VLMProbe: An Open-source Library for VLM Probing and Interpretation

Gexin Huang; Haisheng Fu; Yanting Yang; Chen Zhao; Ruinan Jin; Jiaqi Liu; Xiaoxiao Li

cs.CVOct 21, 2026

VLMProbe: An Open-source Library for VLM Probing and Interpretation

Authors: Gexin Huang, Haisheng Fu, Yanting Yang, Chen Zhao, Ruinan Jin, Jiaqi Liu, Xiaoxiao Li

Organizations: The University of British Columbia · Vector Institute

Abstract

Understanding how vision-language models use visual information requires connecting internal representations to model behavior. We present vlmprobe, an open-source library that integrates lens readouts, attention attribution, causal interventions, and sparse autoencoder (SAE) analysis within a common model interface. Shared adapters, calibration and checkpoint utilities, evaluation pipelines, and visualizations support systematic analysis across architectures. Using this framework, we compare methods across vision-language, language, and vision models through matched case studies and controlled evaluations of pretrained sparse dictionaries. The comparisons show that representation readouts and spatial explanations depend on the analysis objective and observation site, and that local attribution estimates can disagree with measured intervention effects. SAE evaluations further expose distinct trade-offs in reconstruction fidelity, concept readout, and spatial localization. Together, these results support combining multiple forms of evidence to evaluate an interpretation. vlmprobe provides a reusable basis for selecting methods, comparing their conclusions, and testing hypotheses about model behavior. The code is available on github.com/ubc-tea/vlm-probe

Figures & tables

Figure 1: The vlmprobe workflow. Native inputs and explicit analysis settings feed lens, attention, causal and sparse-feature methods; their outputs support token readouts, spatial explanations, sparse feature analysis and measured interventions. Miniature plots illustrate output types rather than experimental results.

Model	Image grid	Architectural feature
SmolVLM-500M-Instruct	$8\times 8$	Compressed image tokens with a Llama decoder
InternVL3-1B-hf	$16\times 16$	InternVL vision path with a Qwen2 decoder
Qwen3-VL-2B-Instruct	$14\times 14$	Normalized Q/K and DeepStack visual additions
Pixtral-12B	$16\times 16$	Pixtral patch encoder with a Mistral decoder
LLaVA-1.5-7B	$24\times 24$	CLIP visual features projected into a Llama decoder

Table 1: Models in the shared-case comparison. Sizes follow public checkpoint names. Native adapters retain each model’s decoder sites and visual-token grid.

Source	Released identifiers	Analysis question
GQA	07302654 / 19225230	Is the shirt black or yellow?
MMVP	Images 1 / 2	Are the butterfly’s wings open or closed?
VSR zero-shot	Test rows 0 / 1	Car behind suitcase; cat under backpack
VisOnlyQA real	Row 0 of two splits	Connected triangle; chart-value extraction

Table 2: Predetermined shared inputs. Each source contributes two cases; image captions from a separate COCO split calibrate the VLM Tuned Lenses.

Family	Methods demonstrated	Matched comparison
Lens	Logit, Tuned, Jacobian, Attention Lens; EmbedLens readout; Patchscope	Residual/head decoding on IOI; visual-token and transfer readouts on GQA
Attention	Rollout, Grad-CAM, ATTATTR, TAM, Beyond-H/C, generic relevance	Seven attribution methods on GQA across five VLMs; native ViT maps and deletion tests
Causal	Activation/attribution patching, knockout, path patching, EAP/EAP-IG, VSV steering	Exact versus approximate replacements; attention knockout; controlled sender paths; edge retention; generation changes across three VLMs
SAE	ReLU, TopK, Gated, JumpReLU, BatchTopK, Matryoshka; RA-SAE and SAE++ readouts	Frozen Pythia dictionaries; source-native spatial readouts across nine vision/VLM backbones; image–question attribution and feature removal in five decoder configurations

Table 3: Comparison groups and the evidence each provides. Groups organize matched inputs and compatible measures; distinct analysis units remain separately interpreted.

