly higher expressive power. This would potentially bias the evaluation in our favor, making it impossible to disentangle whether the improvements come from our core claim, that is respecting the modality gap, or simply from having greater mapping flexibility. Enforcing a one-to-one mapping via the Hungarian algorithm ensures an "apples-to-apples" comparison, proving that simply respecting the geometric structure is enough to outperform s-SAEs.

Furthermore, it is important to highlight that our evaluation already accounts for the many-to-many nature among cross-modal features. Specifically, the trace ratio metric used in our paper explicitly relies on the bridge score, which inherently captures many-to-many relationships. By evaluating this ratio, we quantify the amount of cross-modal interaction energy that is successfully concentrated within one-to-one feature mappings compared to the global interactions. Therefore, rather than completely ignoring the many-to-many dynamics, we actively use them to measure and validate the efficacy of our one-to-one assignments from a statistical and geometric perspective.

W3) Results for SAEs with the same number of neurons as the single SAE appear only in Appendix F.2.2, and only for some experiments. The main paper results compare models with s-SAE having double the number of neurons. They are therefore not directly comparable, and the main-paper presentation is potentially misleading.

The size of a dictionary dictates the upper bound of disentangled, monosemantic concepts an SAE can encode. We used the same size kk for all dictionaries to ensure a fair comparison based on this conceptual capacity, rather than strict parameter count. A s-SAE assumes one atom can express a bimodal concept, meaning a dictionary of size kk can represent up to kk bimodal concepts. Because the modality gap induces i-SAEs to use an atom pair (one image, one text) per cross-modal concept, each independent dictionary must also be size kk to reach that exact same upper bound of kk bimodal concepts. Halving the independent dictionaries just to match the s-SAE's total parameter count would artificially handicap them to k/2k/2 bimodal concepts. Thus, matching the dictionary size provides the fairest baseline to compare their representational quality.

Furthermore, we have included the experiments in Appendix F.2.2 to confirm that our gains stem from resolving the modality gap rather than merely doubling the parameter count. The experiments show i-SAEs have better performance compared to the s-SAEs, and the gains still remain by varying the dictionary size in Appendix F.2.1.

W4) The equations in Section 2 are mostly inline, and the SAE notation does not follow prior work. As a result, it is unclear whether the single SAE uses a per-modality bias, as in [1].

We apologize if the paper’s equations were difficult to understand because of the notation and/or space. To the best of our knowledge there is no standard notation used in the literature for SAE papers on multimodal representations, as their applications in this setting is relatively new. We have made an effort to standardize the notation and structure of previous works [1,2,3,4] and open-source toolboxes (https://github.com/KempnerInstitute/overcomplete). As a result, for all experiments the decoder does not use any bias while the encoder does.

W5) The paper contains several formatting issues:

  • Citations should use \citep{} or \citet{} accordingly (Lines: L. 102 and L. 185 show examples of incorrect formatting).
  • Appendices A, B, C, E, and G and Table 1 are never referenced in the main text.
  • Related work (App. A) and limitations (App. B) are placed entirely in the appendix. At least parts of them belong in the main paper.

We apologize for the formatting issues and proceed with correcting them in the final version of the paper.

W6) Most results are presented only as figures. In Figure 4 the s-SAE is barely visible, and Figure 6 is hard to read. Both should be accompanied by tables. No results include standard deviations.

We have decided to exclude the tables as the paper is already very long. The same reason we deleted other qualitative examples. However, if requested, we can include them as well. Regarding the standard deviations, we decided to remove them as they make the plots harder to read. The current plots instead are clearer and allow us to analyze the overall trend of the various metrics across multiple datasets and contrastive VLM models more easily.

W7) The authors state in Appendix G that The independent SAEs, particularly those with the additional co-activation loss, consistently demonstrate superior autointerpretability performance compared to the shared SAE variants. (L. 2558). The results don’t support this: (a) Results for the group-sparse (co-activation) loss are not reported. (b) On the simulation score, independent SAEs outperform the shared SAE for only 3 of 6 values of on CC3M and 2 of 6 on YFCC15M.

