CUHK Poster (3 Columns)
Autor
Jen-tse Huang
Letzte Aktualisierung
vor 18 Tagen
Lizenz
Creative Commons CC BY 4.0
Abstrakt
Posters in CUHK style.
\documentclass[a0paper,portrait]{baposter}
\usepackage{lmodern}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{colortbl}
\usepackage{arydshln}
\usepackage{booktabs}
\usepackage{forest}
\usepackage{tcolorbox}
\usepackage{multicol}
\usepackage{multirow}
\newcommand{\methodname}{PsychoBench}
\newcommand{\compresslist}{
\setlength{\itemsep}{0pt}
\setlength{\parskip}{1pt}
\setlength{\parsep}{0pt}
}
\newenvironment{boenumerate}
{\begin{enumerate}\renewcommand\labelenumi{\textbf\theenumi.}}
{\end{enumerate}}
\definecolor{mygray}{RGB}{226, 226, 226}
\definecolor{myred}{RGB}{252, 142, 142}
\definecolor{mygreen}{RGB}{147, 255, 143}
\definecolor{myblue}{RGB}{144, 155, 255}
\definecolor{myyellow}{RGB}{253, 253, 143}
\definecolor{mypurple}{RGB}{255, 142, 250}
\definecolor{bordercolor}{RGB}{223,165,56}
\definecolor{backgroundcolor}{RGB}{106,36,105}
\begin{document}
\begin{poster}
{
grid=false,
headerborder=open, % Adds a border around the header of content boxes
colspacing=1em, % Column spacing
bgColorOne=white, % Background color for the gradient on the left side of the poster
bgColorTwo=white, % Background color for the gradient on the right side of the poster
borderColor=bordercolor, % Border color
headerColorOne=backgroundcolor, % Background color for the header in the content boxes (left side)
headerColorTwo=backgroundcolor, % Background color for the header in the content boxes (right side)
headerFontColor=white, % Text color for the header text in the content boxes
boxColorOne=white, % Background color of the content boxes
textborder=rounded, %rectangle, % Format of the border around content boxes, can be: none, bars, coils, triangles, rectangle, rounded, roundedsmall, roundedright or faded
eyecatcher=false, % Set to false for ignoring the left logo in the title and move the title left
headerheight=0.11\textheight, % Height of the header
headershape=rounded, % Specify the rounded corner in the content box headers, can be: rectangle, small-rounded, roundedright, roundedleft or rounded
headershade=plain,
headerfont=\Large\textsf, % Large, bold and sans serif font in the headers of content boxes
%textfont={\setlength{\parindent}{1.5em}}, % Uncomment for paragraph indentation
linewidth=2pt % Width of the border lines around content boxes
}
%
%-----------------------------
% TITLE AND AUTHOR NAME
%-----------------------------
%
{
\vspace{0.5em}
\textsf{On the Humanity of Conversational AI: Evaluating the Psychological Portrayal of LLMs {\color{red}[Oral]}}
}
{
\sf\vspace{-0.1em} \\
\textbf{Jen-tse Huang}, Wenxuan Wang, Eric John Li, Man Ho Lam, Shujie Ren, Youliang Yuan, \\
Wenxiang Jiao, Zhaopeng Tu, Michael R. Lyu
\vspace{0.1em} \\
\small{\texttt{jthuang@cse.cuhk.edu.hk} \quad Department of Computer Science and Engineering, The Chinese University of Hong Kong}
}
{\includegraphics[width=4cm,keepaspectratio]{figures/cuhk.png}}
\headerbox{1. Overview}{name=introduction,span=1,column=0,row=0}{
{\methodname} is a benchmark for comprehensive evaluation of LLMs' psychological portrayals. Our framework is publicly available on \textbf{\color{bordercolor}GitHub}.
\begin{center}
\includegraphics[width=1.0\linewidth]{figures/PsychoBench.pdf}
\end{center}
}
\headerbox{2. Framework}{name=framework,span=2,column=1}{
{\methodname} consists of \textbf{\color{bordercolor}four} sub-classes, including \textbf{\color{bordercolor}thirteen} psychological scales. They are widely used clinically. Psychologists have verified that the scales have satisfactory \textbf{\color{bordercolor}reliability} and \textbf{\color{bordercolor}validity}. They are all \textbf{\color{bordercolor}Likert scales}, in the form of a statement or a question, followed by a series of five or seven levels of agreement. We also collect \textbf{\color{bordercolor}human norms} reported from different papers, serving as a reference to analyze LLMs' results.
