Your suggestions are insightful and will enhance the completeness of our paper!

We used all available resource and devoted efforts to conduct additional experiments. We addressed all your negative comments below and will add them into the revision. If this rebuttal addresses your concerns, we earnestly ask you to consider raising the score and supporting us for acceptance.

Claim 1 & Weakness: We prune the overall presentation of the paper to reduce repetitive expressions and add necessary backgrounds.

Limited by the length of rebuttal, we are not able to list all the changes we made, so only key examples are provided. E.g.:

• line 245-253:

The transferability of SAEs between foundation models and instruction-tuned models has been extensively investigated in text-only contexts[3][4][5], as it demonstrates whether SAEs can capture universal semantic features within LLMs. Similarly, the transferability from MLLMs to corresponding LLMs serves as a critical metric for the quality of features learned by SAE-V. (Followed by original content)

• Fig. 7 caption:

We evaluated SAE-V data filter method on LLaVA-NeXT-7B model, Align Anything dataset, and LLaVA-Bench benchmark. The result show that all SAE-V-based methods significantly outperforms the random selection baseline, while the cosine similarity filter achieved 108% of the full dataset’s performance with only 20% of the data, and the co-occurrence filter peaked at 50% of the data, reaching a score of 108.17.

To address the specific concerns:

• SAE only accepts textual tokens, while SAE-V is designed for both text and image tokens. When evaluating with multimodal input, SAE can only reconstruct text features, whereas SAE-V reconstructs both text and image features, achieving lower reconst. loss (Fig. 4, Col. 3). Moreover, even when evaluating on text tokens, SAE-V surpasses SAE in reconstruction capability (Fig. 4, Col. 1-2), showing effective cross-modal generalization.
• Regarding the "Original" bar in Fig. 3: The reconst. loss measures how well the model predicts the next token compared to ground truth. The original model's predictions are also probability distributions, not exact predictions, which is why the "Original" has non-zero reconst. loss.

Claim 2: We provided additional details, and conducted extra ablation study to compare SAE-V with non-training baseline.

Regarding details of the experiment in Section 3.2:

• Pipeline: Given the 1000 val. set of ImageNet, We score the 24*24 image patches according to the metrics of SAE-V features (Fig. 6 legend). We filter out the patches according to the score and the given ratio (x-axis in Fig. 6). We use LLaVA-NeXT-7B to classify the filtered image, and calculate the accuracy (Not loss value) of LLaVA-NeXT-7B output as y-axis in Fig. 6.
• Target: This experiment is designed to support the claim that SAE-V could preserve the key info from images, and the higher accuracy is, the more key info is preserved through the leftover patches. E.g., in Fig. 5, SAE-V could preserve the most relevant information (the dog) in the image.

We added an additional baseline to this experiment, using attention score of image patches and the original text token to filter the patches. Results:

SAE-V could achieve relatively similar performance using reconstruction instead of original activation of MLLMs.

Claim 3: We updated the presentation and added additional baselines.

Sorry for the ambiguous figure. To clarify, we used LLaVA-Bench[2] to report the MLLM performance, and the y-axis is not percentage, but the exact score on LLaVA-Bench. To demonstrate that our method is superior regardless of benchmark selection, we did an ablation study of benchmarks. For more details see the rebuttal of Reviewer k2no point 1.

As for data selection, we included the CLIP baseline and added a paragraph discussing the related work [1]. The experiment results are as follows:

The experiment shows that although the peak performance is close, CLIP achieve the peak performance using 4 times data than SAE-V, showing the effectiveness of SAE-V.

Reference

[1] Gadre et al., DATACOMP: In search of the next generation of multimodal datasets, 2023.

[2] Liu et al., Visual instruction tuning, 2023.

[3] Kissane et al., Saes (usually) transfer between base and chat models, 2024.

[4] Taras et al., Do Sparse Autoencoders (SAEs) transfer across base and finetuned language models?, 2024.

[5] Gallifant et al., Sparse autoencoder features for classifications and transferability, 2025.

Thanks for your valuable suggestion!

During rebuttal period, we used all available resource and devoted efforts to conduct additional experiments. We addressed all your negative comments below and will add them into the revision. If this rebuttal addresses your concerns, we earnestly and kindly ask you to consider raising the score and supporting us for acceptance.

point 1

I am unsure what benchmarks authors use to report the MLLM performance, so I would like clarification.

We used LLaVA-Bench[1] to report the MLLM performance, and we also test our method on MME benchmark. Here is the result on LLaVA-NeXT-7B and Align Anything dataset:

MME:

This result demonstrates that the superiority of SAE-V is not effected by the selection of benchmarks.

point 2

Why do they only try to fine-tune the LLaVA-NeXT-7B model and not pretrain and fine-tune the model from scratch? This is important to know how does the dataset filtering affect different training stages.

