\documentclass[11pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{enumitem}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[hidelinks]{hyperref}
\hypersetup{pdftitle={SteerViT: 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=0.5em}

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

\begin{enumerate}
\item \textbf{What makes a visual encoder steerable?}

An image contains both a cat and a bookshelf. We first use the prompt
``the cat'' and then ``the bookshelf.''

Explain the difference between changing the text used to score a fixed
CLIP image representation and changing the visual representation itself
in SteerViT. At what stage does the prompt influence the image processing
in each approach?

\newpage
\item \textbf{How does text enter the visual backbone?}

Describe the path from a written prompt to the cross-attention updates
inside SteerViT. Include the text encoder, the trainable text adapter,
and the roles of image queries and text keys and values.

Why is it useful to retain multiple contextualized prompt-token vectors
rather than reduce the entire prompt to one vector?

\newpage
\item \textbf{What does ``frozen visual backbone'' mean?}

SteerViT freezes the original visual Transformer's parameters, yet its
output changes with the prompt.

Explain how both statements can be true. Distinguish between fixed model
weights and fixed intermediate representations. Can gradients pass
through frozen visual layers to train an adapter inserted earlier in
the network?

\newpage
\item \textbf{Why train with segmentation supervision?}

SteerViT is trained using an image, a referring expression, and a mask
identifying the referred object.

Explain how this supervision teaches the model which visual regions
should respond to the prompt. Why might supervising the object's full
spatial extent provide a richer training signal than supervising only
its center point?

\newpage
\item \textbf{How do we distinguish useful visual steering from merely
responding to text?}

A model retrieves the correct object when given an image and an
appropriate prompt. Explain why this result alone does not establish
that the model is a useful steerable visual encoder.

Describe what each of the following tests would reveal:
\begin{enumerate}
\item Keep an image containing two objects fixed and switch the prompt
from one object to the other. Also test an irrelevant prompt. What would
changed retrieval results demonstrate, and why would a performance drop
with an irrelevant prompt alone be insufficient?
\item Retrieve among visually similar objects using a detailed prompt,
such as ``the red mug with a white handle.'' What capability does this
test beyond selecting a broad object category?
\end{enumerate}
\end{enumerate}
\end{document}