PaperlayerView exact revision on arXiv

Source-linked AI summary

Encoded but Unused: Bounding the Sources of VLM Errors

Anat Veikherman, Mor Ventura, Nitay Calderon, Lihi Zelnik-Manor, Roi Reichart

arXiv:2610.29241v1cs.CV

TL;DR

VLM errors may reflect missing visual evidence or evidence that is encoded but not used, a distinction that final answers alone cannot resolve. The paper estimates utilization failures with bounded, probe-based analysis and introduces VELens to trace object and attribute evidence across layers. It estimates that utilization failures account for most errors in both evaluated models.

  • Problem

    VLM error analyses often detect or mitigate incorrect outputs without determining whether required visual evidence was absent or encoded but unused, despite the distinction pointing to different interventions.

    Source: §1 · p. 1

  • Method

    The paper estimates utilization failures with a lightweight probe, derives bounds for probe imperfections, and introduces VELens to rank object and attribute evidence across VLM layers.

    Source: §1 · p. 2, §3.2 · p. 4, §4.1 · p. 5

  • Results

    84.4% of LLaVA errors and 70.0% of Qwen errors are estimated to be utilization failures, with conservative lower bounds of 77.5% and 61.0%, respectively.

    Source: §1 · p. 2, §6.2 · p. 9

  • Takeaways & Limitations

    Across the two studied models, visual utilization is the main bottleneck, suggesting that reliability improvement requires better mechanisms to locate, route, and use already encoded visual information.

    Source: §1 · p. 3, §7 · p. 10

    The demonstrated correction intervention assumes a known object region and is presented as a proof of concept, not a standalone correction method.

    Source: §6.3 · p. 10

Abstract

from arXiv · show

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.

1 INTRODUCTION

VLM errors can arise because required visual evidence is absent from representations or because encoded evidence is not used. The paper formalizes this distinction, estimates utilization failures with VELens, and finds they dominate errors in the two evaluated models.

1 INTRODUCTION

  • Visual-representation failures mean required evidence is not encoded, whereas visual-utilization failures mean it is encoded but not used.The distinction matters because the two failure modes point to different interventions: improving the visual encoder versus locating, routing, and incorporating encoded evidence during generation.Source: §1 · p. 1
  • Because probes can rely on spurious cues or miss present evidence, the paper derives bounds on the true utilization-failure rate using answer-preserving transformations that vary evidence salience.The study also reports that identifying encoded evidence can guide error mitigation.Source: §1 · pp. 2–3
  • VELens projects layer representations into vocabulary space and combines rank-based evidence for the target object and correct attribute to assess whether both are present.It is evaluated on object-attribute questions involving color, material, and shape.Source: §1 · p. 2
  • Across two VLMs and three attributes, VELens outperforms every baseline and achieves joint object-and-attribute ROC-AUCs of 0.95 on LLaVA and 0.91 on Qwen.It also outperforms corresponding baselines in low-resource and unseen-value settings.Source: §1 · p. 2
  • 84.4% of LLaVA errors and 70.0% of Qwen errors are estimated to be visual-utilization failures, with lower bounds of 77.5% and 61.0%, respectively.The paper finds no universal layer at which evidence becomes accessible; this varies by model and attribute.Source: §1 · p. 2

2 RELATED WORK

Prior VLM work detects or mitigates incorrect outputs and studies internal representations, information flow, and attention. This paper distinguishes itself by jointly analyzing object and attribute evidence across layers and separating missing evidence from encoded-but-unused evidence.

2 RELATED WORK

  • Prior error research mainly detects output conflicts or mitigates them during training and inference, while often addressing outputs without identifying their underlying cause.Xiong et al. (2026) analyze utilization but do not establish whether the required evidence was encoded; this paper distinguishes representation from utilization failures.Source: §2 · p. 3
  • Internal analyses use hidden-state classifiers, vocabulary-space projections, or cross-modal attention to study visual information and its influence on generation.Fu et al. (2025) show that information recoverable from intermediate vision-encoder layers may fail to influence predictions.Source: §2 · p. 3
  • Unlike per-layer or per-concept analyses, this work analyzes object and attribute evidence jointly across layers.Prior studies report that object semantics are encoded in visual tokens, become more interpretable across layers, and that attention disruptions correlate with errors.Source: §2 · p. 3

3 QUANTIFYING THE RATE OF VISUAL-UTILIZATION FAILURES

The paper estimates the proportion of VLM errors attributable to encoded-but-unused visual evidence, while accounting for the fact that this latent evidence state cannot be directly observed. For object-attribute questions, the operational probe requires both object and attribute evidence.