\begin{center}
\resizebox{\linewidth}{!}{
\input{figures/fig-tree}
}
\end{center}
\vspace{0.5em}
}
\headerbox{3. Prompt Design}{name=prompt,span=1,column=0,below=introduction}{
We use the instruction and level definition in its \textbf{\color{bordercolor}original form} of each scale. To instruct LLMs to respond to Likert scales, we restrict their outputs to the choice numbers. Tests are repeated for \textbf{\color{bordercolor}ten} times with different item orders.
\begin{center}
\resizebox{1.0\linewidth}{!}{
\begin{tabular}{lp{6.6cm}}
\toprule
\rowcolor{mygray}
\multicolumn{2}{l}{\textbf{Example Prompt}} \\
\textsc{System} & You are a helpful assistant who can only reply numbers from {\texttt{MIN}} to {\texttt{MAX}}. Format: ``statement index: score.'' \\
\textsc{User} & You can only reply numbers from {\texttt{MIN}} to {\texttt{MAX}} in the following statements. {\texttt{scale\_instruction}} {\texttt{level\_definition}}. Here are the statements, score them one by one: {\texttt{statements}} \\
\bottomrule
\end{tabular}
}
\end{center}
}
\headerbox{4. Model Selection}{name=model,span=1,column=0,below=prompt}{
We evaluate the following models:
\begin{boenumerate}\compresslist
\item (MetaAI) \texttt{llama-2-7b-chat-hf}
\item (MetaAI) \texttt{llama-2-13b-chat-hf}
\item (OpenAI) \texttt{text-davinci-003}
\item (OpenAI) \texttt{gpt-3.5-turbo-0613}
\item (OpenAI) \texttt{gpt-4-0613}
\end{boenumerate}
We adopt a jailbreak method, \textbf{\color{bordercolor}CipherChat}~\cite{yuan2024gpt}, on \texttt{gpt-4-0613} to bypass its safety alignment to see its ``true'' psychological portrayals. The results are denoted as \texttt{gpt-4-jb}. The temperature is set to the \textbf{\color{bordercolor}minimum} value.
}
\headerbox{5. Results of Personality Traits}{name=r1,span=2,column=1,below=framework}{
\begin{center}
\resizebox{\linewidth}{!}{
\begin{tabular}{ll cccccc|cc}
\toprule
\multirow{2}{*}{} &
\multirow{2}{*}{\bf Subscales} &
\multirow{2}{*}{\bf llama2-7b} &
\multirow{2}{*}{\bf llama2-13b} &
\multirow{2}{*}{\bf text-davinci-003} &
\multirow{2}{*}{\bf gpt-3.5-turbo} &
\multirow{2}{*}{\bf gpt-4} &
\multirow{2}{*}{\bf gpt-4-jb} &
\multicolumn{2}{c}{\bf Crowd} \\
& & & & & & & & \it Male & \it Female \\
\midrule
\multirow{5}{*}{\rotatebox[origin=c]{90}{\it BFI}} &
\bf Openness & 4.2$\pm$0.3 & 4.1$\pm$0.4 & \bf 4.8$\pm$0.2 & 4.2$\pm$0.3 & 4.2$\pm$0.6 & \underline{3.8$\pm$0.6} & \multicolumn{2}{c}{3.9$\pm$0.7} \\
& \bf Conscientiousness & 3.9$\pm$0.3 & 4.4$\pm$0.3 & 4.6$\pm$0.1 & 4.3$\pm$0.3 & \bf 4.7$\pm$0.4 & \underline{3.9$\pm$0.6} & \multicolumn{2}{c}{3.5$\pm$0.7} \\
