Hide and Seek: HIDE Concepts to SEEK explanations for black box models

What it claims

Hide and Seek (HnS), accepted at ICLR 2027, scores how much a natural-language concept (e.g. "smiling") drives a vision model's output. It does this by hiding that concept in images and measuring the output change, using only forward passes. Beyond the benchmark comparison, the authors report two applied findings:

• Captioning bias. Several state-of-the-art captioning models produced male-centric captions for images containing a tie or a toaster, and female-centric captions for images containing a cake or a hair dryer.
• Segmentation shortcuts. PRUE, which the authors call the current best crop-field segmentation model, relies on relevant concepts such as arable land and also on irrelevant ones such as urban areas.

The authors also claim HnS is the fastest of the compared attribution methods (Appendix A, not in the fetched text).

How it works

1. Encode the concept. The concept prompt (e.g. "a photo of a person smiling") is encoded with SigLIP. Any VLM works, but SigLIP correlated best with ground truth in the VLM ablation.
2. Correct the text vector. Two preprocessing steps are applied: - Modality-gap correction: subtract the mean text embedding of a large, concept-independent word set. - Löwdin orthogonalization: V* = V S^(-1/2) across all concepts, so entangled concepts such as "bald" and "male" get distinct vectors.
3. Retrieve images. Take the top K = 50 images by cosine similarity from an unlabeled image database you supply, for example a face database.
4. Localize the concept. Split each image into SigLIP patches and select patches whose similarity to the concept exceeds τ = μ + 1.5σ, computed over that image's patch scores. That threshold selects about 7% of patches.
5. Hide it. Try 7 fill strategies: blur, zero, white, mean, hue-shift, zero+noise, noise-then-blur. Keep the one where the normalized embedding difference φ(x) − φ(x̃) points most toward the concept vector.
6. Score. - Local attribution = M(x) − M(x̃). - Global attribution = the mean over the K images, excluding badly masked ones (e.g. the whole image hidden), divided by the standard deviation of the local scores. This Cohen's-d-style step pushes down concepts with noisy attributions.

Unlike TCAV and PCBM, which use Δy/Δc, HnS uses only Δy. The authors argue Δc depends on arbitrary positive and negative concept sets and is unstable.

Evidence

• Models: ViT-B-16, ConvNeXt-Tiny and ResNet-18, all ImageNet-pretrained and fine-tuned on each dataset.
• CelebA: classes are Attractive and Male; 26 of the remaining 38 attributes serve as concepts.
• Baselines: TCAV and PCBM.
• Result: HnS achieved the highest correlation with ground-truth attributions on both CelebA and COCO-Stuff.

The fetched text cuts off before the COCO-Stuff setup and the actual correlation values, so no numbers can be given here.

How to use it

What you need:

• Model access: only output probabilities (or an equivalent output for captioning and segmentation) from the black-box model.
• A VLM: SigLIP by default.
• An image database: relevant to your domain.
• A concept list: written as text prompts.

How to read the results:

• Compare rankings, not raw values. The authors say to compare concepts by their relative rankings rather than reading individual attribution values as absolute.
• Image-level explanations: rank concepts by their local scores for that image.

Code availability: code was submitted as anonymous supplementary material. The fetched text gives no public repository, install command or licence.

Limits

• Preprocessing can hurt. Depending on the VLM, modality correction and orthogonalization may be unnecessary or may add noise (Appendix B.1). Check this if you swap SigLIP for another VLM.
• Concepts can be unmeasurable. A concept that the VLM represents poorly, or that is missing from your image database, gives noisy scores. The normalization only down-weights these concepts; it does not fix them.
• Patch-level masking. Localization works on patches with a fixed α. Diffuse or global concepts, and fill artifacts that the model itself reacts to, are not addressed in the fetched text.
• Results not visible here. The ablations (VLM choice, K, α, fill strategies), the speed comparison and the captioning and segmentation experiments are in sections or appendices beyond the fetched text.