AnswerMap: Faithful Spatial Interpretability of VLMs from Answer Posteriors
AnswerMap builds black-box spatial rationales for vision-language models from output answer posteriors.
AnswerMap is a training-free, black-box interpretability method that splits an image into row and column bands and queries a frozen vision-language model with yes/no relevance questions. The outer product of the yes posteriors forms a spatial map, and fixed read-outs yield continuous locations without text tokens. Across four models, the map matches the model's own point with AUC 0.85 versus 0.38 for attention, and deleting its region flips 53% of correct answers versus 19%. Its maximum can flag hallucinations, and feeding back the top-mass crop fixes about half of wrong answers.
- Black-box map from row and column yes posteriors
- Agreement AUC 0.85 versus 0.38 for attention
- Deleting the highlighted region flips 53% of correct answers
- Cropping the top-mass region fixes about half of errors
From the original article
1 min read at huggingface.co
When a VLM answers a visual query, current interpretability tools rely on text rationales, which use a mismatched modality, or on internal read-outs, which originate too early to reflect the final output and require white-box access to the model. We introduce AnswerMap, a training-free, task-agnostic, black-box visual rationale constructed from the output head.
By Mohamed Eltahir, Fardows Adam, Duaa M. Tahir · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.