acceptodds
Under review as a conference paper at ICLR 2027

Encoded but Unused: Bounding the Sources of VLM Errors

Abstract

In this paper, we investigate why vision-language models (VLMs) produce errors by decomposing them into two sources: visual-representation failures, in which the visual evidence required for the correct answer is not encoded, and visual-utilization failures, in which this evidence is encoded but not successfully used. We propose to estimate the visual-utilization failure rate using lightweight probes and derive theoretical bounds on the true rate that account for probe imperfections. We introduce VELens (Visual Evidence Lens), a rank-based probe that traces visual evidence for both a target object and its ground-truth attribute, such as color, material, or shape, across VLM layers. We evaluate VELens on over 6,000 images using LLaVA-1.6-7B and Qwen2.5-VL-7B. Our method achieves ROC-AUC scores of 0.95 and 0.91, respectively, and outperforms alternative probes in both low-resource and unseen-attribute settings. Using it to diagnose errors, we estimate that visual-utilization failures account for 84.4% of LLaVA errors and 70.0% of Qwen errors, with conservative lower bounds of 77.5% and 61.0%, respectively. Overall, our study identifies visual utilization as the dominant bottleneck in VLMs and suggests that a central opportunity for improvement lies in developing better mechanisms to locate, route, and use the visual information they already encode.

Primary area:
interpretability and explainable AI
Keywords:
vision-language models, hallucination, interpretability, probing, logit lens, visual grounding, error analysis, multimodal models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_j4NROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether VLM errors arise because the required visual evidence is not encoded (representation failure) or is encoded but not used (utilization failure), and formalizes the share of utilization failures as an estimand with probe-based bounds. It proposes VELens, a logit-lens, rank-based probe that detects object and attribute evidence across LM layers. Evaluated on color, material, and shape questions from the VAW dataset with LLaVA-1.6-7B and Qwen2.5-VL-7B, it estimates that utilization failures account for 84.4% and 70.0% of errors, respectively.

Critical strengths
  1. The question is well posed and practically relevant: distinguishing missing from unused evidence indicates where improvements should come from.
  2. The paper proposes an interesting and intuitive approach to diagnosing the source of VLM errors.
  3. The rank-based probe is lightweight and remains effective in low-resource settings.
Critical weaknesses
  1. The connection between Assumption 1 and the noise-based experiment is not convincing. Under Assumption 1, the negative transformation ψ−\psi^- should act on the representation, producing inputs whose representations lack the required evidence, so that the probe's non-detection rate on X−X^- can serve as a proxy for representation failures. Gaussian noise does not guarantee this: after noising, some inputs may still encode the evidence (corresponding to utilization failures) while others do not, and this mixture is unknown and varies with λ\lambda. As a result, the lower bound is largely controlled by the noise level. Notably, λ=0.3\lambda = 0.3 is selected because the answer remains identifiable, yet if humans can still identify the attribute, the representation may still encode it as well. The authors should provide human verification of what the noised images retain, report the bounds as a function of λ\lambda, and consider counterfactual edits that remove the evidence by construction (e.g., recoloring the object). Although such edits change the answer, the proof of Theorem 1 relies only on Assumption 1, not on answer preservation.

  2. The estimate μ^ϕ\hat{\mu}_\phi lacks the controls needed to interpret it. The paper should report the probe's detection rate on correctly answered samples, where the evidence is almost surely encoded, as an estimate of the probe's sensitivity. In addition, μ^ϕ\hat{\mu}_\phi depends on the probe thresholds, yet these thresholds are not reported and no sensitivity analysis over them is provided.

  3. The practical claim that the diagnosis can guide error mitigation is not actionable. The probe evaluates M(X,YX)M(X, Y_X), i.e., evidence for the ground-truth answer, and relies on the annotated bounding box to select visual tokens; neither is available at inference time. Thus, per-instance diagnosis and the targeted attention intervention cannot be applied in practice.

  4. The experiments are insufficient to support the general conclusion. The study covers only two 7B models, one dataset, three local attributes, and a multiple-choice format, but concludes that utilization is "the dominant bottleneck in VLMs". Supporting this claim would require larger and more recent models, additional datasets, and open-ended question formats; otherwise, the conclusion should be scoped to the evaluated setting.

Questions for authors

Please refer to the weakness.

