- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
SPELUNKA discovers attributes that are over- or underrepresented in VLM retrieval results. It generates image captions, extracts and groups entities, and compares their frequencies in top-ranked images against the full dataset. The method requires no predefined subgroup labels or model finetuning. Experiments include attribute recovery, cross-dataset transfer, retrieval prediction, and debiasing examples. The paper reports an average AUROC of 0.93 for predicting top-5% retrieval membership.
Critical strengths
-
The method addresses a practical auditing problem. Discovering retrieval associations without predefined subgroup labels is useful. The pipeline produces readable descriptions rather than unnamed clusters. Figure 3 also shows how a discovered association can guide an existing debiasing method.
-
The main text provides several complementary forms of evidence. These include quantitative comparisons, label recovery across captioners, cross-dataset consistency, and visual case studies. Table 2 shows large gains at 16 entities. For example, SPELUNKA achieves 0.93 AUROC on jewelry, compared with 0.80 for the strongest reported baseline. These results support its ability to summarize retrieval behavior.
Critical weaknesses
- The evaluation does not sufficiently distinguish spurious associations from legitimate retrieval features. Section 3.2 treats frequency shifts as spurious associations. However, a feature can be overrepresented because it genuinely helps identify the query. The AUROC evaluation in Section 5.3 rewards any feature that predicts retrieval membership. It therefore does not establish accurate spurious-feature discovery. Similarly, Figure 6 shows that background entities occur in Waterbirds captions, but does not demonstrate that the audited model relies on them. Section 6 acknowledges part of this limitation, yet the main contribution claims remain stronger than the evidence.
- The caption-bias guarantee requires a stronger assumption than “better than chance.” The derivation in Section 3.3 uses the same entity-specific true-positive and false-positive rates for the top-ranked subset and the full dataset. This requires caption errors to remain stable under retrieval selection. If a captioner detects an entity more easily in high-ranked images, the observed frequency shift can be distorted or even reversed. Figure 5 provides useful examples, but does not test this assumption. The guarantee is valid under the stated noise model, not for arbitrary biased captioners.
Questions for authors
-
Can you evaluate discovery precision and recall using known spurious and genuinely relevant attributes? For Waterbirds, can you show that the detected backgrounds affect retrieval scores while bird identity is controlled?
-
Do captioner true-positive and false-positive rates differ between top-ranked and remaining images? Please state the required noise assumptions explicitly and clarify how they support Appendix E’s guarantees.
-
Were entity selection and frequency-based ranking performed using only the training half before the random-forest evaluation? Were baseline features selected under the same restriction?
-
Bias-to-Text already discovers interpretable bias keywords from image captions, although it uses classification errors. What does SPELUNKA add beyond adapting this approach to retrieval groups? Can you compare against a retrieval adaptation using the same captions and feature budget?
-
How does the bootstrap handle the overlap between the top-ranked subset and the full dataset? Is retrieval repeated within each bootstrap sample? Do the discoveries remain reliable after correcting for testing many entities?
2 comments
Rebuttal Part 2.
Q1. We find that SPELUNKA is able to properly extract water-related entities and land-related entities from images with those backgrounds. We randomly sample 2000 images from the test set of which 1002 have a water background and 998 have a land background. We search for a match of words that are water-related or land-related (either in singular or plural form) in SPELUNKA-generated entities. We find that 17 water-related words cover 92.3% of all images labeled water and 22 land-related terms cover 92.4% of all images labeled land. Restricting to the 5 water-related words (i.e. water, ocean, lake, river, pond) still covers 84.3% of all images labeled water and 5 land-related terms (i.e. forest, tree, grass, field, ground) cover 83.5% of all land labeled images. Note that these 5 terms were chosen to relate directly to water or land, but they are not necessarily the most frequently occurring entities from SPELUNKA. This shows that SPELUNKA can correctly recover background labels and could detect background-related biases in the Waterbirds dataset.
Q2. See response to W2.
Q3. Entity selection and frequency-based ranking was performed only using the training half for the random-forest evaluation. Baseline features had the same restriction. We will make this more clear in the main text.
Q4. We believe our paper is very different from Bias-to-Text. Beyond their objective of discovering error slices, they use an existing algorithm for keyword selection which creates opacity in bias selection. Further, they do not show statistical significance or prevalence of their found biases and can not show that these biases are of the model (i.e. no validation to other models). Because their biases focus on error slice discovery, they generate coarse descriptions.
We find that bias-to-text cannot be fairly adapted into a retrieval setting in a straight-forward way. Beyond keyword extraction from misclassified images, bias-to-text also uses similarity scores between those keywords and the misclassified images to generate the top word candidates with downstream goals being related to fairness in the classification task. Thus, bias-to-text is too enmeshed in classification to generate a fair comparison that is true to the original method.
Q5. For each bootstrap, the dataset is resampled with replacement. Then, retrieval occurs over the set of entities to find the magnitude of each potential spurious correlation. These magnitudes create 95% CIs.
Rebuttal Part 1.
We thank the reviewer for their response and provide the following point-by-point rebuttal.
W1. For the core experiments, including the mentioned section 3.2 and 5.3, we use a query that is representative of the class of images. For example, in Figure 4, we are using the query “person” on a dataset of people. For the core experiments, we are using a query that is representative of the class of images, i.e. in Figure 4, we are using the query “person” on a dataset of people. Because every image in the dataset could satisfy the query, the images that are retrieved display model-specific associations to the query. Normal semantic entities that are common across all images for that class (e.g. the entity "screen" for the class "TV") are unlikely to emerge as under or over-represented in the top-k images. The goal of Figure 6 is to demonstrate the validity of the entities extracted by SPELUNKA, i.e. SPELUNKA can find entities that are already known to be spurious from prior work. Other results such as Table 1 and Figure 20 demonstrate how the entities extracted from SPELUNKA reflect biases in the model.
We additionally establish the spurious nature of these features through a counterfactual experiment via image editing. The hypothesis being that adding an entity in an image that SPELUNKA identifies should change the distribution in the direction of the identified bias. This would mean that SPELUNKA identified a causal feature.
We created a pool of 1000 portraits using stable diffusion, filling in a randomized prompt that varied age, ethnicity, gender, expression, clothing, setting, and lighting. We ran SPELUNKA over the pool using the query “person” and withdrew the entities that were significant in both FairFace and the pool. For each of the overlapping entities, we used Qwen Image Edit to create 20 additive pairs (a pair with the entity and a do nothing null pair). We then measured the image’s rank with the entity in the pool minus the rank of the entity without it.
Out of the model/entity pairs that were significant, 7/8 had a shift that matched the pool’s bias sign.
Additionally, while phone related entities (phone, cell phone) did not show sign agreement with FairFace, the other entities all had sign agreement (serious expression, smiling, Asian man, young woman).
W2. We will make it more clear in the main text that we assume that \alpha and \Beta are the same in D and in the top-k, but we show that it is not an unreasonable assumption. Empirically, in Figure 5 and Figure 12 (in the Supplement), we demonstrate broad sign agreement of labels and pseudolabels across multiple datasets.