Throughout the text, we use the monosemanticity scores (MS) and multi-modal monosemanticity scores (MMS) as our main interpretability metrics. We have still added the auto-interpretability scores in Appendix G to understand if the features learned by SAEs are interpretable in the sense that an LLM can provide a human-readable explanation of what a feature represents. As shown in Table 25, i-SAEs generally outperform s-SAEs on average, both overall and when using the InfoNCE loss.

However, we do not consider auto-interpretability scores as an objective metric for comparing SAE architectures. This pipeline evaluates only the textual modality, depends heavily on the underlying LLM, and relies on a stochastic stratified sampling of activating examples. Consequently, it inherently exhibits high variance in both detection and simulation scores, serving merely as a qualitative, directional indicator rather than a definitive measure.

Finally, regarding point (a), since the group-sparse loss already performed poorly on our primary MS and MMS metrics, and given the high computational cost of running multiple LLM inferences, we decided to exclude it from these additional experiments.

W8) Possibly incorrect attribution, as on L. 205 it cites SAE-Bench [5] in the context of limitations of shared cross-modal decomposition. As far as I can tell, [5] does not discuss this. The authors should verify the citation.

The SAE-Bench [6] citation appears in L. 205 in the context of describing the structural limitation of s-SAEs. The SAE-Bench paper highlights that the reconstruction performance quantifies the SAE faithfulness over the original model's activations for preserving its behaviour (see Sections 3.1 and 3.2.1 of the cited paper, “Reconstruction (fidelity)“). Consequently, the single s-SAE reconstruction failure using bimodal atoms is indicative of unfaithfulness of its decomposition for contrastive VLM output representations. We thus believe that citation is appropriately placed.

W9) K denotes the number of active neurons in the SAE and k the number of neurons. Different letters should be used for one of them, as this notation is confusing.

We apologize and we proceed with changing the notation in the final version of the paper.

Q2) The paper compares two auxiliary losses, but the experiments don’t show a clear winner, and the paper doesn’t discuss the comparison. Which loss do the authors recommend, and under what conditions?

Overall, the results trend show clear advantages for the InfoNCE loss over the Group-Sparse loss for both SAE architectures, we thus recommend the InfoNCE loss as the default choice for extracting bimodal concepts in shared SAEs and for encouraging co-activation in independent SAEs.

Q3) Appendix G reports an accuracy metric alongside the simulation score. Is this metric taken from an existing prior work on autointerpretability? If so, which one?

Both metrics used in Appendix G are taken from [5]. As specified in the text, we use the Simulation and Detection scores. The "Accuracy" column stands for the Detection score, as the metric is computed as the accuracy over the detection predictions.

Reviewer_R7s5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors try to tackle the "split-dictionary" problem that arises when a single sparse autoencoder (SAE) is shared across the visual and textual modalities of a contrastive VLM. Prior work addresses this with a co-activation loss on a single shared SAE, which the authors argue compromises reconstruction quality, per-modality interpretability, and zero-shot performance. Instead, they propose a dual architecture that decouples the SAEs by modality (i-SAEs) while retaining the co-activation guidance, and they match atoms across modalities post hoc via Hungarian matching. The approach is evaluated on four VLMs, two datasets, and a wide range of sparsity levels, and it achieves competitive or better results in reconstruction, interpretability, and downstream tasks such as zero-shot classification and retrieval.

Critical strengths
  1. The paper is generally well written and easy to follow. It provides a broad, systematic study of the current main setups for this task: various VLMs, datasets, sparsity levels, and co-activation design choices are included, giving a broad picture of the current state of the art. I believe this would benefit other researchers in the field.

