acceptodds
Under review as a conference paper at ICLR 2027

Cross-Modal Latent Influence: Patch-Level Training Sample Attribution for Vision-Language Models

Abstract

Explaining a vision-language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image-text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decompose training-example influence through these concepts and localize their contributions to specific text phrases and image regions. We evaluate MOSAIC on Qwen and Gemma across captioning, chest X-ray understanding, and visual question answering. The learned features are more localized and coherent without degrading task behavior, and ablating influence-ranked latents changes model outputs more than ablating latents ranked by activation or frequency. We find that visual evidence can substantially reorganize training-data attribution while preserving more of the shared latent structure. MOSAIC thus connects where model behavior comes from with what multimodal evidence carries that influence.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language Models, Sparse Autoencoders, Training Data Attribution

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

This paper aims to investigate how training examples influence predictions in vision-language models, with a focus on the relationship between these influences, internal features, image regions, and textual content. The authors propose MOSAIC, a method that combines sparse autoencoders with training-data influence estimation. By introducing spatial sparsity constraints and cross-modal alignment, the method learns shared latent features across images and text. It then traces the influence of training examples to specific image patches and text phrases. The authors conduct experiments on Qwen and Gemma across image captioning, chest X-ray understanding, and visual question answering tasks. The results show that MOSAIC produces more localized visual features and clearer cross-modal semantic alignment while largely preserving model performance. Ablation experiments further suggest that features ranked by influence scores have a greater impact on model predictions than those selected based on activation values.

Critical strengths
  1. This paper extends training-data attribution to fine-grained visual and textual evidence by connecting influential training examples with sparse latent features, image patches, and text phrases. This provides a more detailed view of how training data relates to model predictions.
  2. The paper provides experiments across different models and tasks. The latent intervention results also show that influence-ranked features have a greater impact on predictions than activation-based baselines, supporting the usefulness of the proposed method.
Critical weaknesses
  1. The paper lacks direct validation of training-data influence through retraining experiments. While the latent intervention results demonstrate the importance of selected features, they do not necessarily confirm the accuracy of the estimated training-example influence. A small-scale retraining experiment would help strengthen the reliability of the attribution results.
  2. The analysis of removing visual inputs shows substantial changes in training-example rankings, but the retrieved candidate sets have very little overlap. Although the authors also examine shared candidates, the results vary across different settings. It remains unclear how much of the observed difference comes from candidate retrieval rather than changes in influence scores. A comparison using a fixed candidate set would help clarify this issue.
Questions for authors
  1. Have the authors considered validating the estimated training-example influence through small-scale retraining experiments? If so, how well do the results agree with the estimated influence scores?
  2. In Section 4.4, removing the image leads to substantial changes in training-example rankings, while Appendix I.2 shows very little overlap between the retrieved candidate sets. Although the authors analyze shared candidates, would the same ranking changes still be observed if influence scores were recomputed using a fixed candidate set under both conditions?
  3. In Section 3.1, the influence scores are computed using gradients with respect to the upper VLM parameters, while the models are fine-tuned using LoRA. What motivated this choice? Would computing influence with respect to the LoRA parameters lead to different training-example rankings?
  4. In Section 3.1, the authors estimate latent contributions using activation-weighted gradients. Do these contributions sum to the estimated influence of the corresponding training example? If not, how large is the difference?
  5. Appendix D.3 shows that SAE layer placement affects model performance. Does changing the SAE layer also affect the training-example rankings or the identified image patches?
4 comments
AuthorsOpenReview ↗

Gentle reminder about our response

Dear Reviewer vhkS,

Thank you again for your thoughtful review. As the discussion period closes on November 18 (AoE), we wanted to kindly follow up on our response.

Following your comments, we conducted additional experiments and clarifications on (1) training-example influence, validated through retraining (5.7–15.7× stronger effects than random) and reweighting (ρ=0.77); (2) fixed candidate sets, confirming our attribution findings across six settings; (3) parameter scope, supporting our upper-VLM choice; (4) latent contributions, clarifying their interpretation and validating rankings (ρ=0.84); and (5) SAE layer placement, showing stronger grounding and intervention results at our selected layer.

If our response has addressed your concerns, we would be grateful if you would consider reflecting this in your assessment. If any concern remains, we would be glad to provide further clarification while the discussion is still open.

Best regards,
The Authors

AuthorsOpenReview ↗

Followup on Rebuttal for Reviewer vhkS

We sincerely thank the reviewer for their time and thoughtful feedback. As the discussion period ends today, we would greatly appreciate a response indicating whether our additional experiments and clarifications address the remaining concerns as well as if there is anything remaining that we could clarify. We kindly ask the reviewer to consider updating their rating if there are no remaining concerns.