3 QUANTIFYING THE RATE OF VISUAL-UTILIZATION FAILURES

  • Visual-utilization failures are errors for which sufficient visual evidence is encoded but not used, while visual-representation failures lack that evidence.The final answer alone cannot distinguish these sources, although they call for different interventions.Source: §3 · p. 3

3.1 PROBLEM FORMULATION

  • The target rate is the proportion of errors whose latent representations contain sufficient signal for the ground-truth answer.A lightweight probe approximates the latent state, but its answer-based training labels make its output a proxy rather than proof of encoded evidence.Source: §3.1 · pp. 3–4

3.2 BOUNDS WITH AN IMPERFECT PROBE

  • Answer-preserving transformations that increase or decrease evidence salience support conservative bounds, provided the probe has nonzero detection after the positive transformation and nonzero non-detection after the negative one.The probe-dominance assumption treats those transformed detection and non-detection rates as conservative proxies for behavior when evidence is genuinely present or absent.Source: §3.2 · p. 4

3.3 VISUAL EVIDENCE IN OBJECT-ATTRIBUTE QUERIES

  • For object-attribute queries, the composite probe detects sufficient evidence only when both target-object evidence and correct-attribute evidence are detected.Attribute evidence means the object-specific representation supports the correct value more strongly than alternatives in the same category.Source: §3.3 · p. 4

4 THE VISUAL EVIDENCE LENS METHOD

VELens implements a composite object-and-attribute probe by projecting visual-token representations into vocabulary rankings across language-model layers. It converts relative target-versus-competitor ranks into layer-wise evidence vectors for lightweight classifiers.

4 THE VISUAL EVIDENCE LENS METHOD

  • VELens implements the composite object-and-attribute probe by projecting intermediate visual-token representations into vocabulary space and tracing evidence across layers.The paper presents it as its most reliable probe and estimate of the utilization-failure rate.Source: §4 · p. 5

4.1 METHOD OVERVIEW AND NOTATION

  • The output head maps each layer's visual-token representation to vocabulary rankings that indicate which concepts it supports.Because visual tokens precede question tokens and causal masking prevents them from attending to the question, their layer-wise representations encode image-available information.Source: §4.1 · p. 5

4.2 VISUAL-EVIDENCE EXTRACTION

  • A local evidence score is the fraction of competing tokens ranked below the best-ranked target token, with larger scores indicating stronger relative evidence.Local scores are aggregated over a spatial region for each layer, then stacked into an evidence vector.Source: §4.2 · p. 6

4.3 PROBE SUPERVISION

  • Logistic-regression probes classify layer-wise evidence vectors, using dataset annotations for attribute supervision and inside-versus-outside object regions for object supervision.The attribute probe labels candidate values present or absent; the object probe learns to distinguish regions containing the object from regions where it is absent.Source: §4.3 · p. 6

5 EXPERIMENTAL SETUP

The experiments compare VELens with classifier-, confidence-, and attention-based probes while keeping the VLM frozen. They use two open-weight VLMs and annotated object regions from VAW, focusing on color, material, and shape.

Baseline Methods

  • The baselines are classifier-, confidence-, and attention-based probes, all using the same mapping from annotated object regions to visual tokens.The classifier baseline needs a separate classifier for each attribute value and cannot handle unseen attributes or serve as an object probe.Source: §5 · p. 6

Pooling Configurations

  • Attribute probes use either one shared estimator across values (All) or one per value (Separate), while object evidence uses one estimator across categories.Aggregation strategies are selected on validation data, then retrained on the full training set and evaluated on held-out test data.Source: §5 · p. 7

Base Models

  • The evaluated VLMs are LLaVA-1.6-7B, connecting a pretrained CLIP vision encoder to a Mistral-7B LM, and Qwen2.5-VL-7B, pairing the Qwen2.5 LM with a natively trained dynamic-resolution vision encoder.Source: §5 · p. 7

Dataset

  • The evaluation uses 6,074 VAW images and focuses on color, material, and shape, which can be determined from the target object itself.VAW scenes contain over 17 objects per image on average.Source: §5 · p. 7

6 RESULTS

VELens performs strongly on visual-evidence detection and generalizes across limited supervision and unseen attribute values. Its estimates indicate that utilization failures dominate errors in both models, while evidence accessibility varies by layer and an attention intervention corrects some evidence-positive errors.