Figure 2: Layer-wise vocabulary readouts for one GQA decision across five VLMs. The input panel shows the question, answer and annotated shirt. All decoder layers retain their measured colors; selected rows show shortened top-1 fragments after native tokenizer decoding. Top-1 search includes all vocabulary languages, regardless of the English input. Undecodable or unrenderable fragments remain unlabeled. Yellow indicates a better rank for the fixed next token, using one shared logarithmic scale. Orange outlines identify answer predictions: Qwen3-VL generates Black while the other models generate Yellow. LLaVA predicts Yellow in two pieces, with the second conditioned on the first. Horizontal labels shorten chat markers to [start]/[end], omit whitespace markers and display line breaks as arrows; columns follow each model’s native template.

Figure 3: Residual and head readouts of the same IOI decision in GPT-2 small. Left: Logit, Tuned, and Jacobian Lens track Danielle at the final prediction position. Right: Attention Lens distinguishes heads within layers 7 and 10. Top-token labels and target ranks reveal agreement at late residual sites alongside substantial head-level variation.

Figure 4: EmbedLens attaches lexical readouts to visual patches in LLaVA-1.5-7B. Amber patch boxes and leader lines connect the nearest vocabulary fragments to their projected visual tokens; green boxes mark the annotated shirt. These are embedding-space neighbors rather than independently detected object labels. The black-shirt example associates with “black,” while yellow-shirt patches also yield unrelated fragments such as “Blue.”

Figure 5: Patchscope reads a source residual through a new text context. In LLaVA-1.5-7B, the source answer-position residual replaces the placeholder after target layer 0 in a four-line color-copy prompt, red -> red , blue -> blue , green -> green , then x -> . Curves show changes in complete-word log probability relative to a same-site zero-vector injection; positive values indicate greater support for the plotted color. This configuration changes word likelihoods but fails to recover the shirt color as its top continuation.

Figure 6: Attention explanations assign different spatial emphasis to the same yellow-shirt input. Rows compare five VLMs; columns compare raw attention and seven explanation methods. Target-conditioned methods explain Yellow. Blue indicates positive values and orange negative ATTATTR values; green boxes mark the shirt. Each map is independently normalized, so intensity compares positions within a map rather than score magnitude across maps. Original photographs are displayed at 20% opacity over opaque heatmaps. Raw attention, Grad-CAM, and ATTATTR inspect the last decoder attention layer; propagation methods span the decoder attention stack. “Zero map” denotes a measured null attribution.

Figure 7: Object-conditioned attribution and deletion on ViT-B/16. Seven maps explain cat and dog in the same author-provided image. Curves measure class probability after replacing top-ranked patches with training-mean RGB. Rollout combines target-independent connections, while other tools can separate the classes; deletion exposes differences that map appearance alone does not resolve.

Figure 8: Attribution patching can miss effects measured by activation patching. Gray-image hidden states replace selected states in the real-image run at one decoder layer. The schematic rows are separate interventions: all image-token states or the states that predict the answer’s token pieces. Teal markers measure changes in the complete-answer log probability; red curves estimate the same changes from gradients. Negative values reduce support for Yellow. The estimates miss large early image-state and late answer-state effects.

Figure 9: Single-position replacement separates question-word effects from prediction-site sensitivity. Each cell replaces one native text-position state with its corresponding gray-image state at the indicated layer, as depicted in the schematic. Columns cover question and assistant-boundary subwords. Red reduces the complete-answer log probability of Yellow; blue increases it. Activation patching measures the replacement effect and attribution patching estimates it, using the same signed scale. Contrast with the joint replacements in Figure 8 .

Figure 10: Attention Knockout reveals route-specific dependence in LLaVA-1.5-7B. Blocking image-to-question connections at layers 8–16 lowers the fixed Yellow first-subword probability from 0.753 to 0.119 and makes Black the next-token prediction. Blocking shirt keys lowers it to 0.076; region sizes differ, so these curves describe the selected masks rather than controlled localization accuracy.

Figure 11: Controlled sender-head paths on a paired butterfly-wing input. Each cell replaces one sender using its reference-image value at the prediction position, with other head outputs frozen and only the recomputed final residual replayed. Other head outputs are frozen during the controlled run; MLP and normalization recompute, and only the resulting receiver is injected in a fresh run. Signed effects compare the same configured receiver across five VLMs. Yellow outlines identify the largest absolute effect within each model; blue decreases and red increases support for Open.

Figure 12: Attribution rankings differ in actual IOI circuit recovery. EAP and input-path EAP-IG select connections in the same GPT-2 graph; each curve point is a fresh retained-edge intervention. Both complete graphs recover the clean margin, but EAP-IG does not uniformly outperform EAP on this single prompt pair.

