AnswerMap: Faithful Spatial Interpretability of VLMs from Answer Posteriors

AnswerMap. Faithful Spatial Interpretability of VLMs from Answer Posteriors. By Mohamed Eltahir and eight others, from King Abdullah University of Science and Technology and Edge Hill University. Published on the twenty-eighth of September. Its code repository is public and real: eighteen source files.

When a vision-language model answers a question about an image, current interpretability tools explain that answer in one of two ways, and both have a drawback. Text rationales use a mismatched modality: words, offered as the explanation for a spatial decision. Internal read-outs come from too early in the computation to reflect the final output, and they also require white-box access to the model, which is not always available.

AnswerMap builds a visual rationale from the output head instead. It needs no training, it is agnostic to the task, and it treats the model as a black box. The image is cut into K bands of rows and K bands of columns. Each band is shown on its own to the frozen model, together with the query, phrased as a yes or no question about relevance. The outer product of the yes posteriors for the rows and the columns gives a spatial map, conditioned on the query. Where both a row and a column say yes, the map lights up.

The map can also produce answers. Define a fixed read-out on top of it, such as the expectation or the maximum, and continuous outputs like a location come out natively. That bypasses the reliance on discrete text tokens for tasks whose outputs are continuous, and it guarantees by construction that the answer depends on the image, since the map is built from nothing but the model's responses to pieces of that image.

A rationale can be confabulated, so the authors validate AnswerMap across four models and three query distributions with two tests. The first asks whether the map agrees with the point the model itself generates. It does, with an AUC of zero point eight five, against zero point three eight for attention.

The second test deletes the map's region from the image. That flips fifty-three percent of correct answers, against nineteen percent when the region picked out by attention is deleted instead. The two tests check faithfulness from opposite ends. One asks whether the map agrees with what the model says. The other asks whether the model actually needs what the map points at. Attention, the point of comparison in both tests, falls well short on each.

Beyond faithfulness, three read-outs show what the map is good for. Its maximum flags hallucinated objects without any generation. Its expectation localizes correctly in cases where the model's own pointing fails. And its region of highest mass, cropped and fed back to the model, fixes half of the model's wrong answers. Each of these uses the same map, read a different way, which is what task-agnostic means in practice here.

AnswerMap is aimed at two readers at once. For interpretability, it is a rationale whose faithfulness is tested rather than assumed, by checking both where it lands and what happens when that region disappears. For anyone building on vision-language models, its read-outs amount to a new output interface for visual tasks, one that does not have to pass through text tokens at all.