acceptodds
Withdrawn from ICLR 2027

Vision-Language Models Can Report Which Image's Visual-Token Activations Were Perturbed

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can a vision-language model (VLM) report which of two images had its internal visual-token activations perturbed while the visible inputs themselves remain unchanged? We study this question in three instruction-tuned VLM checkpoints by perturbing decoder-side activations at candidate-associated visual-token positions and asking the model which candidate was affected. Because successful localization could simply reflect damaged visual information, we separately test whether measured visual functions remain intact and use controls to probe simpler explanations. At operating points satisfying both the category criterion and a secondary counting guard, localization rises from \(0.499\) to \(0.688\) in Qwen2.5-VL and from \(0.505\) to \(0.656\) in Qwen3-VL, with both effects surviving whole-window multiplicity correction. Idefics3 shows a smaller but statistically detectable effect. Norm-restored and crossover controls show that activation-norm inflation and fixed display-side preference are not sufficient explanations, while matched-quality and visible-corruption comparisons bound stronger perturbation-specific interpretations. Finally, in Qwen3-VL, perturbation identity is near-ceiling linearly decodable at both \(p=.30\) and \(p=.60\), yet the fitted probe directions have sharply different steering efficacy: little targeted steering at \(p=.30\) but material bidirectional steering at \(p=.60\), where the counting guard no longer holds. We call this phenomenon perturbation reportability: candidate-specific internal perturbation information can become behaviorally reportable, while linear availability and steering efficacy remain distinct under the tested intervention procedure.

Primary area:
interpretability and explainable AI
Keywords:
vision-language models, mechanistic interpretability, internal representations, activation perturbations, linear probing, causal interventions, representation analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1PD7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper extends the ‘candidate-specific perturbation reporting’ paradigm developed by Fornasiere et al. (2026) for LLMs to VLMs. The experiment was designed such that a perturbation was applied to one of two images, and the model was asked to identify which of the two had been perturbed. The paper ensures image quality by category and counting. Finally, the paper does not provide sufficient support for the hypothesis it puts forward. Experimental results are reported on Qwen2.5-VL, Qwen3-VL and Idefics3.

Critical strengths

This article is honest, and its claims are restrained.

Critical weaknesses

(1) This paper cannot rule out another important explanation (the degradation of image representation); therefore, it cannot substantiate the claim made in the paper: that visual language models can identify which image’s visual features have been perturbed.

Also, category classification and counting cannot sufficiently guarantee that the model’s visual functions are functioning properly.

(2) The steering analysis offers limited insight and novelty. The observation that linearly solvable probes of equal accuracy exhibit different steering effects is not theoretically surprising and does not introduce any new concepts, as linear solvability does not imply causal validity.("Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?", ACL 2021)

The authors also acknowledge that it fails to demonstrate a native monitoring mechanism, consequently, this set of experiments makes little contribution to the paper’s central research question.

Questions for authors

Let the models report which one has undergone perturbation after controlling for the degree of feature degradation in the two images. Only if the model still significantly outperforms chance levels is there strong reason for substantiating the claim made in this paper.

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

This paper studies whether VLMS can tell which of two images has had its internal representations disturbed, without changing the images themselves. The authors test three models and examine both their ability to identify the affected image and their performance on visual tasks. The two Qwen models demonstrate improved identification of affected images compared to random chance, while effectively preserving object recognition and counting accuracy. Further experiments investigate possible explanations for this behavior and whether modifying internal representations can influence the answers of models.

Critical strengths

The main findings are supported by well-controlled experiments.

  • The experimental design is well controlled. The perturbed-image label is independent of the correct answer to the visual task.
  • The crossover tests are convincing. They show that a fixed preference for A or B cannot fully explain the results.
  • The norm-restoration control is useful. It shows that increased activation magnitude is not required for localization.
  • The main Qwen findings are supported by robustness checks. They survive multiple-comparison correction and replicate on independently sampled trials.

The conceptual distinctions are helpful.

  • The paper separates behavioral reporting, linear decoding, and steering, which helps readers understand what each experiment shows and what it does not establish.
Critical weaknesses

It is unclear what cues the model uses to identify the perturbed image.

  • The model may identify the perturbed image by detecting reduced representation quality. The quality-question control is consistent with this explanation. If so, successful localization would not establish an ability beyond general quality assessment.
  • The category and counting tests cover limited visual abilities. Passing them does not rule out degradation in other abilities, such as spatial reasoning, fine-grained attribute recognition, and reading text in images.
  • Although the authors acknowledge this possibility, the experiments do not establish whether quality degradation also explains localization when the tested visual abilities are largely preserved.