We appreciate the reviewer's suggestion about pretraining from scratch. While we fully agree this would provide valuable insights into how our filtering method affects different training stages, the computational resources required for pretraining MLLMs from scratch are prohibitively expensive for academic research team like us.

Pretraining even a 7B parameter model requires hundreds of GPU-days on A100/H100 GPUs, which is unfortunately beyond our current resource constraints. Instead, we focused our experiments on fine-tuning existing models, which still demonstrates the effectiveness of our approach while being computationally feasible (for more details supporting this claim, see the rebuttal of Reviewer b89F W3&W5 and Reviewer k2no point #4.). We commit to conduct at least one experiment at pretrain stage in the future versions of our paper.

point 3

Extending SAE to MLLMs is nothing technically innovative.

While extending SAE to MLLMs may appear straightforward, our work contributes several key innovations:

• Cross-modal feature analysis: We've developed novel methods to identify and analyze features that capture cross-modal interactions (Section2.2).
• Interpreting Multimodal Alignment: Our framework provides unique insights into how MLLMs integrate information across modalities during the alignment process. As shown in Section 3.1.2, SAE-V reveals patterns in feature distribution that directly correspond to model performance on multimodal understanding tasks.
• Self-guided data filtering: Our paper made the first attempt to use mechanistic interpretability methods to perform multimodal data filter using the model's own representations.

Our work provides additional insights and extends the practical applications of multimodal interpretability methods, which is also the main contribution of our paper.

point 4

The findings about the cosine-score and performance are interesting, but I'd have liked to see more models and more datasets being used for analysis.

We conducted experiments on larger models (LLaVA-NeXT-Vicuna-13B) and datasets (MMInstruct) during rebuttal period.

For experiment on larger models, see the rebuttal of Reviewer b89F W3&W5.

For experiments on larger datasets, we selected MMInstruct[2], which contain 200k data pieces, and is 4 times bigger than Align Anything dataset. Based on MMInstruct, we applied SAE-V-based data filter and alignment on Chameleon-7B, and the results are shown below:

It demonstrates that SAE-V paradigm could scale up to larger models and datasets, while maintaining its performance.

Reference

[1] Liu et al. Visual instruction tuning, 2023. [2] Liu et al. Mminstruct: A high-quality multi-modal instruction tuning dataset with extensive diversity, 2024.

Despite some misunderstandings, we conducted more experiments to clarify your valuable concerns.

We conducted additional experiments, addressed your comments, and would add them into revision. If this rebuttal addresses your concerns, we kindly ask you to consider raising the score.

Methods & Evaluation: We added additional paragraph to clarify this potential confusion.

In Eq. 1, we define the operation $S_{\theta}(\cdot)$ as the encoding operation of the SAE-V, whereas $Z_i \in R^{\ l\times n}$ is the feature activation, $l$ is the length of the input to SAE-V, and $n$ denotes the number of features in the SAE-V. Each $z_j \in R^{1\times n}$ represents the activation of a specific token (index $j$) across all features in the SAE-V.

Theoretical Claims: We sincerely apologize for the mistake.

Thanks for pointing this out! Based on our definitions, Eq. 1 should be: $Z = ReLU(H\times W_{enc} + b_{enc})$. We updated the paper accordingly.

Broader Literature: We perform experiments to show the scalability of SAE-V empirically.

Regarding [1], we acknowledge that reduced reconst. loss alone cannot guarantee meaningful features. However, our work empirically validates SAE-V through image patch filtering (Section 3.2), and data filter & alignment (Section 4).

While the theoretical analysis of underlying mechanism remains an open question for future work, we conducted experiments to show that SAE-V has the potential to scale up to larger models and datasets. For more details, see the rebuttal of Reviewer b89F W3&W5 and Reviewer k2no point #4.

References: We acknowledge the novelty of [2], though we are making distinct contributions.

We acknowledge that [2] is a pioneering effort in applying SAE for mechanistic interpretability of VLMs, and we modified our claims accordingly and added [2] into the related works subsection. Compared with [2], our work provides additional insights and extends the practical applications of mech interp. For more details, see the rebuttal of Reviewer k2no point #3.

W1&2: We prune the overall presentation of the paper to reduce repetitive expressions and add necessary backgrounds.

Thanks for your suggestions! We have modified the paper accordingly. For some decent examples, see the rebuttal of Reviewer o9Vp Claim 1.

W3: We conducted deeper analysis of the underlying mechanisms.

We replicated the patch filtering experiments in Section 3.2 on VQA tasks (using A-OKVQA val. set with LLaVA-NeXT-7B), examining both text tokens and image patches:

Key Findings

• Compression rate: Image information demonstrates lower compression rate (more redundancy) than text.
• Accuracy scaling w/ reduced tokens: Text masking shows a roughly linear relation of accuracy and masked token, while image filter maintains performance until 50%, with a more significant drop only appearing at 75%, suggesting that the information of text is more evenly distributed compared to image.

For interpretability beyond reconst. score, see the rebuttal of Reviewer b89F W1.