  2. The negative result on shared bimodal dictionaries is a valuable observation. The analysis showing that the bimodal atoms of SAEs trained with a co-activation loss cannot reconstruct a large share of each modality's activations (Fig. 3) gives concrete evidence that "closing the gap" inside the SAE discards information. This provides solid motivation for the proposed approach.

Critical weaknesses
  1. I am concerned that some of the evaluation metrics are biased in favour of i-SAEs:

    1. For the trace ratio, the permutation P for i-SAEs is chosen to maximize Tr((D_V P)ᵀ D_T), i.e., the geometric factor of the bridge matrix's own diagonal, while the s-SAE is fixed at P = I. The metric is therefore partly optimized for i-SAEs post hoc.
    2. The MS score is computed per modality, and a dictionary trained on a single modality naturally obtains a higher single-modality MS. Since these metrics underpin the claim of better cross-modal interpretability, this bias weakens the paper's main conclusion.
  2. Other concerns:

    1. Matching quality is not quantified, so it is hard to know what fraction of the Hungarian-matched pairs are semantically consistent.
    2. The human study is small (9 participants and 223 responses, recruited from the authors' institution).
    3. The i-SAE has twice the atoms and parameters of the s-SAE in all main results; the equal-parameter comparison is limited to a single model and dataset.
Questions for authors
  1. How is P obtained for the trace ratio: on training or held-out data?
  2. Can you evaluate zero-shot classification and retrieval using only the Hungarian-matched atom pairs?
  3. Please provide equal-parameter results for additional VLMs and datasets.
  4. Per-modality SAEs combined with Hungarian matching seems like an obvious baseline. Could you clarify what is novel about the proposed approach beyond this baseline?
8 comments

Concerns about evaluation metrics

W1.1) For the trace ratio, the permutation P for i-SAEs is chosen to maximize Tr((D_V P)ᵀ D_T), i.e., the geometric factor of the bridge matrix's own diagonal, while the s-SAE is fixed at P = I. The metric is therefore partly optimized for i-SAEs post hoc.

Actually, we argue that the trace ratio metric from eq. (5) positively favors s-SAEs. In fact, since for s-SAEs the text and visual dictionaries are the same, i.e., DV≡DTD_V \equiv D_T, and P=I\mathbf{P}=I, then it results that geometric term is Tr((DVP)TDT)=k\mathrm{Tr}((D_V P)^T D_T) =k, which is the maximum possible. Conversely, for i-SAEs, since the two dictionaries are not ordered in the same way, the alignment term Tr((DVP)TDT)\mathrm{Tr}((D_V P)^T D_T) will in general be always smaller than kk and only in case the permutation matrix PP manages to recover a perfect alignment the same score as s-SAEs is recovered. However, since PP is only optimized post-hoc and not jointly with DVD_V and DTD_T, this upper bound is generally never reached. This disproves that the metric is optimized for i-SAEs through the matrix PP.

W1.2) The MS score is computed per modality, and a dictionary trained on a single modality naturally obtains a higher single-modality MS. Since these metrics underpin the claim of better cross-modal interpretability, this bias weakens the paper's main conclusion.

We respectfully point out that it cannot be taken for granted that a dictionary trained on a single modality obtains a higher single-modality MS. The MonoSemanticity Score (MS) measures the semantic similarity among the examples that activate a given feature. Since i-SAEs are trained on a single modality, they can encode a large amount of information using a relatively small number of active atoms. Consequently, multiple distinct concepts may be grouped within the same feature rather than being represented by separate features. This results in greater semantic heterogeneity among the examples activating a given feature, potentially leading to lower MS scores.
Conversely, this effect may be mitigated in the multimodal output spaces of contrastive VLMs, where the shared representation space can promote more semantically coherent feature activations. Overall, there is no inherent bias favouring single-modality dictionaries in terms of MS. Rather, the score reflects the semantic organization of the learned representations.
Our experimental results show that i-SAEs consistently outperform or match s-SAEs in terms of both image and text MS scores, while also generally reducing the percentage of dead codes in the visual domain. This suggests that i-SAEs lead to better concept discovery.