2 comments
AuthorsOpenReview ↗

comment 1/2

Dear Reviewer j4NR,

Thank you for the clear and constructive comments. We are glad that you found the research question relevant, and VELens’s intuitive, low-resource approach.

Regarding the lower bounds, Weakness 1; Question 1

We agree that Gaussian noise does not guarantee complete evidence removal. However, our analysis does not require this. As stated in §3.2, ψ−\psi_{-} is “designed to make the visual evidence relevant to the query less salient,” rather than eliminate it from every representation. Appendix A formalizes how the transformed probe behavior provides a conservative reference; if evidence remains detectable, the resulting lower bound becomes more conservative.

In the paper we selected 30% noise through manual inspection of validation examples: the target signal was visibly weakened while the object and attribute remained identifiable, whereas at higher noise levels the answer was often no longer recoverable. Following the reviewer’s suggestion, we conducted an additional human verification with five independent annotators on 50 randomly sampled images transformed using the paper’s 30% target-region Gaussian noise. For each image, annotators answered “What is the color of the [object]?” using four options: the ground-truth color and three distractors (25% chance). Mean accuracy was 90.0% (95% CI: 84%-95%), and a strict majority selected the correct color for 48/50 images (96%). Thus, the 30% transformation weakens the target-region signal while preserving the color answer in nearly all evaluated images.

To address the reviewer’s sensitivity concern, we also evaluated 10% noise. LLaVA's lower bounds remain above 50%50\% (59.2%59.2\%–76.2%76.2\% across attributes), supporting the stability of the majority-utilization conclusion. We added these results to Appendix D.4.

Counterfactual recoloring is complementary but changes the correct answer, falling outside our answer-preserving framework.

Regarding controls for interpreting μ^ϕ\hat{\mu}_{\phi}, Weakness 2

Following the reviewer’s suggestions, we add two controls: detection rate on correct answers and threshold sensitivity. As stated in Appendix D.4, both probes use threshold 0.50.5 (we additionally report below the results using thresholds of 0.4 and 0.6 demonstrating that our conclusions remain stable across these choices); we will clarify that it is used for all estimates and bounds on clean and transformed images.

Detection on correct answers and errors:

ModelVLM outcomeAttributeObjectBoth MM
LLaVACorrect94.5%97.7%92.5%
LLaVAError (HH)86.1%98.0%84.4%
QwenCorrect87.2%91.0%80.3%
QwenError (HH)77.3%89.8%70.0%

Correct answers are an approximate sensitivity control because they can also result from guessing or prior knowledge. Joint detection is 8.1 points lower on errors for LLaVA and 10.3 for Qwen, mainly due to the attribute probe; object detection changes little.

Sensitivity to the probe thresholds

We varied both thresholds over {0.4,0.5,0.6}\{0.4, 0.5, 0.6\}, applying the same thresholds to the clean, X+X_{+}, and 30%-noise images. Each cell reports the utilization failure estimate [lower bound, upper bound].

LLaVA

tattr\tobjt_{\mathrm{attr}} \backslash t_{\mathrm{obj}}0.40.50.6
0.490.9% [83.7%, 97.7%]90.2% [82.4%, 96.9%]89.9% [82.1%, 96.6%]
0.584.9% [78.3%, 96.0%]84.4% [77.5%, 95.4%]84.0% [77.3%, 95.0%]
0.676.2% [70.0%, 96.0%]75.8% [69.6%, 95.5%]75.5% [69.4%, 95.1%]

Qwen

tattr\tobjt_{\mathrm{attr}} \backslash t_{\mathrm{obj}}0.40.50.6
0.478.2% [69.3%, 99.7%]76.0% [67.0%, 98.0%]74.7% [65.7%, 98.8%]
0.572.0% [62.7%, 98.2%]70.0% [61.0%, 96.6%]68.7% [59.8%, 96.9%]
0.664.9% [54.5%, 99.0%]63.1% [53.2%, 97.3%]62.0% [52.4%, 97.6%]

The conclusion is stable across all nine threshold settings: the lowest pooled lower bound is 69.4% for LLaVA and 52.4% for Qwen. In the revised paper, we clarify in §6.2 and Appendix D.1 that the fixed 0.50.5 threshold is used for all calculations. We also add the threshold-sensitivity analysis as Tables 7–8 in Appendix D.4, highlighted in purple, holding the probes and transformations from Figure 4 fixed while varying only the thresholds.

