GwendalTsang commited on
Commit
59cd63b
·
verified ·
1 Parent(s): 2194765

Add audited paper table sources

Browse files
paper-source/tables/firstsentence.tex ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \begin{table}[t]
2
+ \centering
3
+ \caption{Hallucination detection performance (AUROC) across five QA datasets for LLaMA-3.1-8B-Instruct and Mistral-7B-Instruct. For each method, we apply first-sentence truncation for the generated answers. Forward horizon $w$ is set to 7.}
4
+ \label{tab:first_sentence_truncation}
5
+ \resizebox{\linewidth}{!}{%
6
+ \begin{tabular}{cccccccccccccc}
7
+ \toprule
8
+ \multicolumn{2}{c}{} & \multicolumn{6}{c}{LlaMA-3.1-8B-Instruct} & \multicolumn{6}{c}{Mistral-7B-Instruct-v0.3} \\
9
+ \multicolumn{2}{c}{} & CoQA & SQuAD & HotpotQA & TriviaQA & PsiLoQA & Avg & CoQA & SQuAD & HotpotQA & TriviaQA & PsiLoQA & Avg \\ \midrule
10
+ \multicolumn{2}{c}{Pred. Entropy} & 0.7300 & 0.7850 & 0.7264 & 0.7897 & 0.7880 & 0.7638 & 0.7396 & 0.7633 & 0.7449 & 0.7734 & 0.7574 & 0.7557 \\
11
+ \multicolumn{2}{c}{LN-Pred. Entropy} & 0.6895 & 0.7425 & 0.7427 & 0.7962 & 0.5742 & 0.7090 & 0.6532 & 0.6564 & 0.6970 & 0.7305 & 0.5045 & 0.6483 \\
12
+ \multicolumn{2}{c}{Semantic Entropy} & 0.7015 & 0.7806 & 0.7108 & 0.8148 & 0.7392 & 0.7494 & 0.5772 & 0.6314 & 0.6518 & 0.7603 & 0.6685 & 0.6578 \\
13
+ \multicolumn{2}{c}{Lexical Similarity} & 0.7902 & 0.8433 & 0.7924 & 0.8614 & 0.8061 & 0.8187 & 0.7081 & 0.7270 & 0.7534 & 0.8171 & 0.6907 & 0.7393 \\
14
+ \multicolumn{2}{c}{LID} & 0.5535 & 0.5731 & 0.5714 & 0.5550 & 0.6648 & 0.5836 & 0.5286 & 0.5784 & 0.5590 & 0.4902 & 0.6600 & 0.5632 \\
15
+ \multicolumn{2}{c}{EigenScore} & 0.7362 & 0.8152 & 0.7271 & 0.8089 & 0.7941 & 0.7763 & 0.7087 & 0.7350 & 0.7357 & 0.7707 & 0.7517 & 0.7404 \\ \midrule
16
+ \multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Hidden-State\\ Probing\end{tabular}} & RankME & 0.8159 & 0.8106 & \cellcolor{rankthree}{0.8013} & 0.8228 & 0.8549 & 0.8211 & 0.7746 & 0.7548 & 0.6931 & 0.6763 & 0.8617 & 0.7521 \\
17
+ & Curvature & \cellcolor{rankthree}{0.8557} & \cellcolor{rankthree}{0.8625} & \cellcolor{ranktwo}{ 0.8119} & \cellcolor{rankthree}{0.8287} & \cellcolor{rankthree}{0.9160} & \cellcolor{rankthree}{0.8550} & \cellcolor{ranktwo}{ 0.8474} & \cellcolor{rankthree}{0.8704} & \cellcolor{ranktwo}{ 0.8088} & \cellcolor{ranktwo}{ 0.8680} & \cellcolor{ranktwo}{ 0.9264} & \cellcolor{rankthree}{0.8642} \\
18
