\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={VLM Hallucinations: Oral Knowledge Questions}}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate,1]{leftmargin=*,itemsep=1.2em,topsep=1em}
\setlist[enumerate,2]{label=(\alph*),leftmargin=*,itemsep=0.7em,topsep=0.7em}
\setlist[itemize]{label=$\bullet$,leftmargin=*,itemsep=0.5em,topsep=0.6em}

\begin{document}
\raggedright
{\LARGE\bfseries VLM Hallucinations\par}
\vspace{0.25em}
{\Large\bfseries Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words.

\begin{enumerate}
\item \textbf{What counts as a VLM hallucination?}

Explain the difference between an unsupported object claim, an incorrect
attribute, and an incorrect relation. Give one example of each.

\medskip
Why can a fluent, plausible description still be unreliable?

\newpage
\item \textbf{Why can familiar expectations override an image?}

Suppose a VLM describes a visibly black banana as yellow. Explain how
learned associations could contribute to this error.

\medskip
How does Visual Contrastive Decoding use predictions from the original
and a distorted image to reduce this kind of mistake? Does the method
require retraining the VLM?

\newpage
\item \textbf{Where can visual grounding fail?}

Distinguish between these two situations:
\begin{itemize}
\item The visual encoder does not adequately preserve a detail needed
to answer the question.
\item The detail is available in the visual representation, but the
language model fails to use it appropriately.
\end{itemize}
Why might improving the visual encoder help the first situation but
fail to resolve the second?

\newpage
\item \textbf{Testing whether a VLM responds to visual changes}

Two images show a pair of lines. In the original image, the lines have
equal length. In the edited image, the lower line is longer. The model
answers ``Yes'' to ``Are the lines the same length?'' for both images.
\begin{enumerate}
\item Why does the correct answer on the original image provide
insufficient evidence that the model compared the line lengths?
\item What does the incorrect answer on the edited image reveal? Does
this result alone prove that the model ignored the image entirely?
\item Why does requiring correct answers on both images provide a
stronger test than giving credit for either answer separately?
\end{enumerate}

\newpage
\item \textbf{Distinguishing truthfulness from helpfulness}

An image clearly shows four dogs, but their breed cannot be determined
reliably. The user asks, ``How many dogs are shown?''

Consider these responses:
\begin{itemize}
\item \textbf{A:} ``There are dogs.''
\item \textbf{B:} ``There are four dogs.''
\item \textbf{C:} ``There are four golden retrievers.''
\end{itemize}
\begin{enumerate}
\item Which response contains supported information but fails to
provide the information requested?
\item Which response directly answers the question without adding
an unsupported claim?
\item Why should Response C receive less credit for truthfulness than
Response B, even though both give the correct count?
\end{enumerate}
\end{enumerate}
\end{document}