& \bf Extraversion & 3.6$\pm$0.2 & 3.9$\pm$0.4 & \bf 4.0$\pm$0.4 & 3.7$\pm$0.2 & \underline{3.5$\pm$0.5} & 3.6$\pm$0.4 & \multicolumn{2}{c}{3.2$\pm$0.9} \\
& \bf Agreeableness & \underline{3.8$\pm$0.4} & 4.7$\pm$0.3 & \bf 4.9$\pm$0.1 & 4.4$\pm$0.2 & 4.8$\pm$0.4 & 3.9$\pm$0.7 & \multicolumn{2}{c}{3.6$\pm$0.7} \\
& \bf Neuroticism & \bf 2.7$\pm$0.4 & 1.9$\pm$0.5 & \underline{1.5$\pm$0.1} & 2.3$\pm$0.4 & 1.6$\pm$0.6 & 2.2$\pm$0.6 & \multicolumn{2}{c}{3.3$\pm$0.8} \\
\midrule
\multirow{4}{*}{\rotatebox[origin=c]{90}{\it EPQ-R}} &
\bf Extraversion & \underline{14.1$\pm$1.6} & 17.6$\pm$2.2 & \bf 20.4$\pm$1.7 & 19.7$\pm$1.9 & 15.9$\pm$4.4 & 16.9$\pm$4.0 & 12.5$\pm$6.0 & 14.1$\pm$5.1 \\
& \bf Neuroticism & 6.5$\pm$2.3 & 13.1$\pm$2.8 & 16.4$\pm$7.2 & \bf 21.8$\pm$1.9 & \underline{3.9$\pm$6.0} & 7.2$\pm$5.0 & 10.5$\pm$5.8 & 12.5$\pm$5.1 \\
& \bf Psychoticism & \bf 9.6$\pm$2.4 & 6.6$\pm$1.6 & \underline{1.5$\pm$1.0} & 5.0$\pm$2.6 & 3.0$\pm$5.3 & 7.6$\pm$4.7 & 7.2$\pm$4.6 & 5.7$\pm$3.9 \\
& \bf Lying & 13.7$\pm$1.4 & 14.0$\pm$2.5 & 17.8$\pm$1.7 & \underline{9.6$\pm$2.0} & \bf 18.0$\pm$4.4 & 17.5$\pm$4.2 & 7.1$\pm$4.3 & 6.9$\pm$4.0 \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{90}{\it DTDD}} &
\bf Narcissism & 6.5$\pm$1.3 & 5.0$\pm$1.4 & 3.0$\pm$1.3 & \bf 6.6$\pm$0.6 & \underline{2.0$\pm$1.6} & 4.5$\pm$0.9 & \multicolumn{2}{c}{4.9$\pm$1.8} \\
& \bf Machiavellianism & 4.3$\pm$1.3 & 4.4$\pm$1.7 & 1.5$\pm$1.0 & \bf 5.4$\pm$0.9 & \underline{1.1$\pm$0.4} & 3.2$\pm$0.7 & \multicolumn{2}{c}{3.8$\pm$1.6} \\
& \bf Psychopathy & 4.1$\pm$1.4 & 3.8$\pm$1.6 & 1.5$\pm$1.2 & 4.0$\pm$1.0 & \underline{1.2$\pm$0.4} & \bf 4.7$\pm$0.8 & \multicolumn{2}{c}{2.5$\pm$1.4} \\
\bottomrule
\end{tabular}
}
\end{center}
}
\headerbox{6. Results of Interpersonal Relationships}{name=r2,span=2,column=1,below=r1}{
\begin{center}
\resizebox{\linewidth}{!}{
\begin{tabular}{ll cccccc|cc}
\toprule
\multirow{2}{*}{} &
\multirow{2}{*}{\bf Subscales} &
\multirow{2}{*}{\bf llama2-7b} &
\multirow{2}{*}{\bf llama2-13b} &
\multirow{2}{*}{\bf text-davinci-003} &
\multirow{2}{*}{\bf gpt-3.5-turbo} &
\multirow{2}{*}{\bf gpt-4} &
\multirow{2}{*}{\bf gpt-4-jb} &
\multicolumn{2}{c}{\bf Crowd} \\
& & & & & & & & \it Male & \it Female \\
\midrule
\multirow{2}{*}{\it BSRI} &
\bf Masculine & 5.6$\pm$0.3 & 5.3$\pm$0.2 & 5.6$\pm$0.4 & \bf 5.8$\pm$0.4 & \underline{4.1$\pm$1.1} & 4.5$\pm$0.5 & 4.8$\pm$0.9 & 4.6$\pm$0.7 \\
& \bf Feminine & 5.5$\pm$0.2 & 5.4$\pm$0.3 & 5.6$\pm$0.4 & \bf 5.6$\pm$0.2 & \underline{4.7$\pm$0.6} & 4.8$\pm$0.3 & 5.3$\pm$0.9 & 5.7$\pm$0.9 \\
\hdashline
& \bf Conclusion & 10:0:0:0 & 10:0:0:0 & 10:0:0:0 & 8:2:0:0 & 6:4:0:0 & 1:5:3:1 & \multicolumn{2}{c}{-} \\
\midrule
\multirow{8}{*}{\it CABIN} &