Other concerns

W2.1) Matching quality is not quantified, so it is hard to know what fraction of the Hungarian-matched pairs are semantically consistent.

To illustrate how Hungarian Matching aligns dictionary atoms to mirror real data characteristics, we evaluated the average similarity of Independent SAEs’ atoms on CC3M for K=8 on CLIP ViT-L/14:

DatasetIndependent SAEInd. SAE + InfoNCEReal Data (Reference)
CC3M0.2260.2270.238
YFCC15M-V20.2310.2720.209

As observed in the table, values are very similar to each other, highlighting consistency to the datasets in their original representation.
To quantify the fraction of semantically consistent pairs, we calculated the Matching Ratio. Using the dataset's average real-sample similarity as a threshold, we found that 31.17% of Independent SAE atom pairs (and 31.34% with InfoNCE loss) have higher similarity. This highlights that not only do we have general semantic consistency, but also there are more than 30% of atom pairs which exhibit more semantic similarity than real-world data. Without Hungarian matching, or with random permutations, average similarities for all configurations fall to near zero, ranging from -0.005 to 0.003, and matching ratios fall to 0.023% ± 0.014% for random permutations. Additionally, these numbers also highlight why the permutation matrix is essential even when using a co-activation loss.

Finally, our multimodal monosemanticity scores (MMS, Figures 6, 13, 22, 26) reinforce our results, demonstrating that independent SAEs learn significantly more semantically consistent cross-modal features than shared SAEs, as the score is meant to properly measure the quality of the semantic mapping across modalities.

W2.2) The human study is small (9 participants and 223 responses, recruited from the authors' institution).

We agree on this point and we acknowledge that the study should be expanded to a larger pool of participants in order to provide a robust statistical analysis. However, this is not meant to be a central contribution of our work, but rather a preliminary investigation to complement the already rich suite of experiments based on proxy metrics and assessment on downstream tasks. Hence, it is presented in the appendix. Nevertheless, the very low p-values do give us reassurance that the results of the study are significant despite the small sample size.

W2.3) The i-SAE has twice the atoms and parameters of the s-SAE in all main results; the equal-parameter comparison is limited to a single model and dataset.

We intentionally matched the dictionary size kk to ensure a fair comparison of conceptual capacity rather than strict parameter count. Because a s-SAE uses one atom per bimodal concept, while the modality gap forces i-SAEs to use an atom pair (one image, one text), setting all dictionaries to size kk gives both architectures the exact same theoretical upper bound of kk representable concepts. Halving the independent dictionaries strictly to match total parameters would artificially handicap them to k/2k/2 concepts. However, we agree with the importance of isolating architectural gains from parameter scaling. To confirm our improvements stem from resolving the modality gap rather than merely doubling the parameters, we have provided the direct equal-parameter comparison in Appendix F.2.1 by varying kk. Crucially, when comparing an Independent SAE with half the dictionary size (k/2k/2) against a s-SAE of size kk (an exact parameter match) the i-SAE still consistently outperforms the Shared SAE. This performance gap remains especially pronounced in the interpretability evaluations, confirming that the architectural separation itself, not the parameter count, drives the demonstrated gains.

Q1) How is P obtained for the trace ratio: on training or held-out data?

To compute the permutation matrix P we only need the dictionaries, which are fixed after training. We use the Hungarian Matching algorithm applied to the dictionaries.

Q2) Can you evaluate zero-shot classification and retrieval using only the Hungarian-matched atom pairs?

We did not quite understand the question, because Hungarian-matched atom pairs are fixed for each sample, so this would lead to equal cosine similarity for all samples. Could you please give us more details about what you would like us to evaluate?

Q3) Please provide equal-parameter results for additional VLMs and datasets.

