@inproceedings{fan-etal-2026-prompt,
title = "Prompt-Induced Recovery Behavior Decomposition for Black-Box Pre-training Data Detection",
author = "Fan, Xin and
Ryoto, Miyamoto and
Mo, Fan and
Chen, Chongxian and
Matsumoto, Tsuneo and
Kido, Fuyuko and
Yamana, Hayato",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2527/",
pages = "57471--57490",
ISBN = "979-8-89176-459-0",
abstract = "Large language models are pretrained on massive, largely undisclosed corpora that may contain copyrighted, privacy-sensitive, or otherwise unauthorized content, making pre-training data detection essential for data-rights auditing, privacy protection, and benchmark contamination analysis. However, most existing methods follow a gray-box setting, relying on internal signals such as token probabilities, logits, or losses, which are unavailable for commercial LLMs that expose only the final generated text. Recent black-box methods instead use word-recovery rate as the detection signal, but overlook that a word{'}s discriminative value can vary substantially across target models and target texts. To address this limitation, we propose ***PIRBD-PDD***, a target LLM- and text-adaptive framework for black-box pre-training data detection via *Prompt-Induced Recovery Behavior Decomposition*. ***PIRBD-PDD*** decomposes the target LLM{'}s recovery behavior into memorization-oriented and context-oriented modes through prompt induction, contrasts their recovery outcomes to identify high-discriminability words, and computes the final detection score from their recovery rate under the original, non-induced prompt. Experiments across diverse LLMs and data domains show that ***PIRBD-PDD*** consistently outperforms state-of-the-art black-box baselines. Further analyses provide empirical evidence for the proposed decomposition and show that high-discriminability words provide more robust and informative signals for pre-training data detection."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fan-etal-2026-prompt">
<titleInfo>
<title>Prompt-Induced Recovery Behavior Decomposition for Black-Box Pre-training Data Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Miyamoto</namePart>
<namePart type="family">Ryoto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fan</namePart>
<namePart type="family">Mo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chongxian</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tsuneo</namePart>
<namePart type="family">Matsumoto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fuyuko</namePart>
<namePart type="family">Kido</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hayato</namePart>
<namePart type="family">Yamana</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Large language models are pretrained on massive, largely undisclosed corpora that may contain copyrighted, privacy-sensitive, or otherwise unauthorized content, making pre-training data detection essential for data-rights auditing, privacy protection, and benchmark contamination analysis. However, most existing methods follow a gray-box setting, relying on internal signals such as token probabilities, logits, or losses, which are unavailable for commercial LLMs that expose only the final generated text. Recent black-box methods instead use word-recovery rate as the detection signal, but overlook that a word’s discriminative value can vary substantially across target models and target texts. To address this limitation, we propose ***PIRBD-PDD***, a target LLM- and text-adaptive framework for black-box pre-training data detection via *Prompt-Induced Recovery Behavior Decomposition*. ***PIRBD-PDD*** decomposes the target LLM’s recovery behavior into memorization-oriented and context-oriented modes through prompt induction, contrasts their recovery outcomes to identify high-discriminability words, and computes the final detection score from their recovery rate under the original, non-induced prompt. Experiments across diverse LLMs and data domains show that ***PIRBD-PDD*** consistently outperforms state-of-the-art black-box baselines. Further analyses provide empirical evidence for the proposed decomposition and show that high-discriminability words provide more robust and informative signals for pre-training data detection.</abstract>
<identifier type="citekey">fan-etal-2026-prompt</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2527/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>57471</start>
<end>57490</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Prompt-Induced Recovery Behavior Decomposition for Black-Box Pre-training Data Detection
%A Fan, Xin
%A Ryoto, Miyamoto
%A Mo, Fan
%A Chen, Chongxian
%A Matsumoto, Tsuneo
%A Kido, Fuyuko
%A Yamana, Hayato
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F fan-etal-2026-prompt
%X Large language models are pretrained on massive, largely undisclosed corpora that may contain copyrighted, privacy-sensitive, or otherwise unauthorized content, making pre-training data detection essential for data-rights auditing, privacy protection, and benchmark contamination analysis. However, most existing methods follow a gray-box setting, relying on internal signals such as token probabilities, logits, or losses, which are unavailable for commercial LLMs that expose only the final generated text. Recent black-box methods instead use word-recovery rate as the detection signal, but overlook that a word’s discriminative value can vary substantially across target models and target texts. To address this limitation, we propose ***PIRBD-PDD***, a target LLM- and text-adaptive framework for black-box pre-training data detection via *Prompt-Induced Recovery Behavior Decomposition*. ***PIRBD-PDD*** decomposes the target LLM’s recovery behavior into memorization-oriented and context-oriented modes through prompt induction, contrasts their recovery outcomes to identify high-discriminability words, and computes the final detection score from their recovery rate under the original, non-induced prompt. Experiments across diverse LLMs and data domains show that ***PIRBD-PDD*** consistently outperforms state-of-the-art black-box baselines. Further analyses provide empirical evidence for the proposed decomposition and show that high-discriminability words provide more robust and informative signals for pre-training data detection.
%U https://aclanthology.org/2026.emnlp-main.2527/
%P 57471-57490
Markdown (Informal)
[Prompt-Induced Recovery Behavior Decomposition for Black-Box Pre-training Data Detection](https://aclanthology.org/2026.emnlp-main.2527/) (Fan et al., EMNLP 2026)
ACL