Figure 15: Pretrained language SAE trade-offs under a common backbone, layer and dictionary width. Each point is one of 18 fixed checkpoints. Left: Pile reconstruction FVU with an independent calibration mean. Middle: next-token NLL increase after native residual reconstruction. Right: held-out AG News macro AUROC using one selected feature per one-versus-rest class; horizontal lines show matched raw-neuron, PCA and random-projection controls. All panels use measured Pile

\overline{L}_{0}

on the horizontal axis. All SAE weights remain frozen during evaluation.

Figure 16: ReLU and TopK SAE features on the same photograph for CLIP-B/32 and DINO-B/16. Rows fix the backbone and observation layer; columns compare dictionaries. Blue heatmaps show feature activation; orange boxes mark people.

		Activity / fidelity		Spatial response
Model	Recipe	$\overline{L}_{0}$	FVU $\downarrow$	$N$	Coverage	Point@1	Shuffled	Mass
CLIP B/32	ReLU	1,413.14	0.0152	1,534	100.0	54.5	19.2	39.4
CLIP B/32	TopK	64.00	0.2163	1,534	66.5	44.7	16.2	58.8
CLIP B/32	Gated	107.80	0.1261	1,534	69.4	47.1	16.3	57.8
DINO B/16	ReLU	1,083.63	0.0063	1,534	97.9	52.7	18.6	38.6
DINO B/16	TopK	1,023.82	0.0085	1,534	65.4	35.5	15.6	44.4

Table 4: Pretrained spatial SAE configurations on COCO. Activity and reconstruction are evaluated alongside category-feature response and localization; shuffled maps provide a spatial control.

Figure 17: Visual-encoder SAE features across six backbones on four photographs. Each column uses one pretrained dictionary at its native observation layer. Blue heatmaps show feature activation; orange boxes mark target objects. All SAE observations precede language decoding.

Figure 18: Question-conditioned decoder SAE attribution across five model configurations. Inputs and reference answers appear at left; heatmaps show positive net contributions from selected SAE feature groups. Orange boxes mark target objects; maps are normalized independently.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Checkpoint	Revision
HuggingFaceTB/SmolVLM-500M-Instruct	a7da5b986cb59b408707209984f360a5f4ad7e47
OpenGVLab/InternVL3-1B-hf	014c0583a0d4bedf29fbe2dbff4f865eb998e171
Qwen/Qwen3-VL-2B-Instruct	89644892e4d85e24eaac8bacfd4f463576704203
mistral-community/pixtral-12b	c2756cbbb9422eba9f6c5c439a214b0392dfc998
llava-hf/llava-1.5-7b-hf	b234b804b114d9e37bb655e11cbbb5f5e971b7a9
Qwen/Qwen3.5-4B (appendix)	851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a

Appendix

Table 5: Pinned pretrained checkpoints for the shared VLM experiments.

Model	Layer	Initial KL	Selected KL	Epoch
SmolVLM	7	14.3279	1.9050	11
SmolVLM	15	11.4524	1.4967	12
SmolVLM	23	6.0436	0.7889	11
LLaVA	7	7.8680	1.3609	12
LLaVA	15	5.9005	0.9137	12
LLaVA	23	4.5529	0.3782	11

Appendix

Table 6: Tuned Lens validation KL in nats per selected position. Initial and selected translators are evaluated on the same held-out caption positions.

Backbone	Recipe	$d$	$m$	Native observation / input normalization
CLIP B/32	ReLU	768	49,152	Block 11 residual; raw
CLIP B/32	TopK	768	49,152	Block 11 residual; per-token sample normalization
CLIP B/32	Gated	768	49,152	Block 9 MLP output; raw
DINO B/16	ReLU	768	49,152	Block 11 residual; per-token sample normalization
DINO B/16	TopK	768	49,152	Block 11 residual; per-token sample normalization
CLIP L/14-336	Matryoshka	1,024	65,536	Block 22 residual; raw

Appendix

Table 7: Published spatial checkpoints and native observation scopes.

d

is activation width and

m

dictionary width. All layer numbers are zero-based source block indices. “Raw” means no extra SAE input normalization; DINOv2’s backbone final normalization is part of its native observation. SAE++ L1 is a level of a jointly trained hierarchy; RA-SAE retains its nonlinear encoder and relaxed archetypal dictionary.