+ & Val Loss & \cellcolor{rankone}\textbf{0.8621} & \cellcolor{ranktwo}{ 0.8865} & \cellcolor{rankone}\textbf{0.8287} & \cellcolor{ranktwo}{ 0.8731} & \cellcolor{rankone}\textbf{0.9238} & \cellcolor{ranktwo}{ 0.8748} & \cellcolor{rankthree}{0.8430} & \cellcolor{rankone}\textbf{0.8784} & \cellcolor{rankthree}{0.8087} & \cellcolor{rankone}\textbf{0.8805} & \cellcolor{rankone}\textbf{0.9216} & \cellcolor{ranktwo}{ 0.8664} \\
19
+ & RGN & 0.7167 & 0.8104 & 0.7292 & 0.7772 & 0.8112 & 0.7689 & 0.8346 & 0.7929 & 0.7631 & \cellcolor{rankthree}{0.8176} & 0.8797 & 0.8176 \\
20
+ & SNR & 0.7602 & 0.6693 & 0.7723 & 0.6408 & 0.8001 & 0.7285 & 0.7453 & 0.7112 & 0.7259 & 0.7115 & 0.8274 & 0.7443 \\
21
+ &ID & \cellcolor{ranktwo}{ 0.8581} & \cellcolor{rankone}\textbf{0.8900} & 0.7723 & 0.8066 & \cellcolor{ranktwo}{ 0.9208} & 0.8496 & \cellcolor{ranktwo}{ 0.8474} & \cellcolor{ranktwo}{ 0.8769} & 0.7631 & \cellcolor{rankthree}{0.8176} & 0.8942 & 0.8398 \\
22
+ & FEPoID & \cellcolor{ranktwo}{ 0.8581} & \cellcolor{rankone}\textbf{0.8900} & \cellcolor{rankone}\textbf{0.8287} & \cellcolor{rankone}{\textbf{0.8782}} & \cellcolor{ranktwo}{ 0.9208} & \cellcolor{rankone}\textbf{0.8752} & \cellcolor{rankone}\textbf{0.8501} & \cellcolor{ranktwo}{ 0.8769} & \cellcolor{rankone}\textbf{0.8153} & \cellcolor{rankone}\textbf{0.8805} & \cellcolor{rankthree}{0.9207} & \cellcolor{rankone}\textbf{0.8687} \\ \midrule
23
+ \end{tabular}}%
24
+ \end{table}
25
+
paper-source/tables/notruncation.tex ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Table generated by Excel2LaTeX from sheet 'no truncation'
2
+ \begin{table*}[t]
3
+ \centering
4
+ \caption{Hallucination detection performance (AUROC) across QA datasets with $w=7$. For representation-based methods, we extract representations at the last generated token. Top-3 results are highlighted, with darker color indicating better performance.
5
+ }
6
+ \label{tab:no_truncation}
7
+ \resizebox{\textwidth}{!}{%
8
+ \begin{tabular}{cccccccccccccc}
9
+ \toprule
10
+ \multicolumn{2}{c}{} & \multicolumn{6}{c}{LlaMA-3.1-8B-Instruct} & \multicolumn{6}{c}{Mistral-7B-Instruct-v0.3} \\
11
+ \multicolumn{2}{c}{} & CoQA & SQuAD & HotpotQA & TriviaQA & PsiloQA & Avg & CoQA & SQuAD & HotpotQA & TriviaQA & PsiloQA & Avg \\ \midrule
12
+ \multicolumn{2}{c}{Pred. Entropy} & 0.5833 & 0.5703 & 0.7103 & 0.6859 & 0.3604 & 0.5820 & 0.7200 & 0.7316 & 0.6303 & 0.6763 & 0.6110 & 0.6738 \\
13
+ \multicolumn{2}{c}{LN-Pred. Entropy} & 0.5781 & 0.5671 & 0.7087 & 0.6774 & 0.3555 & 0.5774 & 0.6528 & 0.6390 & 0.6589 & 0.6828 & 0.4418 & 0.6151 \\
14
+ \multicolumn{2}{c}{Semantic Entropy} & 0.5003 & 0.5518 & 0.4454 & 0.5505 & 0.6076 & 0.5311 & 0.5769 & 0.6409 & 0.6418 & 0.7353 & 0.6853 & 0.6560 \\