AuthorsOpenReview ↗

comment 2/2

Regarding the actionability of the proposed diagnosis and intervention, Weakness 3

VELens is designed for annotated evaluation data. As §6.3 states, the attention intervention is a proof of concept testing whether evidence-positive errors are easier to correct by attending to the target region (compared to evidence-negative). The ground-truth answer identifies errors and defines the tested evidence, while the bounding box selects the target-object tokens (at inference time, target bounding boxes can be easily obtained using segmentation models such as SAM 3); the intervention does not supply the correct attribute value.

Evidence subsetRegionCorrection rate
Evidence-positiveTarget25.2%
Evidence-negativeTarget12.0%
Evidence-positiveRandom matched-size6.5% ± 0.7%

Evidence-positive errors are corrected more than twice as often as evidence-negative errors. For the random control, we sampled five matched-size regions per image on the same evidence-positive errors and obtained only 6.5%±0.7%6.5\%\pm 0.7\% correction. This shows that the 25.2% correction rate on the evidence-positive set is not incidental: directing attention specifically to the region where VELens detected the evidence is substantially more likely to correct the answer than directing it elsewhere in the image. The random-region control is added in blue to Appendix D.6.

Regarding the empirical scope of our conclusions: Weakness 4

Our original evaluation examines two models across three distinct attributes using 6,074 natural images from VAW, a widely used and diverse dataset, and we now extend it to Qwen2.5-VL-72B-Instruct. The same qualitative conclusion holds across every evaluated model and attribute. We believe this consistency provides meaningful evidence about current open-weight VLMs in attribute QA, while agreeing with the reviewer that it does not establish the same conclusion for other datasets, tasks, or question formats. We will therefore revise the abstract, discussion, and all statements of our conclusion to explicitly specify the evaluated setup.

To broaden model scale, we evaluated Qwen2.5-VL-72B-Instruct (∼73.4\sim 73.4B parameters) on the same examples, splits, questions, probes, and thresholds as Qwen2.5-VL-7B and LLaVA. Both share the same vision encoder design, while 72B has a much larger language decoder.

MetricColorMaterialShapeOverall
Attribute AUC, 72B0.8990.8530.7340.849
Object Attention AUC, 72B0.9700.9760.9830.976
Error rate, 72B23.4%21.4%31.8%24.9%
Error rate, 7B24.4%24.2%41.8%28.8%
Utilization failures, 72B60.7%64.3%64.6%63.0%

Without reconfiguration, VELens achieves overall attribute/object AUCs of 0.849/0.976. The 72B error rate falls from 28.8% to 24.9%, and its utilization-failure share from 70.0% to 63.0%, consistent with the larger decoder using evidence missed by 7B. The central result therefore persists at a substantially larger model scale, while VELens also captures the expected improvement in evidence utilization. We will add the 72B experiment to the revised paper. The corresponding Theorem 1 bounds are currently being computed using the same controlled transformations and will be added with the final results.

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

The paper formalizes a distinction between two sources of VLM errors on image-grounded questions: visual-representation failures, in which the required visual evidence is not encoded in the model's internal representations, and visual-utilization failures, in which the evidence is encoded but not successfully used when generating the answer. The authors cast the utilization rate μ_φ := P(φ_X = 1 | H) as a latent quantity (where H is the error event and φ_X ∈ {0,1} is the unobservable "evidence encoded" state) and approximate it with a probe M. Because any probe is imperfect, they prove Theorem 1: under a "probe dominance" assumption (Assumption 1), answer-preserving transformations ψ_+ (object-emphasizing) and ψ_- (object-blurring/noising) give a two-sided bracket 1 − P(M(X_-, Y_X)=0 | H) / P(M(X, Y_X)=0 | H) ≤ μ_φ ≤ P(M(X, Y_X)=1 | H) / P(M(X_+, Y_X)=1 | H). They then introduce VELens, a rank-based probe that projects every language-model layer's visual-token representation through the LM's output head W_V (logit-lens style), computes a per-token rank of target vs. competing vocabulary tokens, aggregates spatially inside/outside the target-object region and across layers via logistic regression. The composite probe is M = M_obj ∧ M_attr. On 6,074 VAW images with LLaVA-1.6-7B and Qwen2.5-VL-7B over color/material/shape, VELens reaches combined ROC-AUC 0.95 (LLaVA) and 0.91 (Qwen) and outperforms classifier-, confidence-, and attention-based baselines; it also generalizes to low-resource (5–10 labels/class) and unseen attribute values. Using it to decompose error sources yields μ̂_φ = 84.4% (LLaVA) / 70.0% (Qwen) with conservative lower bounds 77.5% / 61.0%. As a proof of concept, an attention intervention that boosts text-to-visual attention toward the annotated target region corrects 25.2% of VELens-detected (evidence-positive) errors vs. 12.0% of evidence-negative errors.