The probe and steering experiments provide limited insight into how the model identifies the perturbed image.

  • At p=0.30p=0.30, a separately trained linear classifier identifies the perturbed image with 98.8% accuracy, while qwen3-VL itself achieves about 66%. However, this classifier is an external readout. Its success does not establish that the model uses the same information in the same way when generating its answer.
  • At p=0.60p=0.60, steering changes 20.1% of answers at alpha=2. However, counting accuracy has already fallen from 96.4% to 62.5% (Table 1). This result therefore does not explain how the model identifies perturbations when the tested visual abilities are largely preserved.

The presentation requires substantial effort to follow. I spent about 2-3 hours reading the paper and still found several parts difficult to understand, particularly the dense terminology and distinctions between experimental conditions. My concern is the clarity of the resulting text. This difficulty may partly reflect my own background, but more intuitive explanations and concrete examples would make the paper easier to follow.

Questions for authors
  • Why is preservation assessed using point estimates rather than a confidence-bound criterion? In Table 5, the Qwen2.5-VL counting ratio has a 95% confidence interval of [0.949, 1.004], which crosses the 0.95 threshold. How should this uncertainty affect the interpretation of this setting as function-preserving?

  • The probe directions at p=0.30p=0.30 and p=0.60p=0.60 are fitted separately. Have you tested each direction in both perturbation conditions? Do they suggest that the difference in steering efficacy comes from the fitted direction, the model’s activation state, or both?

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

This paper investigates whether vision-language models can identify which of two images has undergone an internal visual-token activation perturbation while the input images remain unchanged. The authors introduce a candidate-specific perturbation framework using dropout and Gaussian noise, together with behavioral reporting, visual-function preservation tests, and several control experiments. They evaluate three VLMs (Qwen2.5-VL, Qwen3-VL, and Idefics3) and find above-chance perturbation localization, including settings where measured visual capabilities remain largely preserved. Additional experiments examine alternative explanations through activation normalization, candidate swapping, and visible corruption. Finally, linear probing and activation steering analyses reveal that perturbation information can be linearly decodable without necessarily being effective for steering model behavior.

Critical strengths

The paper investigates an interesting research question concerning whether VLMs can identify perturbations to their own internal visual representations. A particular strength is the careful experimental design, which incorporates multiple controls to distinguish perturbation reportability from simpler explanations, including activation-norm changes, response-side biases, and degradation of visual information. The authors also explicitly evaluate whether visual capabilities remain preserved under perturbation, making the findings more interpretable. Together with the statistically significant above-chance localization results across three VLMs, these experiments provide meaningful empirical evidence of internal perturbation reportability while maintaining appropriate boundaries on its interpretation.

Critical weaknesses

1. Limitations of the experimental setup and construct validity. The central finding remains difficult to interpret because the experiments do not sufficiently distinguish perturbation reportability from detecting degradation in visual representations. The paper's own controls show that representation-quality judgments can perform comparably to or better than perturbation identification, while visible corruption can produce similar behavior. Furthermore, visual-function preservation is assessed using relatively limited category recognition and counting tasks, with counting introduced as a post-hoc criterion. Consequently, it remains unclear whether the models genuinely identify internal perturbations or simply exploit changes in representation quality that these evaluations do not fully capture.

2. Limited novelty and mechanistic evidence. The experimental framework closely follows prior work on activation perturbation reporting in language models, making the novelty primarily an extension to the visual modality. While the additional probing and steering experiments are interesting, they provide limited further mechanistic insight. Perturbation identity is almost perfectly linearly decodable even at weaker perturbation strengths, but meaningful causal steering is observed only under stronger perturbations where counting performance has already substantially degraded. Therefore, the experiments do not establish comparable causal effectiveness under conditions where measured visual functionality remains preserved, limiting the depth of the contribution beyond the existing perturbation-reporting framework.

Questions for authors
  1. How do the authors distinguish genuine perturbation reportability from detecting degraded visual representations, given the comparable performance of representation-quality judgments?

  2. Are category recognition and counting sufficient to establish that fine-grained visual information remains preserved?

  3. Why does effective causal steering emerge only at stronger perturbation strengths, where counting performance has already substantially degraded?

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

Related papers

Discussion (0)

Sign in to comment.