SPELUNKA: A Framework for Finding Unanticipated Associations in Vision-Language Models

Cassandra Parent; Isha Agarwal; Kumail Alhamoud; Charikleia Moraitaki; Rafiya Javed; Walter Gerych; Marzyeh Ghassemi

cs.CVOct 13, 2026

SPELUNKA: A Framework for Finding Unanticipated Associations in Vision-Language Models

Authors: Cassandra Parent, Isha Agarwal, Kumail Alhamoud, Charikleia Moraitaki, Rafiya Javed, Walter Gerych, Marzyeh Ghassemi

Organizations: MIT · Google · WPI

Abstract

Vision-Language Models (VLMs) are the backbone of many state-of-the-art multimodal systems, but often encode spurious and harmful associations from their training data. While prior work focuses on removing known biases from VLMs, methods for discovering unanticipated associations for retrieval tasks remain underexplored. We introduce SPELUNKA, an automated, inference-time framework for finding unanticipated distributional biases for retrieval tasks in VLMs without pre-specified labeled subgroups, inference-time human supervision, or model fine-tuning. Given a concept of interest, SPELUNKA generates captions for related images, extracts relevant entities from captions, and selects semantically meaningful associations. A frequency-based detection rule then identifies spurious associations by measuring over- or under-representation of these attributes in a VLM's top-ranked outputs relative to their full dataset distribution. We evaluate SPELUNKA across four datasets, and find that it 1) successfully recovers both synthetic and naturally occurring spurious correlations, 2) achieves an average AUROC of 0.93 across four datasets in identifying images strongly associated with a query, and 3) outperforms baselines in finding features that are strongly associated with a given query.

Figures & tables

Panel of Figure 1 — Figure 1: Spelunka discovers and enables mitigation of unanticipated biases in VLM retrieval. Spelunka differentiates itself by being able to work on tasks that (a) do not have labels and (b) are semantically meaningful for downstream users compared to other methods.

Figure 2: Overview of the Spelunka pipeline. Step 1 generates pseudolabels by captioning each image, extracting entities, and combines common entities into a final set. Step 2 retrieves the top- $k$ images for a query using the VLM and identifies spurious correlations as the difference between the entity distribution in the top- $k$ set and the full dataset distribution.

Figure 3: Spelunka surfaces meaningful bias automatically by comparing entity frequencies between top- $k$ retrieved images and the full dataset, with no labels, task definition, or fine-tuning required. Querying AimV2 for “good” on FairFace overrepresents smiling faces relative to the dataset distribution (top). Projection debiasing, a common technique to remove a bias by orthogonalizing its embeddings to a concept ( Bolukbasi et al., 2016 ) , then aligns the retrieved distribution with the base rate (bottom).

Figure 4: Biases identified by Spelunka on Fairface when querying “person” across models. Images with entities Asian woman , Asian man , and smiling appear more frequently than in the full dataset than for AIMV2.

Entity	Model	FairFace	UTKFace	Agreement
Asian woman	aimv2-large-patch14-224-lit	0.187	0.168	TRUE
Asian man	aimv2-large-patch14-224-lit	0.169	0.119	TRUE
smiling	aimv2-large-patch14-224-lit	0.123	0.060	TRUE
smiling	CLIP-ViT-L-14-laion2B-s32B-b82K	-0.087	0.020	FALSE
smiling woman	siglip-base-patch16-224	0.077	0.150	TRUE
Young girl	aimv2-large-patch14-224-lit	-0.073	-0.051	TRUE

Table 1: Spelunka finds biases that hold across models. From the overlapping entities in FairFace and UTKFace , that met at least a 5% distribution shift and were significant in FairFace , 14/15 had sign agreement with UTKFace .

Panel of Figure 5 — Figure 5: Spelunka recovers ground-truth labels across datasets and captioning models. Pseudolabel frequencies track true label frequencies on (a) Fashion Product jewelry subtypes and on (b) EmoSet ‘happy’ images across CLIP and AIMv2 backbones.

Figure 6: Spelunka recovers the known background spurious correlation in Waterbirds . Without access to bird-type or background labels, Spelunka extracts background entities ( water , ocean , trees , Forest ) at high frequency from the test set.

	# Entities	Domino	RaVL (Retrieve)	Distill	RaVL (Classify)	Spelunka
Fashion, Jewelry	2	0.61 [0.54, 0.68]	0.55 [0.50, 0.58]	0.63 [0.56, 0.71]	0.59 [0.54, 0.63]	0.56 [0.48, 0.65]
	4	0.68 [0.61, 0.75]	0.67 [0.63, 0.71]	0.67 [0.58, 0.75]	0.66 [0.57, 0.66]	0.70 [0.63, 0.77]
	8	0.73 [0.65, 0.79]	0.67 [0.63, 0.70]	0.78 [0.72, 0.84]	0.72 [0.68, 0.75]	0.75 [0.68, 0.81]
	16	0.73 [0.66, 0.80]	0.69 [0.65, 0.72]	0.80 [0.74, 0.87]	0.74 [0.70, 0.77]	0.93 [0.90, 0.97]
EmoSet, Person	2	0.77 [0.72, 0.81]	0.50 [0.47, 0.53]	0.76 [0.71, 0.81]	0.60 [0.57, 0.63]	0.86 [0.81, 0.90]
	4	0.85 [0.82, 0.89]	0.73 [0.71, 0.75]	0.83 [0.79, 0.87]	0.73 [0.71, 0.76]	0.94 [0.91, 0.96]

Table 2: Spelunka outperforms baselines with an average AUROC of 0.93 in identifying images strongly associated with a query using a random forest classifier across

4

datasets and

6

queries with 16 entities (full results in Appendix). Intervals denote

95\%

confidence intervals with

500

bootstrapped samples. Comparisons done with clip-base-patch-16 . Distill and RaVL (Classify) require subtype labels and thus cannot be applied to all datasets.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Example image from ‘TV’ concept in Dollar Street .