Critical strengths
  1. A clean conceptual decomposition with a non-trivial, actionable headline. The visual-representation vs. visual-utilization split is operationally meaningful: it directly maps onto two different research agendas (encoder improvements vs. LM-side routing / attention / decoding mechanisms). The paper puts bounded numbers on this decomposition rather than treating it as a conceptual stance, and the finding that in the evaluated setting utilization failures dominate by a wide margin would, if it generalizes, reshape where the field spends its optimization effort. The POC intervention adds a falsifiable test of the diagnosis — correction rates differ by 2× between the two VELens strata — which is more than a probing paper usually commits to.
  2. Bounds that explicitly account for probe imperfection. Most probing / interpretability papers report AUC and move on. This paper formalizes the gap between what a probe can measure and the latent property it tries to measure, derives Theorem 1 under a stated assumption, and uses two answer-preserving transformations as conservative bracketers. The whole construction — Assumption 1, the ψ_±-based bounds, and the subsequent treatment of μ̂_φ as a point estimate inside bounds — is more careful than the usual "we probed and found X% accuracy" style, and I think it should become a template for probe-based claims in this area.
Critical weaknesses
  1. Scope-vs-claim mismatch. The empirical base is 2 VLMs at 7B, 1 dataset (VAW), 3 attribute types (color/material/shape), 1 question template. The paper's framing claims — "utilization is the main bottleneck" (abstract), "a central opportunity for improvement lies in developing better mechanisms to locate, route, and use the visual information they already encode" (abstract) — are stated at the level of VLMs in general. These claims would need either (a) broader empirical scope (counting, OCR, spatial, chart; larger models; a non-VAW dataset) or (b) much tighter scope language in the abstract and discussion. As written, the paper risks over-claiming a finding whose demonstrated validity is confined to attribute QA on object-rich natural images with 7B open-source VLMs. The gap between LLaVA's 84.4% and Qwen's 70.0% within just this narrow slice already suggests substantial model dependence that is unexplored.
  2. The lower bound of Theorem 1 is only as good as the chosen ψ_-. The paper selects 30% Gaussian noise on validation. At 90% noise the ψ_- construction produces conditional detection 10.4% but an error rate of 74.7% — i.e., ψ_- degrades the input enough to shift the VLM into a very different behavioral regime, which stresses the "answer-preserving" premise of ψ_-. The reported 77.5% / 61.0% conservative lower bounds therefore inherit the fragility of the ψ_- choice: a slightly stricter noise would raise the lower bound, a slightly weaker one would lower it. A sensitivity table of μ̂φ bounds vs. ψ- strength would substantially strengthen the paper.
Questions for authors
  1. Could you report the Theorem 1 upper and lower bounds across a sweep of ψ_- noise strengths (e.g., {10%, 30%, 50%, 70%}) and ψ_+ emphasis strengths? The headline 77.5% / 61.0% lower bounds depend on a validation-selected transformation strength, and a sensitivity sweep would let the reader see the bounds as a function of the transformation rather than as a single point. If the lower bound stays above ~50% across the sweep, the "utilization dominates" claim becomes much more robust.
  2. The attention-intervention POC compares correction rates on evidence-positive (25.2%) vs. evidence-negative (12.0%) errors, both using the annotated target-region bounding box. What is the correction rate when the same attention intervention is applied to a random region of the image, on the same evidence-positive subset? That control would separate "VELens correctly identifies recoverable errors" from "any attention nudge on the gold region helps regardless of VELens prediction."
  3. Given the ~14-point gap between LLaVA (84.4%) and Qwen (70.0%) in utilization-failure share within just two 7B models, do you have any reason to expect the ordering (utilization > representation) to persist at larger scale, closed-source VLMs, or non-VAW tasks (counting, OCR, chart, spatial relations)? Even one data point on, e.g., Qwen2.5-VL-72B or on a counting benchmark would substantially raise external validity.
  4. Table 1's headline row is "Rank-Sep (M_attr) + Attention (M_obj)" — i.e., the final composite probe mixes VELens with the Attention baseline. Pure Rank-based M_obj reaches 0.84/0.81 vs. Attention's 1.00/0.96. Could you report (a) utilization-failure estimates (and bounds) using pure VELens for both components, and (b) using Attention for both components, so that the contribution of each probe component to the headline 84.4%/70.0% is isolated?
  5. The probe is trained on VAW labels and evaluated on P(M=1 | H). Is the probe's calibration on the error subset H similar to its overall calibration? If H-conditional calibration is biased (e.g., the probe tends to overpredict evidence on items where the VLM confidently answers wrong), that would systematically bias μ̂_φ.
  6. Assumption 1 is stated as a probe-dominance condition. Could you comment on when it might fail? E.g., if ψ_+ emphasizes the object with a visual cue (bounding box overlay) that the VLM has never seen in training, M(X_+, Y_X)=1 could be higher than the natural detection rate on true φ_X=1 cases, which would still satisfy Assumption 1's second inequality but would make the upper bound unhelpfully loose.
2 comments
AuthorsOpenReview ↗

