File size: 2,972 Bytes
b650828
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
\begin{table}[t]
    \centering
    \caption{The detailed evaluation result in terms of Normalized Dispersion Score (NDS) for all multimodal judges on \textbf{bias} perspective. The feedback is provided in numerical scale with range [0, 10]. Specifically, we separately report the bias w.r.t. different demographic identifications, i.e. age, gender, race, nationality, and religion. The best performance across all models is bolded.}
    \resizebox{1.0\linewidth}{!}{%
    \begin{tabular}{c|cccccc}
    \toprule
         % & \multicolumn{6}{c}{\bf Occupation} & \multicolumn{4}{c}{\bf Education} \\
         & Age & Gender & Race & Nationality & Religion & \cellcolor{skyblue}Avg \\
         \midrule
         CLIP-v1$^\diamondsuit$ & 73.6 & 75.2 & 73.1 & 79.1 & 78.4 & \cellcolor{skyblue} 75.2  \\
         BLIP-v2$^\diamondsuit$ & 85.3 & 83.6 & 82.7 & 81.8 & 87.5 & \cellcolor{skyblue} 83.6 \\
         PickScore-v1$^\diamondsuit$ & 65.3 & 66.7 & 66.4 & 67.3 & 69.4 & \cellcolor{skyblue} 66.7  \\
         HPS-v2.1$^\diamondsuit$ & 75.8 & 78.2 & 79.5 & 78.6 & 79.3 & \cellcolor{skyblue} 78.2  \\
         ImageReward$^\diamondsuit$ & 73.9 & 73.2 & 70.9 & 73.0 & 80.2 & \cellcolor{skyblue} 73.2 \\
         Aesthetics$^\diamondsuit$ & \bf 85.3 & \bf 85.9 & \bf 86.3 & \bf 85.8 & 86.2 & \cellcolor{skyblue} \bf 85.9  \\
         \midrule
         LLaVA-1.5-7b$^\heartsuit$ & 67.6 & 71.4 & 75.8 & 68.4 & 77.3 & \cellcolor{skyblue} 71.4 \\
         LLaVA-1.5-13b$^\heartsuit$ & 71.9 & 74.8 & 76.6 & 74.0 & 80.6 & \cellcolor{skyblue} 74.8 \\
         LLaVA-NeXT-mistral-7b$^\heartsuit$ & 68.4 & 64.6 & 62.4 & 59.7 & 78.1 & \cellcolor{skyblue} 64.6 \\
         LLaVA-NeXT-vicuna-7b$^\heartsuit$ & 63.2 & 64.1 & 62.5 & 63.8 & 74.2 & \cellcolor{skyblue} 64.1\\
         Instructblip-7b$^\heartsuit$ & 80.8 & 80.6 & 80.3 & 79.0 & 85.4 & \cellcolor{skyblue} 80.6 \\
         MiniGPT4-v2$^\heartsuit$ & 68.1 & 67.2 & 66.2 & 67.0 & 69.3 & \cellcolor{skyblue} 67.2 \\
         Prometheus-Vision-7b$^\heartsuit$ & 47.2 & 42.5 & 37.8 & 40.0 & 54.2 & \cellcolor{skyblue} 42.5 \\
         Prometheus-Vision-13b$^\heartsuit$ & 54.2 & 44.7 & 36.0 & 39.3 & 65.7 & \cellcolor{skyblue} 44.7 \\
         Qwen-VL-Chat$^\spadesuit$ & 62.4 & 62.3 & 62.3 & 63.1 & 58.9 & \cellcolor{skyblue} 62.3 \\
         Internvl-chat-v1-5$^\spadesuit$ & 74.0 & 74.1 & 73.6 & 73.9 & 76.6 & \cellcolor{skyblue} 74.1 \\
         Idefics2-8b$^\spadesuit$ & 55.1 & 59.2 & 61.7 & 62.8 & 51.0 & \cellcolor{skyblue} 59.2 \\
         \midrule
         GPT-4-vision$^\clubsuit$ & 81.2 & 80.2 & 77.6 & 79.9 & \bf 88.2 & \cellcolor{skyblue} 80.2 \\
         GPT-4o$^\clubsuit$ & 81.2 & 82.7 & 82.8 & 83.2 & 86.1 & \cellcolor{skyblue} 82.7 \\
         Gemini Ultra$^\clubsuit$ & 72.6 & 75.8 & 78.4 & 77.0 & 72.3 & \cellcolor{skyblue} 75.8 \\
         Claude 3 Opus$^\clubsuit$ & 63.3 & 66.1 & 67.5 & 66.9 & 66.8 & \cellcolor{skyblue} 66.1 \\
    \bottomrule
    \end{tabular}%
    }
    \label{exp:bias_nds}
\end{table}