\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={Grounded VLMs: Oral Knowledge Questions},pdfsubject={Grad-CAM and Grounding DINO}}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate]{leftmargin=*,itemsep=1.2em,topsep=1em}
\setlist[itemize]{label=$\bullet$,leftmargin=*,itemsep=0.4em,topsep=0.5em}

\begin{document}
\raggedright
{\LARGE\bfseries Grounded VLMs\par}
\vspace{0.25em}
{\Large\bfseries Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words and justify the underlying mechanism.

\begin{enumerate}
\item \textbf{Why is Grad-CAM class-specific?}

An image contains both a cat and a dog. Explain how Grad-CAM can produce
one heatmap for ``cat'' and a different heatmap for ``dog'' from the same
image.

Which quantities remain unchanged, and which must be recomputed when
the target class changes? Does generating the second heatmap require
retraining the CNN?

\newpage
\item \textbf{Does a heatmap prove that a model uses the right evidence?}

A CNN correctly predicts ``cow,'' but its Grad-CAM heatmap mainly
highlights the surrounding grass.

What possible shortcut does this suggest? Why does the heatmap alone
not prove that the grass caused the prediction?

Describe a controlled test using images in which the cow is retained
but the background changes. What result would strengthen the
shortcut-learning hypothesis?

\newpage
\item \textbf{Why does Grad-CAM use convolutional feature maps?}

Why does Grad-CAM need features that retain spatial information about
the image?

Explain why the last convolutional layer is commonly used rather than
an early convolutional layer or a final classification vector. What is
the trade-off between meaningful object features and spatial detail?

\newpage
\item \textbf{How does Grounding DINO combine image and text information?}

Describe the roles of the feature enhancer, language-guided query
selection, and cross-modality decoder in processing an image and a
text prompt.

In your explanation, distinguish:
\begin{itemize}
\item Interaction between image and text features.
\item Selection of promising image positions.
\item Refinement of object queries into detections.
\end{itemize}
What does the model output, and how does the role of text differ from
a detector restricted to a fixed set of trained category labels?

\newpage
\item \textbf{How does deformable attention differ from standard attention?}

Standard attention compares a query with every image position.
Deformable attention instead samples a small number of locations.

Explain this difference in your own words. Why can sampling fewer
locations make attention more efficient when processing a large image
feature map?
\end{enumerate}
\end{document}