comment 1/2

Dear Reviewer dDKM,

We thank the reviewer for the time and effort dedicated to this review. We particularly appreciate your positive assessment of our error decomposition, the bounds accounting for probe imperfections, and the attention intervention used to test the diagnosis.

Regarding the persistence of our conclusions at larger model scales, Weakness 1; Question 3

Our original evaluation examines two models across three distinct attributes on VAW, a widely used and diverse dataset. We now extend this evaluation to Qwen2.5-VL-72B-Instruct and obtain the same qualitative conclusion across all evaluated models and attributes. We agree that these findings should not be generalized to every image, task, or VLM, but their consistency provides meaningful evidence about current open-weight VLMs within the evaluated attribute-QA setting. Following the reviewer’s recommendation, we will explicitly state this scope whenever presenting our conclusions, including in the abstract and discussion.

To test whether utilization remains the larger error source at greater scale, we evaluated Qwen2.5-VL-72B-Instruct ($\sim$73.4B parameters) on the same examples, splits, questions, probes, and thresholds as Qwen2.5-VL-7B and LLaVA. The two models share the same vision-encoder design but have substantially different decoder sizes, enabling a controlled test of scaling.

MetricColorMaterialShapeOverall
Attribute AUC, 72B0.8990.8530.7340.849
Object Attention AUC, 72B0.9700.9760.9830.976
Error rate, 72B23.4%21.4%31.8%24.9%
Error rate, 7B (paper)24.4%24.2%41.8%28.8%
Utilization failures, 72B60.7%64.3%64.6%63.0%

VELens transfers without reconfiguration, achieving overall attribute AUC of 0.849. Its probe-based utilization estimate remains above 50% for every attribute (60.7% color, 64.3% material, 64.6% shape; 63.0% overall). Thus, the ordering requested by the reviewer (utilization >> representation) persists at 72B scale. Together with the paper's results, this gives a consistent qualitative ordering across LLaVA-7B, Qwen-7B, and Qwen-72B. The 72B model makes fewer errors in every category (28.8% to 24.9% overall), while its utilization-failure rate also declines in every category (70.0% to 63.0% overall). This is consistent with the larger decoder using some visual evidence the 7B model failed to utilize; utilization nonetheless remains the larger error source. Extending VELens to other model families, datasets, and tasks (counting, OCR, charts, spatial reasoning) is valuable future work. We will add the 72B experiment to the paper revision. Its Theorem 1 bounds are being computed with the same transformations and will be added with the final results.

Regarding the sensitivity of the Theorem 1 bounds to the transformation strength, Weakness 2; Question 1

Figure 4a reports the LLaVA–Color setting, showing clean evidence detection, detection after object emphasis (upper bound), and after 30% target-region noise (lower bound). The 10% condition provides the new sensitivity analysis. The 90% condition is diagnostic only: it makes the answer unrecoverable (Appendix D.4), so it cannot be used in Theorem 1. We chose 30% noise by visual inspection on validation because it weakened the signal while preserving the answer. We additionally tested weaker 10% noise:

ModelAttribute typeLower bound, 10% noiseLower bound, 30% noise
LLaVAColor76.2%82.2%
LLaVAMaterial72.0%74.7%
LLaVAShape59.2%71.3%
QwenColorIn progress59.3%
QwenMaterialIn progress60.9%
QwenShapeIn progress62.5%

For LLaVA, weaker corruption gives looser bounds, as expected, but all remain above 50% (59.2%–76.2%). The Qwen values are being finalized and will be available within the next few days; preliminary results show the same trend. We added this analysis and table to Appendix D.4 (blue), and will clarify in Figure 4a's caption that it reports LLaVA–Color and 90% noise is diagnostic only.