Method	Source setting	$\overline{L}_{0}$	FVU	$\Delta\mathrm{NLL}$	AUROC@1	AUROC@5
ReLU + L1	$\lambda=0.012$	468.80	0.0399	0.1145	0.9120	0.9329
ReLU + L1	$\lambda=0.020$	227.06	0.0690	0.2493	0.9073	0.9362
ReLU + L1	$\lambda=0.040$	77.71	0.1078	0.5796	0.8869	0.9270
TopK	$k=20$	20.00	0.1081	0.5751	0.8412	0.9247
TopK	$k=80$	80.00	0.0659	0.2080	0.9002	0.9318
TopK	$k=320$	320.00	0.0182	0.0341	0.9239	0.9370

Appendix

Table 8: All 18 pretrained language dictionaries. Rows within each family use source trainer indices 0, 2 and 4.

\lambda

is the published sparsity penalty; JumpReLU settings are training targets, while reported

\overline{L}_{0}

is measured on Pile. FVU uses the independent calibration mean, and

\Delta\mathrm{NLL}

is in nats per prediction. AUROC uses one or five selected features separately for each AG News one-versus-rest class.

Representation	AUROC@1	AUROC@2	AUROC@5	Macro F1@5
Raw neurons (768)	0.8506	0.8929	0.9249	0.8038
PCA (128)	0.8511	0.9105	0.9481	0.8363
Random projection (4,096)	0.8773	0.9105	0.9286	0.7952

Appendix

Table 9: Matched concept-probe controls on the same AG News documents. The three AUROC columns use one, two or five selected features per class; F1 uses five. PCA and standardization are fitted on training documents only.

Backbone / scope	Native grid	Calibration positions	Validation positions
CLIP B/32	$7\times 7$	24,500	49,000
DINO B/16	$14\times 14$	98,000	196,000
CLIP L/14-336	$24\times 24$	288,000	576,000
CLIP B/16	$14\times 14$	98,000	196,000
SigLIP B/16	$14\times 14$	98,000	196,000
DINOv2 B/14	$16\times 16$	128,000	256,000

Appendix

Table 10: Native spatial observation counts on the fixed 500 calibration and 1,000 validation images. CLS is excluded. The Qwen rows retain their actual image-dependent grids and observe different stages.

Category	Feature	$N$	Coverage	Point@1	Mass (active)
Backpack	370	40	92.5	62.5	42.4
Bench	3276	45	68.9	60.0	77.8
Bottle	2167	79	64.6	41.8	56.8
Bowl	574	56	89.3	71.4	67.4
Car	2499	97	100.0	75.3	52.5
Cell phone	1195	36	72.2	50.0	59.2

Appendix

Table 11: Complete held-out localization results for the CLIP ViT-L/14-336 Matryoshka patch dictionary and its train-selected categories.

N

counts image–category pairs. Coverage, point@1 and active-only mass inside category-box unions are percentages; inactive maps count as point@1 misses. Feature IDs preserve the author’s dictionary order.

Dictionary	Scope	Width	$\overline{L}_{0}$	FVU	Cosine
BatchTopK	CLS	16,384	24.35	0.2816	0.9709
Matryoshka	CLS	16,384	23.76	0.2956	0.9696

Appendix

Table 12: CLS dictionaries on 1,000 held-out COCO images. FVU uses a reference mean from 500 calibration images. Cosine is measured at the final projected image embedding after native CLS reconstruction.

VLM	Boundary	Hidden	Features	Source activation
Qwen2.5-VL-3B	L34 output	2,048	65,536	Raw TopK, $k=100$
Qwen3.5-4B	L27 output	2,560	65,536	JumpReLU export
LLaVA-NeXT Llama3-8B	L24 output	4,096	131,072	ReLU + TopK, $k=256$
LLaVA-v1.6-Mistral-7B	L24 input	4,096	32,768	ReLU, scale 8
PaliGemma2-3B mix448	L23 output	2,304	16,384	JumpReLU

Appendix

Table 13: Native pretrained decoder observations. All widths and sparsity budgets come from the published weights/configurations, not new training.

