\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={Chain of Visual Thoughts: Oral Knowledge Questions}}
\setlength{\parindent}{0pt}
\setlength{\emergencystretch}{2em}
\setlist[enumerate,1]{leftmargin=*,itemsep=1.1em,topsep=0.7em}
\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 Chain of Visual Thoughts\par}
\vspace{0.3em}
{\Large\bfseries Oral Knowledge Questions\par}
\vspace{0.6em}
Explain each answer in your own words.

\begin{enumerate}
\item \textbf{What makes a continuous thought different from a text token?}

Explain how COCONUT produces the next reasoning input without selecting
a vocabulary token.

\medskip
Does each continuous thought necessarily correspond to a particular
word or sentence? Why are continuous thoughts harder to inspect than
textual chain-of-thought?

\newpage
\item \textbf{Why does CoVT learn several types of visual thought tokens?}

Consider these questions:
\begin{itemize}
\item ``How many objects are present?''
\item ``Which object is closer to the camera?''
\item ``What is the outline of this object?''
\end{itemize}
For each question, identify a relevant type of visual information learned
by CoVT. Why might one type of visual representation be insufficient for
all three tasks?

\newpage
\item \textbf{Why is learning to generate visual thought tokens not enough?}

Suppose CoVT generates visual thought tokens that can be decoded into
accurate object masks and depth maps.
\begin{enumerate}
\item Does this guarantee that the model can correctly answer ``Which
object is closer to the camera?'' Explain the difference between
representing useful visual information and using it to answer a question.
\item What additional ability should the \textbf{reasoning stage} teach
after the model has learned to generate visual thought tokens?
\item In the \textbf{efficient-reasoning stage}, some groups of visual
thought tokens are randomly dropped. How might this help the model
avoid depending too heavily on one type of visual information?
\end{enumerate}

\newpage
\item \textbf{Why transfer intact-image latent thoughts to the corrupted path?}

In CrystaL, the corrupted path receives both a degraded image and latent
thoughts from the intact path.
\begin{enumerate}
\item What is the intended role of the transferred latent thoughts?
\item Why does degrading the image create pressure for the model to
use those thoughts?
\item Why is this different from simply training a model to answer
questions about noisy images without transferring any latent information?
\end{enumerate}

\newpage
\item \textbf{Does consistency guarantee correct visual reasoning?}

Suppose CrystaL's intact and corrupted paths produce almost identical
answer distributions and similar attention to latent tokens.
\begin{enumerate}
\item Does this establish that the answer is correct? Give an example
in which both paths agree but are wrong.
\item Why does CrystaL also use answer cross-entropy losses rather than
relying only on consistency between the paths?
\item To test whether transferred latent thoughts carry useful visual
evidence, what would you expect to happen if they were removed or
replaced with thoughts from an unrelated image? Explain why this is a
useful experimental check.
\end{enumerate}
\end{enumerate}
\end{document}