For more datasets, see the rebuttal of Reviewer k2no point #4.

W4: We presents adaptive parameter selection strategies to make our methods effective and pratical.

Thank you for raising this consideration! We tested hyperparameter selection using a 1/20 subset of Align-Anything dataset:

The overall trend closely resembles Fig. 7, confirming that alignment metrics on a small val. set resemble the distribution on the complete dataset, enabling efficient hyperparameter selection.

Additionally, Section 4 have showed that our method outperforms using the complete dataset across most hyperparameter settings, making more refined parameter tuning (such as naive elbow) helpful but not strictly necessary.

W5

While we don't provide formal guarantees, our method is grounded in established principles of dictionary learning that have been validated by the community. We also performed comprehensive ablation study and open-sourced our code to confirm replicability. (see rebuttals above).

Question

Thanks for pointing out! As is shown in the supp. material (code/SAELens-V/scripts/cosimilarity.py, line 179-189), we actually used point-to-point cosine. We modified Eq. 7, and we verified that pair-wise consine performs similarly as the point-to-point in actual dataset selection:

Reference

[1] Heap et al. Sparse Autoencoders Can Interpret Randomly Initialized Transformers, 2025.

[2] Zhang et al. Large Multi-modal Models Can Interpret Features in Large Multi-modal Models, 2024.

We deeply appreciate your thoughtful insights that will significantly strengthen our paper's overall presentation.

In the rebuttal period, we used all available resources and devoted efforts to conduct additional experiments. We addressed all your negative comments below and will add them into the revision. If this rebuttal addresses your concerns, we earnestly and kindly ask you to consider raising the score and supporting us for acceptance.

W1: We defined these concepts with specific metrics, and we acknowledge that more rigorous definitions would strengthen our presentation.

We define a feature as interpretable when it activates for semantically related inputs across modalities. This definition is aligned with the automated interpretability score[1], but since there are currently no benchmarks for multimodal interpretability, we would display an example instead. (e.g., Rebuttal Fig. 1 shows an interpretable feature with its semantic meaning being consistent.)

As for quality, a feature is of high quality when it activates text and image patches that are semantically similar (given by Eq. 7), focusing more on the similarity across modalities instead of overall consistency. The feature in the example above would also be of high quality.

W2: We provided examples for the failure modes, as well as statistical evidence supporting that these failure modes are scarce.

We agree that there are failure cases where SAE-V didn't behave well on multimodal data, and here is an example: Rebuttal Fig. 2. In this example, SAE-V fails to capture the most informative patches because of its similarity with the background.

However, we need to mention that SAE-V is effective in most cases. Shown in Fig. 6, all SAE-V-based methods achieve high accuracy when preserving 75% or 50% patches, and cosine similarity score method maintains high accuracy even when only 25% patches are preserved. For more baselines and ablations, see the rebuttal of Reviewer bBFh Weakness 3 and Reviewer o9Vp Claim 2.

W3&W5: We tested our method on additional models to prove that our method could generalize and scale up.

To prove that SAE-V and its data filtering paradigm could generalize to other multimodal models and scale up to larger models, we replicated SAE-V and its data filtering method on LLaVA-NeXT-Vicuna-13B and LLaVA-NeXT-Vicuna-7B (in the paper, we used LLaVA-NeXT-Mistral-7B). Unfortunately, due to rebuttal time constraints and compute limitations, we were unable to test our method on larger models and more architectures, and we only tested our data filter method in a 5-fold manner rather than 10-fold in the paper.

The interpretability metrics of SAE and SAE-V on both models are shown in the table below:

In all MLLMs and metrics, SAE-V consistently outperforms SAE, demonstrating its superior capability across different architectures, sizes, and semantic complexity.

The alignment experiment results of SAE-V are shown in the table below:

LLaVA-NeXT-Vicuna-13B:

The results show that SAE-V-based data filter outperforms the random selection baseline, and reached the highest performance of 116.67 with 40% data.

We believe that our experiments across 7B, 13B models and three different architectures (Chameleon, LLaVA-NeXT-Mistral, and LLaVA-NeXT-Vicuna) provide sufficient evidence to prove SAE-V's potential to generalize across architectures, scale to larger models, and handle more complex semantic spaces. We commit to adding experiments on at least one 30B-scale model in future versions of our work.

W4: We reported our training time and computation cost, and demonstrated the effectiveness of SAE-V.

Our SAE-V training was completed on 8*A800 GPUs. Using 100k multimodal data samples, each training typically takes around 21 hours, which is comparable to training an SAE on 7B model using the same amount of data.

The effectiveness of SAE-V lies in its generalization capabilities: as reported in Section 3.1.2 and Fig. 4, SAE-V trained on MLLMs demonstrates strong generalization capability to the corresponding LLMs. Therefore, a single training can yield an SAE-V that is applicable to both multimodal models and text-only models.

Reference

[1] Bills et al., "Language models can explain neurons in language models", 2023.