Thanks to the authors for their detailed response and additional experiments. The new experiments have substantially addressed my previous concerns and strengthened my confidence in the reliability of the proposed method. However, the fine-grained attribution results still exhibit some sensitivity to SAE layer placement and the parameter scope used for influence estimation. Therefore, I retain some reservations about their stability across different settings. Accordingly, I have increased my Soundness score from 2 to 3 while maintaining my overall rating unchanged.

AuthorsOpenReview ↗

Response to Reviewer vhkS

We thank the reviewer for their thoughtful follow-up and updated assessment. We appreciate their recognition that our additional experiments have substantially addressed the original concerns.

We would like to clarify the remaining concern regarding stability. Sensitivity to SAE layer placement and parameter scope is expected and does not, by itself, indicate unreliable attribution. Different SAE layers encode different representations, and their ability to preserve task-relevant information can vary substantially. In particular, when a different layer yields degraded task performance, changes in attribution rankings are expected rather than necessarily indicative of instability. Similarly, different parameter scopes measure influence through different parameter spaces. Identical rankings across these settings are therefore neither theoretically expected nor necessary for reliability.

Importantly, cross-configuration ranking agreement should not be conflated with attribution reliability. The more meaningful criterion is whether the identified influences predict actual model behavior. Our controlled retraining experiments provide direct evidence: under both parameter scopes, removing influential training examples produces loss changes in the predicted direction, approximately 5–15× larger than random removal. Across SAE layers, layer 15 achieves stronger grounding, necessity, and sufficiency results, providing empirical justification for its selection.

We therefore believe that the observed sensitivity reflects meaningful differences in the underlying representations rather than evidence of unreliable attribution. Our experiments demonstrate that the selected configuration produces behaviorally meaningful attributions, while the variation across configurations is consistent with their distinct representational and parameter scopes.

We hope this clarification addresses the reviewer's remaining reservation and would be grateful if they would reconsider their overall assessment in light of the additional evidence.

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

The paper proposes MOSAIC to connect influential training examples with image regions and text phrases. It combines a shared sparse autoencoder, spatial regularization, phrase–region alignment, and EK-FAC influence estimation. Experiments cover two VLMs and three tasks. They show improved feature localization and stronger effects from influence-ranked feature interventions. Case studies illustrate the attribution pipeline. However, the experiments provide stronger evidence for feature importance than for accurate training-data attribution.

Critical strengths
  1. The problem is important and the proposed connection is useful. Linking training examples to specific visual and linguistic evidence could improve VLM auditing. The framework provides a concrete way to inspect this connection.

  2. The feature-level evidence is substantial. Evaluation spans two architectures and three tasks, with quantitative results and case studies. On GQA with Qwen, visual footprint decreases from 0.688 to 0.120. On GQA with Gemma, removing 256 influence-ranked features flips 38.3% of predictions, versus 8.3% for activation ranking. These are large observed differences that support the usefulness of the learned features.

Critical weaknesses
  1. The experiments do not establish the claimed training-data attribution. Removing or retaining test-time latents shows that those latents affect predictions. It does not show that the identified training examples caused those effects. No controlled training-example reweighting or retraining experiment validates the estimated influence. The numerical curvature check does not fill this gap. A target-loss gradient × activation baseline is also missing, so the benefit of using training influence to rank features remains unclear. Patch-level faithfulness is additionally unresolved for Gemma. Appendix H acknowledges that its shared key–value states prevent the patch-level grounding test at layer 19. The official implementation confirms that upper layers reuse earlier key–value states. This suggests a visual-information path that bypasses the SAE. Intervening on text positions could therefore explain the reported effects without validating image-patch mediation.

  2. The image-removal conclusion is confounded by candidate retrieval. Table 7 reports candidate-pool overlaps of only 0.021–0.145. Thus, the negative example-level correlations mix changes in retrieval with changes in influence. For CheXpert/Gemma, the correlation is −0.797 over the union but +0.850 among shared candidates. The latter exceeds the latent correlation of +0.795. The claim that examples change more than concepts therefore needs a comparison that recomputes both conditions over the same candidate set.

Questions for authors
  1. How well do the estimated signed influences predict actual loss changes under controlled training-example reweighting or retraining?

  2. Are Gemma’s shared key–value states recomputed during SAE interventions? If not, how do layer-19 image latents affect the answer loss?