Image	Question	Canonical / foil	Instance	Source
14439	What color is the jacket worn by the foreground child to the right of the large kite?	black / white	1733380	C
14439	What color are the trousers worn by the foreground child to the right of the large kite?	blue / red	1733380	C
263463	What part of the car surrounds the dog’s reflection?	side mirror / windshield	145584	C
263463	What color is the frame of the car’s side mirror in the foreground?	black / white	145584	C
263463	What animal is reflected in the car’s side mirror?	dog / cat	10416	C
532575	What shape is the dog tag?	heart / circle	4318	V

Appendix

Table 14: Fixed decoder questions. C denotes an investigator-constructed question; V denotes original VQAv2. The selected COCO instance ID binds all feature and pixel controls for that question.

Model	Question	FVU	Relative error	Mean $L_{0}$
Qwen2.5-VL-3B	Child jacket	0.225	0.253	100.0
Qwen2.5-VL-3B	Mirror part	0.228	0.256	100.0
Qwen2.5-VL-3B	Tag shape †	0.221	0.254	100.0
Qwen3.5-4B	Child jacket	0.455	0.549	124.7
Qwen3.5-4B	Mirror part	0.489	0.563	118.0
Qwen3.5-4B	Tag shape †	0.407	0.534	124.9

Appendix

Table 15: Dictionary fidelity at the image positions used in the illustrated decoder analyses. Values use the same formal image, question and answer-format instruction as the interventions. FVU is position-centered within the canonical pass; relative error is

\|D(z)-h\|_{2}/\|h\|_{2}

, and mean

L_{0}

counts active coordinates per image position.

Model	Question	$S$	$\Delta_{8}$	$\overline{\Delta}_{\rm rnd}$
Qwen2.5-VL-3B	Child jacket	4.972	0.008	-0.0001
Qwen2.5-VL-3B	Child trousers	4.314	0.787	0.0006
Qwen2.5-VL-3B	Mirror part	1.982	0.027	-0.0002
Qwen2.5-VL-3B	Mirror frame	5.274	0.713	-0.0001
Qwen2.5-VL-3B	Reflected animal	5.725	0.229	0.0005
Qwen2.5-VL-3B	Tag shape †	2.452	0.572	0.0004

Appendix

Table 16: Decoder feature deletions across eight fixed questions and five model configurations.

\dagger

marks original VQAv2 questions; other questions are constructed.

Question	Qwen2.5-VL	Qwen3.5	LLaVA Llama3	LLaVA Mistral	PaliGemma
Child jacket	black	black	Brown	Brown	brown
Child trousers	black	blue	Blue	Blue	blue
Mirror part	mirror	mirror	Mirror	Mirror	mirror
Mirror frame	black	black	Black	Black	black
Reflected animal	dog	dog	Dog	Dog	dog
Tag shape †	heart	heart	Heart	Heart	heart

Appendix

Table 17: Greedy generated content from the original image and question. Display whitespace is trimmed; candidate scoring is separate.

Model	Question	FVU	$\Delta_{\rm rec}$	$\Delta_{\rm box}$	$\overline{\Delta}_{\rm px}$	Max overlap
Qwen2.5-VL-3B	Child jacket	0.225	0.001	2.766	-0.200	0.000
Qwen2.5-VL-3B	Child trousers	0.225	0.006	0.789	0.468	0.000
Qwen2.5-VL-3B	Mirror part	0.228	0.010	2.169	2.234	0.926
Qwen2.5-VL-3B	Mirror frame	0.228	-0.011	1.808	1.386	0.854
Qwen2.5-VL-3B	Reflected animal	0.228	-0.021	4.627	0.229	0.000
Qwen2.5-VL-3B	Tag shape †	0.221	-0.002	3.238	0.314	0.690

Appendix

Table 18: Reconstruction and pixel controls. FVU is centered over canonical-pass image positions within each example. Score drops compare full reconstruction, target-box masking and the mean of three same-area translations. Max overlap is the largest translated-mask fraction overlapping the target.

