acceptodds
Withdrawn from ICLR 2027

Same Causal Question, Different Lenses: Evidence-to-Answer Dependence in Vision-Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A vision–language model can answer a question correctly without using the image region that contains the answer. We introduce a region-anchored audit that connects visual evidence to answer-side computation and tests how much of that effect a proposed internal description reproduces. On compact-evidence questions, we compare answer-region masking with same-area control masking, restore clean hidden states during the masked run, and test candidate descriptions through paired clean-state restoration and masked-state writeback. Across Gemma3, Qwen2.5-VL, and LLaVA1.5, answer-region masks cause greater target-score loss than matched controls, and aligned hidden restoration selectively recovers that loss. At the tested interfaces, fitted sparse reconstructions, selected native channels, and example-shared low-rank subspaces do not reproduce their corresponding complete-reference effects. In contrast, a frozen LLaVA complete hidden-state interface restores 95.8% and removes 95.0% of the mean natural clean-to-mask score change on 60 examples with result-blind-reviewed masks, passing all four paired-bootstrap bounds. The audit thus identifies a shared visual answer dependence while distinguishing evidence-sensitive internal structure from a description that faithfully reproduces its measured answer effect.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision-language models, mechanistic interpretability, causal tracing, activation patching, sparse transcoders, visual grounding, visual question answering

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_Y5YpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether internal interventions in a VLM can reproduce the change in its answer score when visual evidence is masked. It compares evidence masks with same-area controls, then swaps internal states between clean and masked runs in both directions. On selected OK-VQA examples, the tested sparse reconstructions, selected channels, and PCA subspaces fail their respective criteria across Gemma3-4B, Qwen2.5-VL-7B, and LLaVA-1.5-7B. A full hidden-state intervention at LLaVA layer 30 passes a separate confirmation on 60 reviewed masks, recovering 95.8% and removing 95.0% of the mean clean-to-mask score change.

The evaluation is careful and the effect-matching criterion is useful. I lean toward rejection because the strongest confirmed finding is a full-state match at one late LLaVA layer, while missing reference results limit the broader compact-versus-complete comparison. The confirmation also leaves the broader evidence-specific interpretation unresolved. I could move toward acceptance if the missing results substantiate the compact-versus-complete comparison. A control analysis on the confirmation set would strengthen the broader evidence-specific interpretation.

Critical strengths
  1. The evaluation separates selection from confirmation. The LLaVA intervention and decision rule are frozen before testing on 60 masks reviewed without model outputs. All four bootstrap bounds pass (Table 10), and a post-hoc tolerance sweep passes at the tested tolerances from 15% to 30%, but fails at 10% (Appendix D.6).

  2. The tests check whether interventions reproduce the target effect rather than simply change the answer score. Gemma's positive-only channel selection substantially overshoots the full-channel reference (Table 8). The reconstruction checks also help explain the sparse tools' failures, and the authors appropriately limit their conclusions to the tools tested.

Critical weaknesses
  1. Table 6 omits the complete feed-forward effect at the same intervention sites. All Qwen and LLaVA sparse and residual effects are within ±0.022 nats, but without this reference it is unclear whether they miss a substantial answer effect. In Table 8, the positive-and-negative groups have mean effects within the matching interval. Gemma fails the local-link test, but Qwen passes it; the quantitative results explaining Qwen’s remaining effect-matching and outside-channel failures are missing. The initial full-state references for Gemma and Qwen themselves fail the matching test (Table 9), limiting what the subspace failures at those interfaces establish. Table 15 strengthens the PCA result by repeating the tests at matched late-layer interfaces. LLaVA remains the clearest case supporting the compact-versus-complete comparison.

  2. The early experiments compare evidence masks with same-area controls, while the later matching tests use the clean-to-evidence-mask change (Appendix C.3). The abstract describes the 95% result correctly, but that result does not show whether the intervention reproduces the additional effect of masking evidence rather than comparable image content. A control-mask comparison on the 60 confirmation examples would connect these two parts of the paper.

Questions for authors
  1. Could you report the complete feed-forward effects at Table 6's exact sites and the estimates and uncertainty behind Table 8's failed criteria? Please also give the mean clean-to-mask score changes underlying Tables 3 and 9 so the reported ratios can be assessed.

  2. Is the patch applied at the input or output of layer 30? For the four pre-answer positions in Table 10, how much of the natural score change, restoration and writeback comes from the first answer token? Do the same four positions pass at earlier layers?

  3. On the same 60 examples, could you compare evidence-minus-control restoration and writeback with evidence-minus-control score loss, using paired uncertainty?

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

This paper investigates how localized patches of visual evidence impact model responses. The authors mask human-annotated “answer-bearing” regions from OK-VQA that are critical for answering the question correctly. They compare the impact of ablating these regions against random masks with the same area. Next, they perform bidirectional activation patching and use sparse transcoders, MLP channels, and low-rank subspaces to investigate which internal representations account for the resulting answer-score changes. They find that masking answer-bearing regions causes greater answer-score loss than masking random regions. They also find that restoring complete hidden states can recover this loss, while their tested sparse and low-rank representations generally cannot reproduce the same effects.

Critical strengths
  • The general question of how VLMs use localized visual evidence is interesting. The area-matched ablation is a principled choice and OK-VQA is a good starting place for this task since you have human-annotated regions.

  • I like the use of bidirectional activation patching to investigate how visual evidence contributes to answer generation.

