acceptodds
Withdrawn from ICLR 2027

VISTA: Visual Information Stress Testing for Representation Shifts and State-Dependent Influence in Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

What changes inside a vision-language model when a visual category remains recoverable under perceptual stress? We study matched color-desaturation and shape-defocus sequences across eight checkpoints, combining frozen readouts, native candidate scoring, and visual-state interventions. We find that severe clean-readout mismatch has a simple correctable component: a training-estimated, class-agnostic translation restores Shape readout accuracy to 0.979–1.000 in the frozen IID evaluation, while representation discrepancies are only partly removed. On a separate, explicitly linked evaluation cohort, all categories remain recoverable by a shared reader, yet Color stress consistently reduces true-class relative evidence and candidate accuracy; Shape changes the evidence in model-dependent directions while endpoint decisions remain near ceiling. Finally, recipient stress strengthens a fixed clean–stress visual contrast in Qwen–Color but weakens it in LLaVA–Shape. These opposite modulations persist across same-family, cross-family, and finite-interval constructions, with natural-direction amplitude advantages over matched controls; other settings are less consistent. The findings characterize translation-correctable readout mismatch, attribute-dependent answer changes, and heterogeneous conditional visual influence rather than a uniform loss of categorical information.

Primary area:
interpretability and explainable AI
Keywords:
Vision-Language Models, Mechanistic Interpretability, Representation Probing, Activation Interventions, State-Dependent Visual Influence, Multimodal Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ws47OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper discusses how Vision-Language Models (VLMs) operate under input shifts, i.e., changing properties of the input image without altering its target 'class'. They construct a toy dataset in which a VLM classifies differently colored shapes. At test time, they vary the saturation and blur of the input image and test whether the VLM predictions remain the same. To understand the source of change in predictions, they look at the VLM's image encoder representations and text decoder predictions.

Critical strengths

None

Critical weaknesses
  • The scope of the study is very narrow. The authors aim to test the performance of VLMs under input shifts using a toy dataset of colored shapes, varying saturation and blur levels. This toy setting is far from any practical application. By using only single shapes, the analysis does not consider cases where confounders can play a crucial role. For instance, the presence of multiple objects requires the model to localize predictions and isolate the contributions of each object. Additionally, by only considering two very simple input shifts (color, blur), it is difficult to draw any meaningful conclusions about the models in question.

  • My main issue with the paper is that there are no clear takeaways from the study:

    1. When the input shift is big, the image representation differs significantly from the original, but it can be recovered with a simple mean correction. This does not answer anything about the image encoder, let alone give us any insight into how the VLM operates. Additionally, some VLMs (LLaVA) seem to be less affected by specific shifts, which means that the image representation is somehow different, but the authors do not elaborate further on that finding.

    2. The text decoder predictions change but do not completely fail under these shifts, and the larger the shift, the worse the prediction gets. This is an expected result but again does not provide any insight into why this is happening and why there are discrepancies across the different VLMs.

    3. In their last experiment, the authors directly replace some of the activations of a non-shifted image with the activations of its shifted counterpart. This is not really connected to the other two evaluations, and it is unclear what its goal is. The results they show (Section 4.3) are hard to interpret, as they do not establish exactly what question they are trying to answer.

    Overall, the results come across as a collection of experiments with no coherent story to convey. Having read Section 4 multiple times, I still struggle to understand what the takeaway message is supposed to be.

  • The writing of the paper makes it impossible to follow. The authors keep introducing terms, such as scaffold group and recipient stress, which have not been properly defined in the paper. Some of these terms may be common in VLM works, but for a general reader, it requires a lot of back-and-forth to try and connect the various concepts. Single-line conclusion paragraphs such as "Thus, these settings show both consistent signed modulation and larger absolute modulation along the natural direction than along the matched controls." (line 428) should be self-explanatory since they are trying to convey a straightforward conclusion/result to the reader. In their current state, they are riddled with ambiguous terms and confuse the reader even more.

Questions for authors
  • How did you choose saturation and blur as the two shifts under which you evaluate the VLMs? There are many other image-level augmentations to apply that do not modify the target properties (e.g., crops, rotations, translations, pixelization) and could serve as additional experiments.

  • Why did you choose to use the same reader (linear layer or MLP) across all models, and does that play a role in the results of Table 1? Each of the models tested probably utilizes a different latent dimension for its encoding blocks, and these discrepancies could affect the prediction accuracy for the Fixed condition. If Qwen is using a much bigger latent dimension, then the MLP could more easily overfit to the relatively small training set, explaining the drop in performance when using the fixed condition.

  • Did you try not pooling the visual tokens and instead using the CLS when it is available? Does that improve the performance under shifts, considering that the CLS is probably trained to ignore some of those shifts (through self-supervised training)?

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

This paper introduces VISTA, a controlled experimental framework for studying how category-preserving perceptual stress affects frozen vision-language models. It uses two axes: Color via chroma reduction and Shape via Gaussian defocus. The authors found severe clean-readout mismatch is largely corrected by a class-agnostic translation, while paired representation discrepancies remain. Color stress reduces native true-class evidence and candidate accuracy, whereas Shape changes continuous evidence in model-dependent directions while endpoint decisions remain near ceiling. Recipient stress modulates fixed visual contrasts in opposite directions for Qwen–Color and LLaVA–Shape, with cross-family and finite-interval consistency and natural-direction amplitude advantages over random and orthogonal controls.

