@inproceedings{xia-aykent-2026-three,
title = "Three Numbers and a Precondition: Auditing Multi-Source {LLM} Benchmarks for Source-Identity Leakage",
author = "Xia, Tian and
Aykent, Sarp",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.114/",
pages = "2484--2502",
ISBN = "979-8-89176-439-2",
abstract = "LLM benchmarks often pool rows from several sources (judges, retrieval configurations, training frameworks) and report one headline score. When label rates differ by source, a detector can earn that score by recognizing a row{'}s source, not its quality, and such a score is no promise about a new source. Our audit reports three AUCs of one detector: given features plus source ($\mathrm{AUC}_\text{full}$, the headline), features alone, and source alone. Their ratio $L=\mathrm{AUC}_\text{source}/\mathrm{AUC}_\text{full}$ shows how close the source alone comes to the headline. The precondition: $L$ is read only above a default floor, $\mathrm{AUC}_\text{full}\geq 0.65$, because two near-chance AUCs give a ratio near 1 although nothing leaks. On four RAGBench configurations (adherence rates 40.2{\%} to 94.2{\%}) the headline is 0.843 and the configuration label alone reaches 0.811 ($L=0.96$), above every per-query feature. When five LLM judges grade the same hard-math answers, $L=0.84$ looks similar, but the headline (0.594) fails the floor: judge identity predicts nothing. The floor, not $L$, separates the two panels. Six further audits with the floor and a 0.90 band on $L$ frozen add the other outcomes, including a floor-passing panel whose source carries nothing."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xia-aykent-2026-three">
<titleInfo>
<title>Three Numbers and a Precondition: Auditing Multi-Source LLM Benchmarks for Source-Identity Leakage</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tian</namePart>
<namePart type="family">Xia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sarp</namePart>
<namePart type="family">Aykent</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>LLM benchmarks often pool rows from several sources (judges, retrieval configurations, training frameworks) and report one headline score. When label rates differ by source, a detector can earn that score by recognizing a row’s source, not its quality, and such a score is no promise about a new source. Our audit reports three AUCs of one detector: given features plus source (AUC_\textfull, the headline), features alone, and source alone. Their ratio L=AUC_\textsource/AUC_\textfull shows how close the source alone comes to the headline. The precondition: L is read only above a default floor, AUC_\textfull\geq 0.65, because two near-chance AUCs give a ratio near 1 although nothing leaks. On four RAGBench configurations (adherence rates 40.2% to 94.2%) the headline is 0.843 and the configuration label alone reaches 0.811 (L=0.96), above every per-query feature. When five LLM judges grade the same hard-math answers, L=0.84 looks similar, but the headline (0.594) fails the floor: judge identity predicts nothing. The floor, not L, separates the two panels. Six further audits with the floor and a 0.90 band on L frozen add the other outcomes, including a floor-passing panel whose source carries nothing.</abstract>
<identifier type="citekey">xia-aykent-2026-three</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.114/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2484</start>
<end>2502</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Three Numbers and a Precondition: Auditing Multi-Source LLM Benchmarks for Source-Identity Leakage
%A Xia, Tian
%A Aykent, Sarp
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F xia-aykent-2026-three
%X LLM benchmarks often pool rows from several sources (judges, retrieval configurations, training frameworks) and report one headline score. When label rates differ by source, a detector can earn that score by recognizing a row’s source, not its quality, and such a score is no promise about a new source. Our audit reports three AUCs of one detector: given features plus source (AUC_\textfull, the headline), features alone, and source alone. Their ratio L=AUC_\textsource/AUC_\textfull shows how close the source alone comes to the headline. The precondition: L is read only above a default floor, AUC_\textfull\geq 0.65, because two near-chance AUCs give a ratio near 1 although nothing leaks. On four RAGBench configurations (adherence rates 40.2% to 94.2%) the headline is 0.843 and the configuration label alone reaches 0.811 (L=0.96), above every per-query feature. When five LLM judges grade the same hard-math answers, L=0.84 looks similar, but the headline (0.594) fails the floor: judge identity predicts nothing. The floor, not L, separates the two panels. Six further audits with the floor and a 0.90 band on L frozen add the other outcomes, including a floor-passing panel whose source carries nothing.
%U https://aclanthology.org/2026.aacl-long.114/
%P 2484-2502
Markdown (Informal)
[Three Numbers and a Precondition: Auditing Multi-Source LLM Benchmarks for Source-Identity Leakage](https://aclanthology.org/2026.aacl-long.114/) (Xia & Aykent, AACL-IJCNLP 2026)
ACL