6.1 EVALUATING THE VISUAL-EVIDENCE PROBE

  • VELens outperforms baselines at every tested low-resource size, while Rank-All generalizes better than Separate to unseen values.Separate can use a few labeled examples for a new value; All can transfer when no labels are available.Source: §6.1 · pp. 7–8, Figure 3 · p. 8
  • 0.95 on LLaVA and 0.91 on Qwen: Rank-Separate plus the Attention object probe achieves the best combined visual-evidence ROC-AUC.Rank-Separate leads attribute detection across all three categories on both models; Attention leads object detection.Source: §6.1 · p. 7, Table 1 · p. 8

6.2 UTILIZATION FAILURES ARE THE MAIN BOTTLENECK

  • Increasing object noise lowers detected evidence and raises LLaVA’s error rate, while emphasizing the object reverses both trends.At 90% noise, evidence detection falls from 86.5% to 10.4% and error rate rises from 39.2% to 74.7%; the positive transformation yields only modest improvement, making the upper bound less informative.Source: §6.2 · p. 8, Figure 4 · p. 9
  • 84.4% for LLaVA and 70.0% for Qwen: estimated utilization-failure rates, with lower bounds of 77.5% and 61.0%, respectively.The pattern is consistent across attributes, and Figure 4 presents the estimated utilization and representation failures and rate bounds.Source: §6.2 · pp. 8–9, Figure 4 · p. 9

6.3 TRACING VISUAL EVIDENCE AND IMPROVING ITS UTILIZATION

  • Evidence generally strengthens in upper layers, but its accessibility trajectory differs across attributes and models, so no universal layer emerges.Qwen’s material evidence is strong early, while color and object evidence rise near the end; LLaVA’s object evidence increases more gradually.Source: §6.3 · p. 9
  • The attention intervention corrects 25.2% of evidence-positive errors versus 12% of evidence-negative errors, but requires a known object region.The authors describe it as a proof of concept rather than a standalone correction method.Source: §6.3 · pp. 9–10

7 DISCUSSION

The discussion frames VELens as a probe for tracing visual evidence and distinguishing whether errors reflect missing or unused evidence. The study reports results on a bounded set of models and attribute questions, while describing possible extensions and reproducibility provisions.

Discussion

  • VELens traces object and attribute visual evidence across VLM layers, and the example in Figure 5 shows an answer changing from wood to glass after increased attention to the target region.The example is an evidence-positive error in which VELens detects both object and attribute evidence.Source: §7 · p. 10, Figure 5 · p. 10

Scope and extensions

  • The analysis covers attribute questions about color, material, and shape, but the authors describe VELens as potentially extendable to other question types.They give attribute verification as an example, since it requires evidence for the object and queried value.Source: §7 · p. 10

AI use disclosure

  • The authors report using generative AI for code, dataset reformatting, text editing, and illustrative images, with all assisted work reviewed.They state that AI was not used to generate data, develop theory or proofs, propose hypotheses, design or critique experiments, or interpret results.Source: AI USE STATEMENT · p. 10

Study context

  • The study uses publicly released open-weight VLMs and the public VAW dataset, with no human subjects, personally identifying data, or new data release..Source: ETHICS STATEMENT · p. 10

Reproducibility

  • The authors state that code for VELens, baselines, and evaluation will be made public and describe the models, dataset processing, prompts, probe choices, transformations, and proof materials.The stated reproducibility materials include the complete proof of Theorem 1 with its assumptions.Source: REPRODUCIBILITY STATEMENT · pp. 10–11

A PROOF OF THE VISUAL-UTILIZATION FAILURE BOUNDS

The proof derives bounds on the utilization-failure rate by relating probe behavior on original errors to behavior after answer-preserving transformations. Its key assumption treats detection after strengthening evidence and non-detection after weakening evidence as conservative proxies for the latent evidence state.