While the k/2k/2 experiment (equal-parameter count) is a proof-of-concept experiment that the modality-specific nature of dictionaries lead to higher performance and interpretability results, we highlight that the main focus should remain on the dictionary size equal for all SAEs. The main takeaway should be that as the modality gap exists in contrastive VLM output spaces, then a geometric prior should be used to interpret their activations effectively. In the case of single s-SAEs, bimodal concepts do not reside in the modality subspaces determined by the modality gap, ultimately degrading the faithfulness of the decomposition and its interpretability outcomes. We instead advocate for a solution where the bimodal atoms reside in these separate subspaces, meaning the requirement of having a single shared concept represented by a pair of atoms, one per modality. This provides a new way of thinking about bimodality.
Please, refer also to the answer in W2.3.

Q4) Per-modality SAEs combined with Hungarian matching seems like an obvious baseline. Could you clarify what is novel about the proposed approach beyond this baseline?

We agree that our architectural approach is technically straightforward, and we acknowledge that prior work [1,2,3,4] has investigated using a single s-SAE across multiple modalities. However, our primary novelty is conceptual and empirical rather than architectural: we demonstrate that the use of a single s-SAE is fundamentally misaligned with the geometric structure of contrastive VLMs.

While training a single SAE might seem appealing, it encounters severe structural limitations due to the modality gap. Efforts to bypass this gap using a single s-SAE require significant compromises, such as restricting the dictionary to only bimodal atoms [2] or using a split-dictionary [1]. Because a single s-SAE is forced to represent a bimodal concept using a single atom in an intermediate subspace, it struggles to capture the distinct geometry of each modality, sacrificing both faithfulness to the original model activations and interpretability.

In light of these limitations, our core contribution lies in a fundamental change of perspective. Rather than attempting to engineer away the modality gap, we propose embracing the inherent geometric structure of the VLM. By explicitly introducing a geometric prior instantiated through per-modality SAEs combined with Hungarian Matching, we allow bimodal atoms to live natively in their respective subspaces. In doing so, we provide the following clear scientific contributions:

  • Identifying the failure of shared architectures: we demonstrate that s-SAEs fail to provide faithful interpretations, highlighting the severe implications of architectural biases that ignore the modality gap;
  • Establishing the necessity of a geometric prior: we show that respecting the geometric structure of representation spaces is a fundamental prerequisite for faithful interpretability;
  • Conducting a focused study on interpretability and faithfulness metrics: unlike previous works that largely neglect these aspects, we provide a rigorous and comprehensive evaluation over multiple VLMs, datasets, sparsity levels and expansion ratios. Through our experiments, we demonstrate that a conceptually simple approach, such as i-SAEs without any co-activation loss, can significantly outperform existing, more complex solutions.

While our resulting architecture is simple, we view this as a significant strength. An easily implementable solution that rigorously respects the model's geometry is strictly more practical and effective than an overly engineered one. The fact that this conceptually simple approach significantly outperforms existing paradigms is exactly the core message of our paper. We believe our work establishes a solid foundation for representation-level interpretability in this setting and serves as a useful basis for the community to build upon.

References

[1] Isabel Papadimitriou, Huangyuan Su, Thomas Fel, Sham M. Kakade, and Stephanie Gil. Interpreting the linear structure of vision-language model embedding spaces. In Second Conference on Language Modeling, 2025.
[2] Grégoire Dhimoïla, Thomas Fel, Victor Boutin, and Agustin Martin Picard. Cross-modal redundancy and the geometry of vision–language embeddings. In The Fourteenth International Conference on Learning Representations, 2026.
[3] Valérie Costa, Thomas Fel, Ekdeep Singh Lubana, Bahareh Tolooshams, Demba E. Ba. From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025.
[4] Chiraag Kaushik, Davis Barch, and Andrea Fanelli. Decomposing multimodal embedding spaces with group-sparse autoencoders. In The Fourteenth International Conference on Learning Representations, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.