Cristian Minoccheri

Seemingly minor architectural choices in vision-language models, like mean vs. attention pooling, drastically impact performance on different CT enterography tasks, suggesting a need for task-specific representation engineering.