Assumption 1

  • Probe dominance assumes that, conditional on an error, detection on the positive transformation and non-detection on the negative transformation conservatively proxy detection of the corresponding latent evidence state.The positive transformation strengthens relevant visual signal; the negative transformation weakens it without changing the ground-truth answer.Source: A PROOF OF THE VISUAL-UTILIZATION… · p. 14

Proof of Theorem 1

  • The proof conditions probabilities on the error event and defines the utilization-failure rate as the probability that required evidence is encoded.The complementary probability corresponds to a visual-representation failure.Source: A PROOF OF THE VISUAL-UTILIZATION… · pp. 14–15
  • The transformed-input detection and non-detection rates are used to establish the theorem’s bounds on the latent utilization-failure rate.The proof obtains the lower bound from probe non-detection on the negative transformation and uses detection on the positive transformation for the upper bound.Source: A PROOF OF THE VISUAL-UTILIZATION… · p. 15

B VELENS METHOD: COMPLEMENTARY DETAILS

VELens contrasts target concepts with competing tokens, then aggregates visual-token evidence into layer-level scores. Different aggregation operators summarize overall support, robustness, or extreme responses.

B.1 OBJECT AND ATTRIBUTE TOKEN SET CONSTRUCTION

  • Object evidence measures how strongly each visual token supports the target object against the entire vocabulary, while attribute evidence compares the target value with alternatives in its category.For attributes, the score is the fraction of alternative values ranked below the best-ranked target variant.Source: §B.1 · pp. 15–16

B.2 AGGREGATION STRATEGIES

  • Layer-level scores aggregate local evidence scores across a set of visual tokens at each language-model layer.The aggregation can be applied to inside- or outside-region token sets.Source: §B.2 · p. 16
  • The four aggregation operators capture different evidence distributions: mean reflects overall support, median robustness to outliers, and minimum or maximum extreme responses.The attention baseline instead aggregates over heads after averaging each concept’s query tokens.Source: §B.2 · p. 16

C EXPERIMENTAL SETUP: COMPLEMENTARY DETAILS

The attention baseline replaces vocabulary-rank evidence with cross-modal attention, while the logit lens projects visual-token representations into vocabulary space. These analyses illustrate how evidence is measured and how its depth varies across objects.

C.1 ATTENTION BASELINE

  • The attention baseline extracts frozen-model attention in one forward pass and compares attention allocated to the annotated object region with attention elsewhere.Concept-token attention ratios are normalized by region size to form inside and outside densities.Source: §C.1 · pp. 17–18
  • For attribute evidence, the baseline compares inside and outside attention densities; uniform image attention gives an inside proportion of 1/2, and larger values indicate preferential attention to the object region.Layer scores average over concept tokens and aggregate attention heads.Source: §C.1 · p. 18
  • For object evidence, the baseline keeps inside and outside attention densities as separate features, labeling inside-region vectors positive and outside-region vectors negative.The layer scores are stacked into separate inside and outside feature vectors.Source: §C.1 · p. 18

C.2 LOGIT LENS

  • The logit lens applies the vocabulary output head at each layer, and visual-token rankings depend only on the image because causal masking blocks access to the question.The projection connects visual-token representations to vocabulary concepts.Source: §C.2 · p. 18
  • In Figure 7, sampled plate-region tokens decode to plate, white, and round only from middle layers onward, while hat and bowl are top-ranked in the earliest layers.The figure shows each sampled token’s top-ranked vocabulary item across layers, with color intensity indicating top-1 probability.Source: Figure 7 · p. 19, §C.2 · p. 19

C.3 DATASET PREPROCESSING

The preprocessing filters ambiguous object-attribute labels, canonicalizes attribute values, and constructs reproducible questions with randomized option order. Models answer using frozen weights and greedy decoding.

C.3 DATASET PREPROCESSING

  • The dataset filters instances with multiple present values in a category or conflicting attributes across same-category object instances, then normalizes fine-grained values into 11 colors, 15 materials, and 5 shapes.Canonicalization reduces label sparsity and improves alignment with vocabulary-level representations.Source: §C.3 · p. 19, Table 3 · p. 21
  • Each eligible VAW instance yields one question per attribute category with a ground-truth value; its annotated region selects VELens tokens but is not shown to the model.Questions use the annotated object name and a shared template for both models.Source: §C.3 · p. 19, Figure 8 · p. 20
  • Canonical answer options are randomly permuted once and cached, so runs reuse identical prompts and option order.The canonical option sets are specified for each attribute category.Source: §C.3 · pp. 19–20, Table 2 · p. 20
  • The VLM remains frozen and answers are generated with greedy decoding; the option list constrains output to a known vocabulary for direct comparison with VAW labels.Predictions therefore require no free-text matching step.Source: §C.3 · p. 20

D.1 DEVELOPMENT-SET CONFIGURATION SELECTION

The appendix reports how probe configurations were selected and evaluates evidence scoring, controlled image transformations, and attention intervention. It also describes how the evidence formulation extends to other question classes when their answers decompose into grounded visual components.