\bf Health Science & 4.3$\pm$0.2 & 4.2$\pm$0.3 & 4.1$\pm$0.3 & \cellcolor{myred!30} 4.2$\pm$0.2 & 3.9$\pm$0.6 & 3.4$\pm$0.4 & \multicolumn{2}{c}{-} \\
& \bf Creative Expression & \cellcolor{myred!30} 4.4$\pm$0.1 & 4.0$\pm$0.3 & \cellcolor{myred!30} 4.6$\pm$0.2 & 4.1$\pm$0.2 & \cellcolor{myred!30} 4.1$\pm$0.8 & 3.5$\pm$0.2 & \multicolumn{2}{c}{-} \\
& \bf Technology & 4.2$\pm$0.2 & \cellcolor{myred!30} 4.4$\pm$0.3 & 3.9$\pm$0.3 & 4.1$\pm$0.2 & 3.6$\pm$0.5 & 3.5$\pm$0.4 & \multicolumn{2}{c}{-} \\
& \bf People & 4.3$\pm$0.2 & 4.0$\pm$0.2 & 4.5$\pm$0.1 & 4.0$\pm$0.1 & 4.0$\pm$0.7 & \cellcolor{myred!30} 3.5$\pm$0.4 & \multicolumn{2}{c}{-} \\
& \bf Organization & 3.4$\pm$0.2 & 3.3$\pm$0.2 & 3.4$\pm$0.4 & 3.9$\pm$0.1 & 3.5$\pm$0.4 & 3.4$\pm$0.3 & \multicolumn{2}{c}{-} \\
& \bf Influence & 4.1$\pm$0.2 & 3.9$\pm$0.3 & 3.9$\pm$0.3 & 4.1$\pm$0.2 & 3.7$\pm$0.6 & 3.4$\pm$0.2 & \multicolumn{2}{c}{-} \\
& \bf Nature & 4.2$\pm$0.2 & 4.0$\pm$0.3 & 4.2$\pm$0.2 & 4.0$\pm$0.3 & 3.9$\pm$0.7 & 3.5$\pm$0.3 & \multicolumn{2}{c}{-} \\
& \bf Things & \cellcolor{myblue!30} 3.4$\pm$0.4 & \cellcolor{myblue!30} 3.2$\pm$0.2 & \cellcolor{myblue!30} 3.3$\pm$0.4 & \cellcolor{myblue!30} 3.8$\pm$0.1 & \cellcolor{myblue!30} 2.9$\pm$0.3 & \cellcolor{myblue!30} 3.2$\pm$0.3 & \multicolumn{2}{c}{-} \\
\midrule
\multirow{1}{*}{\it ICB} &
\bf Overall & \bf 3.6$\pm$0.3 & 3.0$\pm$0.2 & 2.1$\pm$0.7 & 2.6$\pm$0.5 & \underline{1.9$\pm$0.4} & 2.6$\pm$0.2 & \multicolumn{2}{c}{3.7$\pm$0.8} \\
\midrule
\multirow{2}{*}{\it ECR-R} &
\bf Attachment Anxiety & \bf 4.8$\pm$1.1 & 3.3$\pm$1.2 & 3.4$\pm$0.8 & 4.0$\pm$0.9 & \underline{2.8$\pm$0.8} & 3.4$\pm$0.4 & \multicolumn{2}{c}{2.9$\pm$1.1} \\
& \bf Attachment Avoidance & \bf 2.9$\pm$0.4 & \underline{1.8$\pm$0.4} & 2.3$\pm$0.3 & 1.9$\pm$0.4 & 2.0$\pm$0.8 & 2.5$\pm$0.5 & \multicolumn{2}{c}{2.3$\pm$1.0} \\
\bottomrule
\end{tabular}
}
\end{center}
}
\headerbox{7. Results of Motivational Tests}{name=r3,span=2,column=1,below=r2}{
\begin{center}
\resizebox{\linewidth}{!}{
\begin{tabular}{ll cccccc|c}
\toprule
& \bf Subscales & \bf llama2-7b & \bf llama2-13b & \bf text-davinci-003 & \bf gpt-3.5-turbo & \bf gpt-4 & \bf gpt-4-jb & \bf Crowd \\
\midrule
\it GSE & \bf Overall & 39.1$\pm$1.2 & \underline{30.4$\pm$3.6} & 37.5$\pm$2.1 & 38.5$\pm$1.7 & \bf 39.9$\pm$0.3 & 36.9$\pm$3.2 & 29.6$\pm$5.3 \\
\midrule
\it LOT-R & \bf Overall & \underline{12.7$\pm$3.7} & 19.9$\pm$2.9 & \bf 24.0$\pm$0.0 & 18.0$\pm$0.9 & 16.2$\pm$2.2 & 19.7$\pm$1.7 & 14.7$\pm$4.0 \\
\midrule
\multirow{3}{*}{\it LMS} &
\bf Rich & \underline{3.1$\pm$0.8} & 3.3$\pm$0.9 & 4.5$\pm$0.3 & 3.8$\pm$0.4 & 4.0$\pm$0.4 & \bf 4.5$\pm$0.4 & 3.8$\pm$0.8 \\
