@inproceedings{wang-etal-2026-one-trace,
title = "One Trace Is Not Enough: Paired Probing for Runtime Backdoor Detection in Limited-Black-Box {LLM}s",
author = "Wang, Yansu and
Chen, Feiyan and
Zhang, Rui and
Huang, Kaizhu",
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.2111/",
pages = "48027--48043",
ISBN = "979-8-89176-459-0",
abstract = "Backdoored generative LLMs can appear benign on normal inputs yet switch to unsafe or target-specific behavior when hidden triggers are activated, making request-level runtime detection a useful complement to model repair and offline auditing. Existing limited-black-box detectors typically rely on a single decoding trace, while text-only probing methods reveal susceptibility but often depend on specific probes or model-level audit goals. We introduce ProbeShift, a limited-black-box paired-probing framework that treats detection as repeated measurement of the same request under controlled prompt variations. It compares a baseline generation with templated probing generations and aggregates confidence dynamics, task-conditioned behavior cues, and paired response shifts. On safety-oriented BackdoorLLM DPA settings with LLaMA-2-7B-Chat, ProbeShift achieves remarkable performance: whilst maintaining high average recall rate of 97{\%}, it reduced the FPR by 81{\%} on average compared to the SOTA method ConfGuard based on single-track detection. These results show that paired probing yields a more stable and practical runtime signal than single-trace monitoring alone."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-one-trace">
<titleInfo>
<title>One Trace Is Not Enough: Paired Probing for Runtime Backdoor Detection in Limited-Black-Box LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yansu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feiyan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rui</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kaizhu</namePart>
<namePart type="family">Huang</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>Backdoored generative LLMs can appear benign on normal inputs yet switch to unsafe or target-specific behavior when hidden triggers are activated, making request-level runtime detection a useful complement to model repair and offline auditing. Existing limited-black-box detectors typically rely on a single decoding trace, while text-only probing methods reveal susceptibility but often depend on specific probes or model-level audit goals. We introduce ProbeShift, a limited-black-box paired-probing framework that treats detection as repeated measurement of the same request under controlled prompt variations. It compares a baseline generation with templated probing generations and aggregates confidence dynamics, task-conditioned behavior cues, and paired response shifts. On safety-oriented BackdoorLLM DPA settings with LLaMA-2-7B-Chat, ProbeShift achieves remarkable performance: whilst maintaining high average recall rate of 97%, it reduced the FPR by 81% on average compared to the SOTA method ConfGuard based on single-track detection. These results show that paired probing yields a more stable and practical runtime signal than single-trace monitoring alone.</abstract>
<identifier type="citekey">wang-etal-2026-one-trace</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2111/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>48027</start>
<end>48043</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T One Trace Is Not Enough: Paired Probing for Runtime Backdoor Detection in Limited-Black-Box LLMs
%A Wang, Yansu
%A Chen, Feiyan
%A Zhang, Rui
%A Huang, Kaizhu
%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 wang-etal-2026-one-trace
%X Backdoored generative LLMs can appear benign on normal inputs yet switch to unsafe or target-specific behavior when hidden triggers are activated, making request-level runtime detection a useful complement to model repair and offline auditing. Existing limited-black-box detectors typically rely on a single decoding trace, while text-only probing methods reveal susceptibility but often depend on specific probes or model-level audit goals. We introduce ProbeShift, a limited-black-box paired-probing framework that treats detection as repeated measurement of the same request under controlled prompt variations. It compares a baseline generation with templated probing generations and aggregates confidence dynamics, task-conditioned behavior cues, and paired response shifts. On safety-oriented BackdoorLLM DPA settings with LLaMA-2-7B-Chat, ProbeShift achieves remarkable performance: whilst maintaining high average recall rate of 97%, it reduced the FPR by 81% on average compared to the SOTA method ConfGuard based on single-track detection. These results show that paired probing yields a more stable and practical runtime signal than single-trace monitoring alone.
%U https://aclanthology.org/2026.emnlp-main.2111/
%P 48027-48043
Markdown (Informal)
[One Trace Is Not Enough: Paired Probing for Runtime Backdoor Detection in Limited-Black-Box LLMs](https://aclanthology.org/2026.emnlp-main.2111/) (Wang et al., EMNLP 2026)
ACL