Model	Question	$n_{8}$	$\bar{n}_{\rm rnd}$	$\\|\delta h_{8}\\|_{2}$	$\overline{\\|\delta h_{\rm rnd}\\|_{2}}$
Qwen2.5-VL-3B	Child jacket	207/207	14.3/14.3	652.70/652.70	46.65/46.65
Qwen2.5-VL-3B	Child trousers	295/295	16.7/16.7	1767.97/1767.97	54.76/54.76
Qwen2.5-VL-3B	Mirror part	296/296	16.0/16.0	901.80/901.80	41.48/41.48
Qwen2.5-VL-3B	Mirror frame	204/204	19.3/19.3	1825.23/1825.23	32.23/32.23
Qwen2.5-VL-3B	Reflected animal	119/119	16.7/16.7	1022.65/1022.65	32.66/32.66
Qwen2.5-VL-3B	Tag shape †	282/282	14.0/14.0	2148.47/2148.47	22.29/22.29

Appendix

Table 19: Actual deletion budgets. Each entry reports canonical/foil passes.

n

counts edited nonzero coordinates across spatial positions; random columns average three groups. ID-count matching does not match coordinate count or perturbation energy.

Figure 19: A second native ViT control separates elephant and zebra attribution. Identical image preprocessing and deletion budgets allow map differences to be checked against the effect of removing selected patches.

Figure 20: Exact and first-order residual replacement for the black-shirt question. LLaVA’s measured effects are markedly smaller than on Yellow, whereas SmolVLM remains sensitive to the gray-image reference. Both methods use the same native sites and complete Black score.

Figure 21: Text-position patching for the black-shirt input. Exact replacements identify intermediate question-word and later prediction-position effects; the aligned first-order maps provide a direct approximation check rather than a separate attribution target.

VLM Interpretability

Vision-language model interpretability explains how image and language representations shape predictions, evidence use, and multimodal reasoning.

See all 268 papers

May 27, 2026

The Abstraction Gap in Vision-Language Causal Reasoning

Vision-language models (VLMs) generate fluent causal explanations, but current evaluations cannot distinguish linguistic plausibility from faithful causal reasoning. We introduce a dual-probe methodology that isolates these properties. The Text-Only Probe measures linguistic quality. The Chain-Text Probe requires models to first generate explicit causal chains. The Abstraction Gap (AG) metric quantifies the normalized performance difference. Evaluating eight VLMs on CAGE (Causal Abstraction Gap Evaluation), a benchmark of 49,500 questions across 5,500 images spanning Pearl's causal hierarchy, we find seven models exhibit AG exceeding 0.50 with text scores of 6--8 but chain scores below 2.5. Fine-tuning on 45,000 chain-annotated examples fails to close the gap. However, one model achieves near-zero AG. The capability exists within current VLM architectures and depends on pretraining and architectural choices. CAGE provides a diagnostic tool for assessing faithful causal reasoning in VLMs.

Chinh Hoang, Mohammad Rashedul Hasan

Department of Electrical and Computer Engineering, University of Nebraska–Lincoln, Lincoln, Nebraska, USA.

VLM Evaluation Causal Reasoning in Language Models VLM Reasoning Faithfulness in VLMs Reasoning for VLMs

Ask PDF

Aug 11, 2026

Where To Look? : Causal Tracing of Vision Encoders in VLM

Vision-language models can describe an image with remarkable accuracy, yet a more fundamental question remains unanswered: what visual information actually drives their answers? In this work, we investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region. Extending the analysis to larger vision-language models reveals a similar pattern across models and corruption settings, suggesting that strong multimodal performance does not necessarily imply spatially localized causal representations. We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures. Together, our experiments expose a gap between seeing, using, and reasoning over visual structure, and provide a causal framework for studying how visual information is transformed, preserved, and ultimately used by modern vision-language models.

Naren Kumar S, Tirth Bhatt, Mayank Singh

LINGO Research Group, Indian Institute of Technology Gandhinagar, India

Large Vision-Language Models VLM Interpretability VLM Reasoning

Ask PDF

Oct 21, 2026

Encoded but Unused: Bounding the Sources of VLM Errors

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.

Anat Veikherman, Mor Ventura, Nitay Calderon +2

Faculty of Electrical and Computer Engineering, Technion · Faculty of Data and Decision Science, Technion

VLM Interpretability Visual Evidence Sensitivity in VLMs

Ask PDF

Abstract

VLM Interpretability

Explore similar work

The Abstraction Gap in Vision-Language Causal Reasoning

Where To Look? : Causal Tracing of Vision Encoders in VLM

Encoded but Unused: Bounding the Sources of VLM Errors