15
+ \multicolumn{2}{c}{Lexical Similarity} & 0.6780 & 0.5988 & 0.7294 & 0.6838 & 0.4082 & 0.6196 & 0.7071 & 0.7169 & 0.6946 & 0.7547 & 0.5781 & 0.6903 \\
16
+ \multicolumn{2}{c}{LID} & 0.5059 & 0.5281 & 0.5171 & 0.4989 & 0.5994 & 0.5299 & 0.5274 & 0.5688 & 0.5518 & 0.4970 & 0.6447 & 0.5579 \\
17
+ \multicolumn{2}{c}{EigenScore} & 0.5247 & 0.5300 & 0.5987 & 0.5882 & 0.5080 & 0.5499 & 0.7092 & 0.7508 & 0.6587 & 0.7273 & 0.7246 & 0.7141 \\ \midrule
18
+ \multirow{7}{*}{\begin{tabular}[c]{@{}c@{}}Hidden-State\\ Probing \end{tabular}} & RankME & \cellcolor{ranktwo}{ 0.6598} & 0.6071 & 0.7040 & 0.7114 & 0.7478 & 0.6860 & 0.7743 & 0.7341 & 0.6909 & 0.6699 & 0.8277 & 0.7394 \\
19
+ & Curvature & \cellcolor{rankthree}{0.6323} & \cellcolor{ranktwo}{ 0.6183} & \cellcolor{rankthree}{0.7413} & 0.7366 & \cellcolor{rankthree}{0.7565} & \cellcolor{rankthree}{0.6970} & \cellcolor{rankone}\textbf{0.8492} & \cellcolor{ranktwo}{0.8553} & \cellcolor{rankthree}{0.7940} & \cellcolor{ranktwo}{ 0.8368} & \cellcolor{ranktwo}{ 0.9005} & \cellcolor{ranktwo}{ 0.8472} \\
20
+ & Val Loss & \cellcolor{rankone}\textbf{0.6705} & \cellcolor{rankthree}{0.6164} & \cellcolor{ranktwo}{ 0.7682} & \cellcolor{rankone}\textbf{0.7861} & \cellcolor{ranktwo}{ 0.7836} & \cellcolor{ranktwo}{ 0.7250} & \cellcolor{rankthree}{0.8283} & \cellcolor{rankone}\textbf{0.8679} & \cellcolor{ranktwo}{ 0.7968} & \cellcolor{rankone}\textbf{0.8496} & \cellcolor{rankthree}{0.8861} & \cellcolor{rankthree}{0.8457} \\
21
+ & RGN & 0.5993 & 0.6130 & 0.7040 & \cellcolor{ranktwo}{ 0.7859} & 0.7373 & 0.6879 & 0.7090 & \cellcolor{rankthree}{0.7553} & 0.7493 & \cellcolor{rankthree}{0.7868} & 0.8562 & 0.7713 \\
22
+ & SNR & 0.5240 & 0.5699 & 0.7009 & 0.5567 & 0.6624 & 0.6028 & 0.7475 & 0.6870 & 0.7620 & 0.7008 & 0.8207 & 0.7436 \\
23
+ & ID & \cellcolor{rankone}\textbf{0.6705} & 0.6130 & 0.6932 & 0.7073 & 0.7373 & 0.6843 & \cellcolor{rankthree}{0.8283} & \cellcolor{rankone}\textbf{0.8679} & 0.6993 & \cellcolor{rankthree}{0.7868} & 0.8533 & 0.8071 \\
24
+ & FEPoID& \cellcolor{rankone}\textbf{0.6705} & \cellcolor{rankone}\textbf{0.6377} & \cellcolor{rankone}\textbf{0.7807} & \cellcolor{rankthree}{0.7516} & \cellcolor{rankone}\textbf{0.7862} & \cellcolor{rankone}\textbf{0.7253} & \cellcolor{ranktwo}{ 0.8466} & \cellcolor{rankone}\textbf{0.8679} & \cellcolor{rankone}\textbf{0.7982} & \cellcolor{rankone}\textbf{0.8496} & \cellcolor{rankone}\textbf{0.9031} & \cellcolor{rankone}\textbf{0.8531} \\ \bottomrule
25
+ \end{tabular}}%
26
+
27
+ \end{table*}