Atheer Abdulqader Alboloshi

تخصص الذكاء الاصطناعي · دفعة 2026

AnswerMap: Faithful Black-Box Spatial Interpretability for Vision Language Models

AnswerMap is a training-free, black-box method for spatial interpretability in vision language models. Given an image and any free-text query, it produces a heatmap over that image showing where the model itself locates the evidence, from which two read-outs are taken: a single point (where) and a scalar (how strongly the model commits to that region). The map is elicited from the model's own answers rather than its internals. It poses 2K yes/no questions about the rows and columns of the image, takes the yes posterior from first-token logits, and combines the two marginals by an outer product. A K×K map costs 2K queries, needs no weights, gradients, or attention, and exists before any answer is generated.

اقرأ التقرير النهائي (PDF)