  3. Does the conclusion hold when influence is recomputed over the same candidate set in both conditions? How are missing candidates handled in the current union correlation?

  4. What gains does MOSAIC provide over combining Yu et al.’s latent-mediated influence with Liao et al.’s structured sparse autoencoder? Can a controlled comparison isolate the benefits of the proposed alignment and normalization changes?

  5. Please provide dataset subset sizes, SAE learning rates and training duration, LoRA settings, and the retrieved candidate count. What are the sample counts and uncertainty estimates for the intervention results?

1 comment
AuthorsOpenReview ↗

Gentle reminder about our response

Dear Reviewer zbUC,

Thank you again for your thoughtful review. As the discussion period closes on November 18 (AoE), we wanted to kindly follow up on our response.

Following your comments, we conducted additional experiments and clarifications on (1) training-data influence, validated through retraining (5.7–15.7× stronger effects than random) and reweighting (ρ=0.77); (2) Gemma's K/V architecture, demonstrating patch-level faithfulness; (3) fixed candidate sets, confirming our attribution findings across six settings; (4) controlled comparisons, improving detection AUROC from 0.55 to 0.82; (5) gradient × activation baselines, showing stronger causal effects from influence-ranked latents; and (6) implementation details, including reproducibility settings and confidence intervals.

If our response has addressed your concerns, we would be grateful if you would consider reflecting this in your assessment. If any concern remains, we would be glad to provide further clarification while the discussion is still open.

Best regards,
The Authors

Reviewer_Uo32OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper is a combination of two correctly cited previous works. On the one hand, the manuscript extends (Yu et al., 2026), which leverages the fact that (training sample) influence functions can be soundly computed in the orthogonal space of SAE latents (due to lower token level correlation) to VLMs. On the other hand, the method in (Liao et al., 2026) is employed to achieve spatially structured latent representations. Novelty: first, phrase–region alignment ensures the same latent coordinate captures corresponding linguistic and visual evidence via a phrase region alignment objective; second, postiion-wise attribution. The result allows both for training sample influence computation on VLMs and spatial interpretability of activations.

Critical strengths

The paper correctly extends recent advancements in influence computation to VLMs, with the useful fact that the latent representation becomes spatially grounded, allowing to map latents back to image regions.

Enough ablations are included to understand the influence of all design choices.

Critical weaknesses

The scale of the contribution is uncertain. While known methods are correctly combined and extended to VLMs, the usefulness of this is not properly discussed. Section 4.2 and 4.3 show the method works, but the metrics naturally benefit spacial grounding. If the claim is phrase-region alignment etc. improve the representation space w.r.t. the vanilla method (Base SAE), can you guarantee this is not simply due to the increased loss signal imposing stronger regularity on SAE latents (as the decrease in dead might suggest), and that in fact visual grounding is necessary for better representations? Besides, I’m missing a use-case that showcases how influence computation benefits from visual grounding.

Section 3 is hard to follow, and has a lot of relevant information pushed to the appendices.

Questions for authors

Even using EK-FAC, for which dataset size does the cost of storing gradients become prohibitive? Is this only feasible for fine-tuning datasets?

Regarding Multimodal Influence Decomposition, position-wise attribution gives you in which image/text tokens a feature is expressed, but what guarantees the feature is not expressed both for image and text tokens? Perhaps the decomposition in the title is a bit misleading.

2 comments
AuthorsOpenReview ↗

Gentle reminder about our response

Dear Reviewer Uo32,

Thank you again for your thoughtful review. As the discussion period closes on November 18 (AoE), we wanted to kindly check whether you have had a chance to read our response.

Following your comments, we conducted additional experiments and clarifications on (1) visual grounding, demonstrating improvements beyond regularization (27–28% multimodal latents vs. 11%) and practical mislabeled-example detection (AUROC 0.95); (2) dataset size, showing how we store training examples; (3) multimodal feature representation, clarifying how shared latents support modality-specific influence attribution.

If our response has addressed your concerns, we would be grateful if you would consider reflecting this in your assessment. If any concern remains, we would be glad to provide further clarification while the discussion is still open.

Best regards,
The Authors

AuthorsOpenReview ↗

Followup on Rebuttal for Reviewer Uo32

We sincerely thank the reviewer for their time and thoughtful feedback. As the discussion period ends today, we would greatly appreciate a response indicating whether our additional experiments and clarifications address the remaining concerns as well as if there is anything remaining that we could clarify. If our responses have addressed your concerns, we would be grateful if you would consider reflecting this in your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.