Critical weaknesses
  • I have serious concerns about the writing and presentation of this paper, including the possibility that it is largely AI generated. I am concerned that this paper may have been written and designed using an AI research agent. I cannot verify this, and the use of AI is not itself my concern. My concern is that the paper reads like it has not undergone sufficient human review. The writing is extremely verbose yet does not communicate information clearly. There is substantial jargon throughout the abstract and introduction that makes a relatively simple experiment unnecessarily difficult to understand. For example, the authors introduce terms like "dependence-visibility gap" without clearly explaining what they mean. The plots also have significant presentation issues. Figures 5 and 6 contain overlapping numerical values, and some labels appear cut off that make me think an AI agent generated the plots without human review. These are basic issues that should have been caught by inspecting the figures before submission. I want to emphasize that even if the paper was entirely human-written, I think substantial work is needed to improve the writing, presentation, and scientific contribution.

  • I do not think this paper has a sufficiently interesting research question or contribution. Ultimately, this paper presents one relatively straightforward experiment. The authors mask image regions that humans have annotated as containing the answer, compare this against masking random regions, and perform activation patching on the resulting model states. The finding that masking the answer-bearing region is more impactful than masking random regions is entirely expected. I think there are much more interesting experiments that could be done with this setup. For example, apply the masking experiments to a dataset with strong world-knowledge priors, such as Visual CounterFact [1]. This would help investigate how much information comes from the image versus learned knowledge. It would also be interesting to investigate partial masking. For example, in Figure 1, a human could likely answer the question by seeing only a single word of text in the image. Can VLMs do the same? Are particular regions within the annotated area more important than others?

  • The choice of mechanistic analysis is not well-motivated. I do not understand why the authors focus on MLPs and hidden states when the task is fundamentally about retrieving specific information from image regions. Why not investigate attention heads? For a task like OK-VQA, where the model needs to retrieve information from specific visual regions, attention seems like a much more principled choice compared to MLPs.

  • Why do the authors not report VLM accuracy on the original images, images with the human-annotated regions masked, and images with random regions masked? This seems like one of the most basic experiments to include. Figure 2 reports additional answer-score loss. I would much rather see how many questions the models actually answer correctly under each masking strategy. I also think these results would vary significantly by task. For example, questions that can be answered using language priors should respond very differently to tasks where the model needs to read specific visual content.

  • I do not think this paper meets the bar for ICLR. The main result is largely obvious, and I do not think the mechanistic analysis provides a particularly meaningful explanation of how VLMs retrieve visual information.

[1] Golovanevsky et al. 2025. Pixels Versus Priors: Controlling Knowledge Priors in Vision-Language Models through Visual Counterfacts.

Questions for authors
  • Why did you perform analysis on MLPs and not attention heads?

  • I would like to see this extended to partial masking and other tasks.

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

The paper asks a simple question: when a VLM relies on a specific image region to answer, which internal description actually accounts for that reliance? The authors mask the human-annotated answer region in OK-VQA images and compare it to masking a random region of the same size. They then restore clean hidden states into the masked run to see where the lost score can be recovered. Finally, they test whether several compact descriptions (transcoder features, selected MLP channels, low-rank subspaces) can reproduce the effect in both directions: restoring it when written into the masked run, and removing it when written into the clean run.

On Gemma3, Qwen2.5-VL, and LLaVA-1.5, masking the answer region hurts more than masking a random region, and restoring hidden states recovers that extra loss. None of the compact descriptions reproduce the effect. Only the full hidden state at LLaVA layer 30, just before the answer, passes on a held-out set of 60 examples (about 95% in both directions).

Critical strengths
  1. Rigor: Matched control masks, restoration controls, two-direction tests, a frozen interface, and masks reviewed without seeing model results. This is more careful than most VLM interpretability work, and since the paper’s main contribution is a standard for evaluation, that care matters.

  2. A useful negative result on sparse tools. It shows clearly that an interpretable-looking feature structure does not mean the features explain the model’s behavior. This is a relevant warning for VLM circuit-tracing work.

Critical weaknesses
  1. The main positive result is expected by construction. The only thing that passes is the full hidden state at the second-to-last layer, right before the answer. The answer is computed directly from those states, so restoring them should recover nearly everything. This says little about how the model uses visual evidence, yet the abstract and conclusion present it as a key finding.

  2. The compact candidates are set up to fail. The sparse reconstruction subtracts two separately truncated reconstructions, giving error worse than doing nothing. The native channels are picked per image using the correct answer. The subspaces are fit on only 40 images. Their failure tells us about these specific choices, not whether a compact explanation exists, so the “dependence-visibility gap” claim goes beyond the evidence. The small sample sizes, a reused 240-image set, and a post hoc three-model comparison add to this.

Questions for authors
  1. Does the full hidden state pass at most late layers in all three models? If so, the LLaVA result reflects how residual networks work, not a specific finding. A layer sweep on the 60-mask set would settle this.

  2. Is the sparse reconstruction error caused by picking top-K features separately for the clean and masked runs? What happens with a shared feature set or no truncation?

  3. On the 60-mask set, only 29 of 60 Gemma examples show a positive score drop. Does Gemma actually rely on the annotated region here, and is its reversal driven by a few examples?

  4. At the layer used for the sparse test, how much does writing back the full feed-forward change recover? If very little, that layer can’t answer the question.

  5. How many layer and position choices were tried before picking LLaVA layer 30?

  6. Can you test native channels selected without using each image’s correct answer?

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

Related papers

Discussion (0)

Sign in to comment.