AuthorsOpenReview ↗

comment 2/2

Regarding attention intervention on random area; Question 2

Thank you for suggesting this control. We applied the same attention intervention to random regions on the same evidence-positive errors as the reported 25.2% result. For each image, we independently sampled five matched-size regions, keeping all other settings fixed, and report the mean correction rate across trials.

Evidence subsetIntervention regionCorrection rate
Evidence-positiveAnnotated target region25.2%
Evidence-positiveRandom matched-size region6.5% ± 0.7%

Target-region attention corrects 25.2% of errors versus 6.5% for random regions (18.7% less). This shows that the 25.2% correction rate on the evidence-positive set is not incidental: directing attention specifically to the region where VELens detected the evidence is substantially more likely to correct the answer than directing it elsewhere in the image. We added this control to the Attention Intervention appendix (blue).

Regarding utilization estimates with alternative probe combinations; Question 4

In Table 1, we selected Rank-Sep for MattrM_{\mathrm{attr}} and Attention for MobjM_{\mathrm{obj}} because this combination achieved the best joint ROC-AUC on both models. Following the reviewer’s suggestion, we use the two methods (our Rank-based probe and the Attention-based probe) to examine how each component affects the utilization estimate. We therefore evaluate the reported hybrid combination:

ModelMattrM_{\mathrm{attr}}MobjM_{\mathrm{obj}}Utilization estimateLower boundUpper bound
LLaVARank-SepAttention84.4%77.5%95.4%
LLaVARank-SepRank78.7%74.9%98.0%
LLaVAAttention-SepAttention66.3%43.1%86.3%
QwenRank-SepAttention70.0%61.0%96.6%
QwenRank-SepRank60.0%52.9%94.4%
QwenAttention-SepAttention63.8%44.9%96.0%

Rank-based probes for both components still yield majority-level conservative lower bounds (74.9% LLaVA, 52.9% Qwen), so the conclusion does not depend on the Attention object probe. Replacing Rank with Attention for the object component increases these bounds to 77.5% and 61.0%. Conversely, using Attention for the attribute component lowers them to 43.1% and 44.9%. Consistent with Table 1, Rank-Sep is stronger for attribute detection, while Attention improves object evidence detection component.

Regarding the probe’s calibration on VLM errors; Question 5

As discussed in the paper, the true visual-evidence state ϕX\phi_{X} is unobserved and M(X,YX)M(X,Y_{X}) is only a proxy. Thus, this analysis cannot establish directly whether the VLM encoded the evidence. We nevertheless test whether VELens behaves differently on model errors by comparing its evidence-detection rate on VLM-correct and VLM-error examples, using the same held-out set and decision threshold.

ModelVLM correct, P(M=1∣H=0)P(M=1\mid H=0)VLM error, P(M=1∣H=1)P(M=1\mid H=1)
LLaVA94.5%86.1%
Qwen87.3%76.9%

VELens detection decreases by 8.4 percentage points for LLaVA and 10.4 points for Qwen on the error subset, confirming that these examples are more challenging. Thus, the observed shift is toward detecting less evidence on model errors, rather than systematically overpredicting it. This decreases P(M=1∣H=1)P(M=1\mid H=1), making μ^ϕ\hat{\mu}_{\phi} more conservative. Moreover, Theorem 1 uses controlled transformations and probe behavior measured directly on the same H=1H=1 subset, rather than assuming that overall probe performance transfers unchanged to model errors.

Regarding potential failure modes of the probe-dominance assumption; Question 6

Probe dominance can fail if a transformation changes probe behavior for an unintended reason. For example, if ψ+\psi_{+} introduces an artificial cue that increases detection beyond its rate on natural ϕX=1\phi_{X}=1 cases, the relevant dominance inequality is violated and the upper bound may become artificially small, rather than merely loose. Similarly, if ψ−\psi_{-} suppresses detection because of out-of-distribution artifacts rather than weakened evidence, the lower bound may become artificially high.

