\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate,1]{leftmargin=*,itemsep=1.1em,topsep=0.7em}
\setlist[enumerate,2]{label=(\alph*),leftmargin=*,itemsep=0.6em,topsep=0.5em}
\begin{document}
\raggedright
{\LARGE\bfseries AdaptVis and Entropy-Gradient Grounding\par}
\vspace{0.3em}
{\Large\bfseries Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words.
\begin{enumerate}
\item \textbf{Why is more image attention not necessarily better?}

A VLM assigns most of its attention to text tokens. We increase the total
attention assigned to image tokens, but its spatial-reasoning accuracy
does not improve.
\begin{enumerate}
\item Why is this outcome possible?
\item For the question \lq\lq{}Where is the book relative to the candle?'',
what should useful visual attention capture beyond simply assigning
more weight to the image?
\end{enumerate}

\newpage
\item \textbf{How does AdaptVis decide whether to sharpen or smooth attention?}

Explain the sequence of steps AdaptVis follows, starting with the
baseline model's answer confidence and ending with a new answer.

\medskip
Why might smoothing help when the model focuses on the wrong region?
Why could sharpening be harmful when the model is confidently wrong?

\newpage
\item \textbf{How can a training-free method use backpropagation?}

Entropy-Gradient Grounding backpropagates next-token entropy to
visual-token embeddings while keeping the VLM frozen.
\begin{enumerate}
\item What quantity is differentiated, and with respect to what?
\item What does a large gradient magnitude tell us about a visual token?
\item Why does computing these gradients not constitute training the model?
\end{enumerate}

\newpage
\item \textbf{Why use multiple crops and retain the global image?}

A question about a document requires information from two distant regions.
\begin{enumerate}
\item Why might one crop miss information needed for the answer?
\item How can multiple enlarged crops help?
\item Why should the original image remain available even when both
relevant regions have been cropped?
\end{enumerate}

\newpage
\item \textbf{What are the two different entropies in Entropy-Gradient Grounding?}

The method uses both \textbf{next-token entropy} and \textbf{spatial entropy}.
\begin{enumerate}
\item For next-token entropy, what outcomes form the probability
distribution, and how does the method use this entropy to locate evidence?
\item For spatial entropy, the probabilities are the relative areas of
connected active regions. What does this entropy measure, and how does
it help decide when to stop refining crops?
\item If spatial entropy decreases, does that necessarily mean the model's
answer has become more correct? Explain.
\end{enumerate}
\end{enumerate}
\end{document}