Figure 8: Example image from ‘home’ concept in Dollar Street .

Figure 9: Example image from ‘topwear’ concept in Fashion Product .

Figure 10: Example image from ‘jewelry’ concept in Fashion Product .

Figure 11: Example image from FairFace .

Panel of Figure 12 — Figure 12: Pseudolabels and true label frequency for sub categories in the Jewelry concept. We find that Spelunka is able to accurately identify true labels.

Figure 13: Frequency shift results for the ‘face’ concept in FairFace .

Figure 14: Frequency shift results for the ‘jewelry’ concept in Fashion Data .

Figure 15: Frequency shift results for ‘home’ concept in Dollar Street .

Figure 16: Frequency shift results for ‘street view’ concept in Dollar Street .

Figure 17: Frequency shift results for ‘TV’ concept in Dollar Street .

Figure 18: Frequency shift results for the ‘scene’ concept in EmoSet .

Figure 19: Frequency shift results for the ‘person’ concept in EmoSet .

	# Entities	Domino	RaVL (Retrieve)	Distill	RaVL (Classify)	Spelunka
Fashion, Jewelry	2	0.61 [0.54, 0.68]	0.55 [0.50, 0.58]	0.63 [0.56, 0.71]	0.59 [0.54, 0.63]	0.56 [0.48, 0.65]
	4	0.68 [0.61, 0.75]	0.67 [0.63, 0.71]	0.67 [0.58, 0.75]	0.66 [0.57, 0.66]	0.70 [0.63, 0.77]
	8	0.73 [0.65, 0.79]	0.67 [0.63, 0.70]	0.78 [0.72, 0.84]	0.72 [0.68, 0.75]	0.75 [0.68, 0.81]
	16	0.73 [0.66, 0.80]	0.69 [0.65, 0.72]	0.80 [0.74, 0.87]	0.74 [0.70, 0.77]	0.93 [0.90, 0.97]
EmoSet, Person	2	0.77 [0.72, 0.81]	0.50 [0.47, 0.53]	0.76 [0.71, 0.81]	0.60 [0.57, 0.63]	0.86 [0.81, 0.90]
	4	0.85 [0.82, 0.89]	0.73 [0.71, 0.75]	0.83 [0.79, 0.87]	0.73 [0.71, 0.76]	0.94 [0.91, 0.96]

Appendix

Table 3: Comparison of methods with clip-base-patch-16 . Distill and RaVL (Classify) require sub labels and thus cannot be applied to all datasets.

Panel of Figure 20 — Figure 20: Spelunka identifies grayscale bias and enables its correction. On Dollar Street TVs with 50% of images randomly converted to grayscale, querying clip-vit-base-patch-16 for “expensive” underrepresents grayscale images in the top-20 retrieved (a); projection debiasing on the black-and-white concept restores the base-rate distribution (b).

VLM Evaluation

Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.

See all 999 papers

Scaling Vision-Language Models Is Not Enough to Mitigate Bias
Jul 30, 2026Ioannis Sarridis, Ioannis Kompatsiaris, Symeon PapadopoulosVLM Evaluation VLM Robustness
Information Technologies Institute, CERTH Thessaloniki, Greece
Partition-Aware Unlearning for Removing Spurious Correlations in Large Vision-Language Models
Aug 30, 2026Aditi Sarker, Nazreen Shah, Rafi Ibn Sultan +3VLM Evaluation VLM Unlearning
Department of Computer Science, Wayne State University, MI, USA · Department of Electronics & Communications Engineering, IIIT Delhi, India · Institute for AI and Data Science (AIDaS), Wayne State University, MI, USA
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
May 28, 2026Cheolhong Min, Jaeyun Jung, Daeun Lee +5Representation Learning Visual Spatial Reasoning
Seoul National University · The Ohio State University · NVIDIA

Jul 30, 2026J/K move · Enter open · S save

Scaling Vision-Language Models Is Not Enough to Mitigate Bias

Ioannis Sarridis, Ioannis Kompatsiaris, Symeon Papadopoulos

Information Technologies Institute, CERTH Thessaloniki, Greece

VLM Evaluation VLM Robustness Vision-Language Models

Open paper AskPDF arXiv

Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases). Across these settings, the Spearman correlation between model scale and performance weakens as evaluation shifts from ImageNet (

ρ{=}0.68

) to single-attribute (

ρ{=}0.48

) and further to multi-attribute (

ρ{=}0.05

) bias benchmarks. In contrast, properties of the training data (size and quality) show more consistent relationships with worst-group accuracy across both bias benchmarks. Notably, curated datasets yield improvements of up to 25% over uncurated alternatives at a comparable scale. Finally, the effect of architectural choices (e.g., patch size, image resolution) is highly context-dependent, varying with the nature of the benchmark, including the type of bias and its spatial distribution within images.

Abstract

VLM Evaluation

Explore similar work

Scaling Vision-Language Models Is Not Enough to Mitigate Bias

Partition-Aware Unlearning for Removing Spurious Correlations in Large Vision-Language Models

Why Far Looks Up: Probing Spatial Representation in Vision-Language Models