Our ψ+\psi_{+} blurs only the background, leaving target-object pixels unchanged and adding no bounding box. For ψ−\psi_{-}, we chose 30% noise on validation to weaken evidence without destroying answer recoverability; we exclude 90% noise because it does. Figure 4a shows detection changing monotonically with transformation strength. These choices support, but do not guarantee, the assumptions: if either inequality fails, its bound is no longer theoretically guaranteed.

Reviewer_KY4qOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a significant research problem concerning error attribution in Vision-Language Models (VLMs). The methodology is well-conceived, and the experimental results offer valuable insights. However, the authors’ interpretation—that probe detection results imply visual information has been fully encoded by the model—relies on strong, unverified assumptions, particularly when subsequently estimating the rate of "visual utilization failure." Key aspects, such as the validity of the theoretical boundaries, the evaluation approach for joint probes, and the overall reliability of the error attribution, require more rigorous validation.

Critical strengths
  1. This paper moves beyond simply identifying whether a Vision-Language Model (VLM) has made an error; it further distinguishes between cases where visual information fails to be encoded and cases where it is encoded but not correctly utilized. This approach to error attribution provides more targeted guidance for future model improvements and sets this work apart from traditional research on hallucination detection and mitigation.
  2. VELens leverages the relative rankings of visual tokens—across different layers—within the vocabulary space, incorporating evidence related to objects and their attributes. Compared to training a classifier directly on hidden features, this design offers a more intuitive semantic interpretation and enables the analysis of how various visual concepts evolve across layers within the model.
  3. Experimental results demonstrate that "Rank-Separate" achieves strong ROC-AUC scores in tasks involving the detection of color, material, and shape across two different models. Furthermore, the authors conduct experiments under low-resource training conditions and test the model on previously unseen attribute values, confirming the method's robustness across varying levels of supervision.
Critical weaknesses
  1. Probes are trained under the supervision of ground-truth attribute labels; the ability to predict the correct attribute does not imply that the original VLM can directly recover and utilize this information from the corresponding hidden representations. Probes might instead rely on attribute priors, correlations with object categories, or other non-causal features. Consequently, current experiments demonstrate only that the correct attributes are, to some extent, decodable—not that the representations contain sufficient evidence to enable the model to arrive at the correct answer.
  2. The theoretical lower bound relies on the "Probe-Dominance Assumption," which is a strong assumption that lacks sufficient empirical validation.
  3. The ROC-AUC evaluation of joint probes fails to adequately verify the model's capacity to jointly detect evidence for both objects and attributes.
  4. The limited scope of the experimental tasks constrains the generalizability of the conclusions.
Questions for authors
  1. See Weakness1
  2. Can the authors provide empirical evidence to validate the "probe-dominance assumption" upon which the theoretical lower bound relies? Furthermore, to what extent would the theoretical conclusions be affected if this assumption does not hold?
  3. How does the ROC-AUC evaluation demonstrate that "joint probes" are capable of detecting evidence related to both objects and attributes simultaneously, rather than relying primarily on one or the other? Can the authors provide a more direct evaluation method to verify their true joint-probing capability?
  4. Could additional experiments be conducted across a more diverse range of tasks, datasets, and models?
2 comments
AuthorsOpenReview ↗

comment 1/2

Dear Reviewer KY4q,

We thank the reviewer for their thoughtful and encouraging feedback. We are glad that you appreciated our distinction between missing and unused visual evidence, the intuitive and interpretable design of VELens, and its robustness across models, supervision levels, and unseen attributes. Below, we address each of your comments and questions in detail.

Regarding whether probe decodability demonstrates sufficient internal evidence, Weakness 1; Question 1

We agree with the reviewer that a probe’s ability to decode an attribute does not by itself demonstrate that the VLM can recover and use that information. This reflects a broader challenge in interpretability: there are no ground-truth labels indicating whether a model’s internal representations contain sufficient evidence. We explicitly acknowledge this in lines 158–161: “In practice, however, MM is not trained on ϕX\phi_X, which is unobserved, but on labels indicating whether YY is correct for XX, so M(X,YX)M(X, YX) is only a proxy: it supports but does not prove ϕX\phi_X in either direction, as a positive prediction may rely on imperfect or spurious cues, and a negative one may fail to extract a present signal. To account for this uncertainty, we next derive bounds on μϕ\mu_\phi.”

