失敗モード：再現率の見逃しとモダリティ間ハルシネーションのリスク

This chapter examines a failure in which the router misses an image that should have been improved. The presenters use a case where the visible count does not match the written description to explain why recall and faithfulness matter.

What a recall miss means

Here, a recall miss means that the router passes an image as acceptable even though it should have been caught and sent to the appropriate processing path because it has a problem. Recall is the proportion of cases that should be caught that are actually caught. When a miss occurs, the problem remains before the later enhancement step.

Six wings and a description of eight

In the presenters' example, only six chicken wings are visible in the image. However, the description attached to the image says "eight chicken wings." The router passes this combination and allows the image to proceed to enhancement. In other words, it fails to detect the mismatch between the image and the description.

Figure: A slide titled "Routing Failures: Recall Miss." It places an image of six chicken wings beside a description of an eight-piece dish, while the dish match is shown as passing.

This material shows that an image can look plausible without its actual content matching the written description. The slide marks "8 pieces" as passing for an image that contains only six pieces. This is therefore not merely an image-quality problem; it is a failure to check content during routing.

Compare the two modalities

The image and the description are different modalities. The image represents what is actually visible, while the description or metadata represents the expected content. If we evaluate only one of them, we may check whether the image looks natural or whether the description has the right form, but we cannot tell whether both refer to the same dish. The router needs to compare what appears in the image with what the description or metadata requires.

The danger of editing in the missing two

After this miss, an editing model may try to match the description's "eight" by adding two wings that were not visible. This is the risk of cross-modal hallucination: the model creates content that it did not observe. Even an enhancement process loses faithfulness to the original image if it creates new content to satisfy the text.

Image quality and content correctness are different

The important point in this example is to separate visual appeal from content correctness. An image can have natural colors and composition and look attractive, but it is not correct if the dish count differs from reality. Conversely, being faithful to the original image does not simply mean making only a small pixel-level change. It also means not arbitrarily changing the dish type, count, or visible elements.

Evaluate recall and faithfulness separately

A recall miss is a failure of the router to catch an image with a problem. If two wings are later added to that image, the edited result also fails faithfulness. The first is a routing question—"Which path should this image have taken?" The second is a generation-result question—"Does the edited content remain faithful to the original image and description?" They occur in the same example, but they should be recorded as separate evaluations.

Production lessons

• The router uses not only the image but also the description and metadata.
• When the image and text disagree, a visually appealing result is not enough to send the case safely to enhancement.
• Instead of asking the editing model to fill in uncertain content, evaluate both recall and faithfulness and stop the problem early.