Critical strengths

The paper has a very precise question to be asked: what happens inside a VLM when a visual category remains decodable but the model’s own behavior changes? The matched stimulus families hold category and scene configuration fixed while varying a single perceptual parameter. This within-family design is a genuine methodological advance over standard corruption benchmarks. The translation-correctability result is elegant and well-controlled.

Critical weaknesses

The “opposite modulation” finding rests on two model–attribute cells. This is rather limited support. The translation-correctability result, while clean, may be less surprising. For an affine reader, subtracting a class-agnostic mean shift is equivalent to adjusting the decision bias by −Wû_s, as the authors note in Appendix B.5. The finding therefore demonstrates a large bias component in the clean–stress mismatch, useful and worth reporting, but does not by itself reveal a structural property beyond “the stressed mean is displaced.” Some quantitative claims need tighter reporting. The abstract states “translated Shape-L2 accuracy reaches 0.979–1.000.” This range mixes IID evaluation with Confirmation results and is not accompanied by intervals. Similarly, the discrepancy-reduction ranges are reported without per-checkpoint values or uncertainty. Given the paper’s otherwise strong statistical discipline, these summary ranges feel under-specified. The paper does not report free-generation behavior at scale. Candidate scoring with a fixed five-way answer space is a controlled and appropriate choice for the evidence analysis, but it does not establish that the observed evidence changes matter for open-ended behavior. Also it would be quite helpful if the authors test some late versions of the MLLMs to see where things stand.

Questions for authors

Can you provide a forest plot or summary table across all six model–attribute cells and both margins, with intervals, to make the evidential balance for the opposite-modulation claim transparent at the main-text level? Can you report sum log-probability or per-token margins as a robustness check for Shape, particularly to address the InternVL–Shape sign discrepancy?

Reviewer_Wg6tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propsoes VISTA, a controlled experimental framework for examining how visual perturbations affect internal representations and their influence on VLM answers. VISTA studies how VLMs respond to synthetic color and shape perturbations. Probing uses eight checkpoints, and the native-evidence and intervention analyses use three. The paper reports a probe mismatch that a translation can correct, answer changes that depend on the attribute, and recipient-state effects that vary across settings.

Critical strengths

The experimental comparisons are well controlled and cover multiple model families.

  • Image pairs keep the category and scene layout fixed while changing color saturation or shape blur. This helps isolate the effect of the tested perturbation.
  • Probes are trained and tested on separate scaffold groups. This reduces the risk that performance reflects memorization of shared scene layouts.
  • Probe predictions and the model’s own candidate scores are compared on the same examples. This helps distinguish recoverable category information from the model’s ability to answer correctly.
  • Probing covers eight checkpoints from three model families, while the native-evidence and intervention analyses cover three representative checkpoints.

The translation experiment provides a concrete explanation for part of the probe failure.

  • The correction uses a mean shift estimated from training pairs. It applies the same shift to every class and keeps the clean-trained probe fixed.
  • After correction, shape-probe accuracy reaches 97.9–100% across all eight checkpoints on the IID evaluation
  • This shows that failure of a clean-trained probe does not necessarily mean that category information has been lost. A simple shift in the representations can explain much of the failure.
  • The paper also measures the representation differences that remain after correction. This avoids treating restored probe accuracy as evidence that the representations have become identical.
Critical weaknesses

The translation finding provides limited insight beyond probe calibration.

  • The correction is specific to each stress condition. Eq1 estimates a separate mean shift for each stress level. For a linear probe, applying this translation is equivalent to adjusting its bias.
  • Probe failure does not necessarily reflect failure of the model itself. On Confirmation Shape L2, Qwen3.5-2B’s clean-trained probe has a balanced accuracy of 0.200, while the model’s native candidate accuracy is 1.000 (Tables 1 and 2).
  • The broader explanatory value remains unclear. Translation is a useful readout check, but the paper does not show that the correction transfers across stress conditions or explains changes in native predictions.

The controls do not establish that the intervention effects are specific to perceptual stress.

  • A nonzero interaction does not by itself identify a stress-specific mechanism. A nonzero \delta_\tau in Eq 6 shows that target-ROI tokens and the remaining tokens do not act additively. Other structured activation changes could also produce such interactions.
  • The direction controls address amplitude rather than stress specificity. Random and orthogonal controls test the size of the effect. They do not establish whether task-irrelevant natural changes or alternative token partitions of the same size would produce similar interactions.
  • Agreement across constructions is useful but does not provide independent replication across recipient samples. All three constructions reuse the same recipients.
  • The strongest opposite-sign results do not isolate the role of model or attribute. Qwen–Color and LLaVA–Shape differ in both respects. In the other four settings, the evidence depends on the margin or intervention construction.
  • The uncertainty estimates do not account for the full set of comparisons. All intervals are pointwise. They do not provide simultaneous coverage across the tested settings and analyses.
Questions for authors
  1. Does a translation fitted at one stress level transfer to unseen stress levels or perturbations without refitting?
  2. How do the recipient-state interactions compare with those from matched, task-irrelevant natural changes and from token partitions of the same size?

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

Related papers

Discussion (0)

Sign in to comment.