\documentclass[11pt]{article}

\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}

\title{Vision Transformers: Oral Knowledge Questions}
\author{}
\date{}

\begin{document}

\maketitle

\begin{enumerate}[leftmargin=*, itemsep=1.2em]
    \item Why can a Vision Transformer not determine the original spatial
    arrangement of image patches from patch content alone? Explain how
    positional embeddings solve this problem.

    \item Explain the roles of queries, keys, and values without simply
    reciting the attention equation. What determines the attention weights,
    and what information is actually combined?

    \item Inside a Vision Transformer encoder block, how do multi-head
    self-attention and the MLP perform different jobs? Which dimension does
    each operation mix?

    \item Explain how the final $[\mathrm{CLS}]$ token becomes an image-class
    prediction. In
    \[
        p(y=c\mid\text{image})
        = \operatorname{softmax}(\text{logits})_c,
    \]
    what does $c$ mean?

    \item If the patch size is reduced from $16\times16$ to $8\times8$, what
    happens to spatial detail, sequence length, and attention cost? Explain the
    trade-off rather than giving only the formulas.
\end{enumerate}

\end{document}