Several results support this proxy: detection requires both object and attribute evidence in the target region, increases when that region is emphasized, and decreases when its visual signal is degraded (in §6.2 and Figure 4a). Moreover, attending to the target region corrects 25.2% of evidence-positive errors, compared with 12.0% of evidence-negative errors, showing that detected errors are more decodable (as shown in §6.3).

Our evidence is therefore not decodability alone, but the consistent behavioral controls and conservative lower bounds of 77.5% and 61.0%.

Regarding probe-dominance assumption, Weakness 2; Question 2

Directly verifying Assumption 1 would require observing whether ϕX\phi_X contains sufficient evidence—the unknown quantity we aim to estimate. We therefore test its observable implications. Figure 4a shows that detection increases from 86.5% to 88.8% when the target is emphasized and decreases to 43.0%, 23.4%, and 10.4% under increasing noise. This monotonic response supports, but cannot prove, the assumption.

If the negative-dominance condition fails, Theorem 1 no longer guarantees a conservative lower bound; the probe estimates and transformation results remain descriptive, but the theoretical conclusion weakens. Appendix A states this condition and its role explicitly.

AuthorsOpenReview ↗

comment 2/2

Regarding the joint detection of object and attribute evidence, Weakness 3; Question 3

Our composite probe follows directly from the questions: answering “What is the color of the car?” requires evidence for both the target object and its attribute. We define:

M(X,Y)=Mattr(X,Y)∧Mobj(X).M(X,Y)=M_{\mathrm{attr}}(X,Y) \land M_{\mathrm{obj}}(X).

The reported configuration uses Rank-Sep for MattrM_{\mathrm{attr}} and Attention for MobjM_{\mathrm{obj}}; Table 1 evaluates both separately before their combination. Their joint outcomes on model errors are:

ModelAttributeAttribute and object evidenceAttribute evidence onlyObject evidence onlyNeither
LLaVAColor86.5%2.4%11.1%0.0%
LLaVAMaterial84.1%2.5%12.7%0.7%
LLaVAShape81.3%0.0%18.7%0.0%
QwenColor67.9%14.1%16.7%1.3%
QwenMaterial73.2%4.7%15.0%7.1%
QwenShape69.5%3.0%26.3%1.2%

Across both models and all attributes, the object probe detects evidence for most errors, leaving attribute evidence as the component less often detected. This is expected because locating evidence for the target object is easier than distinguishing its specific attribute value from competing values in the same category. This difference emphasizes the importance of the joint criterion: object evidence alone is insufficient, so both the target object and its queried attribute must be detected.

Regarding the additional experiments, Weakness 4; Question 4

To evaluate VELens across a more diverse range of models and broaden our conclusions to a substantially larger scale, we evaluate Qwen2.5-VL-72B-Instruct ($\sim$73.4B parameters) using the same examples, splits, questions, probes, and thresholds as for Qwen2.5-VL-7B and LLaVA. Both Qwen models share the same vision encoder design, while the 72B model has a substantially larger language decoder.

MetricColorMaterialShapeOverall
Attribute AUC, 72B0.8990.8530.7340.849
Object Attention AUC, 72B0.9700.9760.9830.976
Error rate, 72B23.4%21.4%31.8%24.9%
Error rate, 7B (paper)24.4%24.2%41.8%28.8%
Utilization failures, 72B60.7%64.3%64.6%63.0%

Under the same setup, VELens achieves overall attribute and object AUCs of 0.849 and 0.976. The 72B model’s error rate decreases from 28.8% to 24.9%, while its utilization-failure rate decreases from 70.0% to 63.0%. This pattern is consistent with the larger language decoder successfully using encoded evidence in cases where the 7B model fails. Importantly, utilization failures still constitute the majority of errors for every attribute and 63.0% overall. Thus, the central finding persists at a substantially larger scale, while VELens captures the improvement in evidence utilization. We will add the 72B results and clarify the evaluated scope in the revised paper. The corresponding Theorem 1 bounds, computed using the same transformations, will be included with the final results.

Comments on the paper

2 comments

Dear Reviewers,

We are currently in the discussion phase. The authors have responded to the comments, and I would greatly appreciate it if you could take a moment to check the rebuttal and leave any follow-up comments or questions. If the response has addressed your concerns, please also consider updating your review accordingly.

Your feedback is very valuable to both the authors and the decision process. Thank you again for your help.

Best regards, AC

Dear Reviewers, we have responded to your questions and added new experiments and analyses. Please let us know if these revisions address your concerns or if any further clarification would be helpful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.