D.1 DEVELOPMENT-SET CONFIGURATION SELECTION

  • Color and material yield comparably strong rank-based evidence in both models, whereas shape is consistently the most challenging attribute.The appendix attributes shape’s difficulty to needing the object’s outer boundary; configurations were selected on a training-derived development split before refitting on the complete training split and testing once.Source: §D.1 · pp. 21–22

D.2 JOINT OBJECT-AND-ATTRIBUTE AUC

  • The joint object-and-attribute probe ranks examples with a signed score: attribute support is positive, rejection is negative, and weak object evidence yields scores near zero.The ROC-AUC labels whether the attribute is annotated as present on the target object, pooling color, material, and shape (N = 3,128).Source: §D.2 · p. 25

D.3 ROC CURVES

  • On LLaVA, the rank-based probe dominates every baseline across ROC thresholds, while on Qwen the classifier-based baseline is comparable to it.The ROC curves show the attribute-evidence comparisons for color, material, and shape.Source: §D.3 · p. 25, Figure 9 · p. 25

D.4 CONTROLLED TRANSFORMATIONS

  • The positive transformation blurs surroundings while preserving the target object, and the negative transformation corrupts the object region at λ=0.3 while preserving the answer.The selected corruption remains answer-preserving; at λ=0.9 the object is largely destroyed and the correct answer is no longer recoverable.Source: §D.4 · pp. 25–27, Figure 10 · p. 26

D.5 LAYER-WISE VISUAL-EVIDENCE DEVELOPMENT

  • Figure 11 plots layer-wise AUC for object and attribute evidence in both models, tracing evidence that becomes accessible at different depths.Source: Figure 11 · p. 27

D.6 ATTENTION INTERVENTION

  • A 4.74% damage rate at α=2 keeps collateral errors below 5%, so the intervention uses that strength to increase attention to target-object tokens.Damage is measured on answers that were correct before intervention; α=1 yields 3.26% and α=3 yields 9.99%.Source: §D.6 · p. 27

E.1 A TAXONOMY OF VISUAL QUESTIONS

  • VELens applies directly when answering evidence can be decomposed into identifiable grounded components, rather than according to question wording alone.The appendix frames question types by whether evidence concerns one concept and region, multiple concepts, relations across regions, or global scene information.Source: §E.1 · p. 29, E VELENS BEYOND THE EVALUATED QUESTION… · p. 29

E.2 QUESTION CLASSES THAT NATURALLY FIT THIS FORMULATION

  • The formulation covers verification, object presence and localization, and compositional single-object questions by adapting candidate sets, image-region scope, or score aggregation to the required evidence.Verification with closed candidate values keeps the inputs unchanged; open-ended verification requires refitting the attribute estimator. Presence uses whole-image evidence, localization retains per-token scores, and multiple properties of one object share its region and are combined conjunctively.Source: §E.2 · p. 29

E.3 QUESTION CLASSES REQUIRING MORE SUBSTANTIAL EXTENSIONS

The controlled setting does not cover all visual question types: relational, counting, global-scene, and knowledge-dependent questions require evidence beyond fixed components tied to individual concepts. Extending the approach to these cases would require structured evidence for relations, instances, scene-level information, or visual and parametric knowledge together.

E.3 QUESTION CLASSES REQUIRING MORE SUBSTANTIAL EXTENSIONS

  • Relational and comparative questions require evidence for relations between entities, so detecting each entity separately does not establish the answer.Extending VELens would require components over tuples of regions.Source: §E.3 · p. 30
  • Counting questions require individuating and aggregating instances without double-counting evidence, while global-scene questions depend on distributed scene-level evidence.Using local scores for these tasks would require grouping instances and counting or summarizing those groups.Source: §E.3 · p. 30
  • Knowledge-dependent questions combine visual evidence with parametric knowledge, making visual-representation and visual-utilization failures less identifiable.An incorrect answer may reflect missing visual evidence, missing knowledge, or failure to combine the two.Source: §E.3 · p. 30
  • The experiments establish VELens in a deliberately controlled setting, while its natural scope is questions whose answer evidence composes structured, visually grounded concepts.The paper identifies relations, instance sets, and mixed visual and parametric evidence as paths toward broader diagnosis.Source: §E.3 · p. 30

Related papers

Loading 2610.29241v1…