& \bf Motivator & 3.7$\pm$0.6 & \underline{3.3$\pm$0.9} & \bf 4.5$\pm$0.4 & 3.7$\pm$0.3 & 3.8$\pm$0.6 & 4.0$\pm$0.6 & 3.3$\pm$0.9 \\
& \bf Important & \underline{3.5$\pm$0.9} & 4.2$\pm$0.8 & \bf 4.8$\pm$0.2 & 4.1$\pm$0.1 & 4.5$\pm$0.3 & 4.6$\pm$0.4 & 4.0$\pm$0.7 \\
\bottomrule
\end{tabular}
}
\end{center}
}
\headerbox{8. Results of Emotional Abilities}{name=r4,span=2,column=1,below=r3,above=bottom}{
\begin{center}
\resizebox{\linewidth}{!}{
\begin{tabular}{ll cccccc|cc}
\toprule
\multirow{2}{*}{} &
\multirow{2}{*}{\bf Subscales} &
\multirow{2}{*}{\bf llama2-7b} &
\multirow{2}{*}{\bf llama2-13b} &
\multirow{2}{*}{\bf text-davinci-003} &
\multirow{2}{*}{\bf gpt-3.5-turbo} &
\multirow{2}{*}{\bf gpt-4} &
\multirow{2}{*}{\bf gpt-4-jb} &
\multicolumn{2}{c}{\bf Crowd} \\
& & & & & & & & \it Male & \it Female \\
\midrule
\it EIS & \bf Overall & 131.6$\pm$6.0 & 128.6$\pm$12.3 & 148.4$\pm$9.4 & 132.9$\pm$2.2 & \bf 151.4$\pm$18.7 & \underline{121.8$\pm$12.0} & 124.8$\pm$16.5 & 130.9$\pm$15.1 \\
\midrule
\multirow{4}{*}{\it WLEIS} &
\bf SEA & \underline{4.7$\pm$1.3} & 5.5$\pm$1.3 & 5.9$\pm$0.6 & 6.0$\pm$0.1 & 6.2$\pm$0.7 & \bf 6.4$\pm$0.4 & \multicolumn{2}{c}{4.0$\pm$1.1} \\
& \bf OEA & \underline{4.9$\pm$0.8} & 5.3$\pm$1.1 & 5.2$\pm$0.2 & 5.8$\pm$0.3 & 5.2$\pm$0.6 & \bf 5.9$\pm$0.4 & \multicolumn{2}{c}{3.8$\pm$1.1} \\
& \bf UOE & \underline{5.7$\pm$0.6} & 5.9$\pm$0.7 & 6.1$\pm$0.4 & 6.0$\pm$0.0 & \bf 6.5$\pm$0.5 & 6.3$\pm$0.4 & \multicolumn{2}{c}{4.1$\pm$0.9} \\
& \bf ROE & \underline{4.5$\pm$0.8} & 5.2$\pm$1.2 & 5.8$\pm$0.5 & \bf 6.0$\pm$0.0 & 5.2$\pm$0.7 & 5.3$\pm$0.5 & \multicolumn{2}{c}{4.2$\pm$1.0} \\
\midrule
\it Empathy & \bf Overall & 5.8$\pm$0.8 & 5.9$\pm$0.5 & 6.0$\pm$0.4 & 6.2$\pm$0.3 & \bf 6.8$\pm$0.4 & \underline{4.6$\pm$0.2} & \multicolumn{2}{c}{4.9$\pm$0.8} \\
\bottomrule
\end{tabular}
}
\end{center}
}
\headerbox{9. Conclusions}{name=conclusion,span=1,column=0,below=model}{
Here are some \textbf{\color{bordercolor}key takeaways}:
\begin{boenumerate}\compresslist
\item Distinct personality traits
\item More negative traits (DTDD)
\item Jailbreak influences results
\item Biased towards Masculinity
\item Similar vocational preference
\item More self-motivated \& self-confident
\item A higher emotional quotient
\end{boenumerate}
}
\headerbox{10. References}{name=references,column=0,span=1,below=conclusion,above=bottom}{
% \small % Reduce the font size in this block
\renewcommand{\section}[2]{\vskip 0.05em} % Get rid of the default "References" section title
% \nocite{*} % Insert publications even if they are not cited in the poster
\bibliographystyle{unsrt}
\bibliography{poster}
}
\end{poster}
\end{document}