Inside one decision
How the model reaches its answer
MedGemma is a vision-language model that reads a chest X-ray and outputs findings. Pick a case on the left; this map shows the frozen-attention attribution for that one finding: the features that contributed most, flowing left to right into the answer.
What am I looking at?
- Each dot is a feature: a specific, sparse pattern we isolated from inside the model (see the Features tab for named examples). The dot on the far right is the answer (the finding, e.g. cardiomegaly).
- Warm dots come from the X-ray image; cool dots come from the question it was asked.
- Lines are each feature's effect on the answer. The model is really deciding yes-or-no on the finding: orange makes it lean yes (the finding is present), blue makes it lean no. Thicker lines mean a stronger effect.
- Columns run left to right, from the earliest to the latest stage of the model's processing. Hover a dot to light up its connections; click it for details and the other cases it shows up in.
click a dot for details · scroll to zoom · drag to pan
The building blocks of the reasoning
What the model has learned to notice
These are visual patterns MedGemma has learned to represent. By taking the model apart, we pulled out individual features, each one responding selectively to a specific thing in the image. Every card is one feature, and the thumbnails are the X-ray patches that switch it on. Most turn out to be real anatomy: lung bases, shoulder joints, tubes and wires, even the “AP PORTABLE” text stamped on portable films.
loading features…