@inproceedings{park-etal-2026-star,
title = "{STAR}: Detecting Inference-time Backdoors in {LLM} Reasoning via Prior-Posterior Amplification",
author = "Park, Seong-Gyu and
Park, Sohee and
Lee, Jisu and
Na, Hyunsik and
Choi, Daeseon",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2010/",
pages = "42780--42797",
ISBN = "979-8-89176-462-0",
abstract = "Recent LLMs increasingly rely on explicit reasoning mechanisms such as Chain-of-Thought, but these reasoning traces also create a new attack surface for inference-time backdoors that steer the model toward malicious reasoning paths without modifying its parameters. Such attacks can remain linguistically coherent, making them difficult to detect with surface-, consistency-, or confidence-based signals. We propose $\textbf{STAR}$, a lightweight detection framework that compares input-conditioned token probabilities with prior-only reasoning probabilities. STAR measures token-level prior-posterior amplification and applies CUSUM to identify persistent anomalies in the generated reasoning process. Experiments across four LLMs (8B-70B), five reasoning benchmarks, and three inference-time backdoor attacks show that STAR achieves strong detection performance, with an average AUROC of 0.980, and substantially lower TFLOPs than high-overhead baselines such as Onion. STAR also remains effective under top-k logit access and the evaluated burn-in-aware adaptive attacks, while exposing a trade-off between evasion attempts and attack success."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="park-etal-2026-star">
<titleInfo>
<title>STAR: Detecting Inference-time Backdoors in LLM Reasoning via Prior-Posterior Amplification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Seong-Gyu</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sohee</namePart>
<namePart type="family">Park</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jisu</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hyunsik</namePart>
<namePart type="family">Na</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daeseon</namePart>
<namePart type="family">Choi</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Recent LLMs increasingly rely on explicit reasoning mechanisms such as Chain-of-Thought, but these reasoning traces also create a new attack surface for inference-time backdoors that steer the model toward malicious reasoning paths without modifying its parameters. Such attacks can remain linguistically coherent, making them difficult to detect with surface-, consistency-, or confidence-based signals. We propose STAR, a lightweight detection framework that compares input-conditioned token probabilities with prior-only reasoning probabilities. STAR measures token-level prior-posterior amplification and applies CUSUM to identify persistent anomalies in the generated reasoning process. Experiments across four LLMs (8B-70B), five reasoning benchmarks, and three inference-time backdoor attacks show that STAR achieves strong detection performance, with an average AUROC of 0.980, and substantially lower TFLOPs than high-overhead baselines such as Onion. STAR also remains effective under top-k logit access and the evaluated burn-in-aware adaptive attacks, while exposing a trade-off between evasion attempts and attack success.</abstract>
<identifier type="citekey">park-etal-2026-star</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2010/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42780</start>
<end>42797</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T STAR: Detecting Inference-time Backdoors in LLM Reasoning via Prior-Posterior Amplification
%A Park, Seong-Gyu
%A Park, Sohee
%A Lee, Jisu
%A Na, Hyunsik
%A Choi, Daeseon
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F park-etal-2026-star
%X Recent LLMs increasingly rely on explicit reasoning mechanisms such as Chain-of-Thought, but these reasoning traces also create a new attack surface for inference-time backdoors that steer the model toward malicious reasoning paths without modifying its parameters. Such attacks can remain linguistically coherent, making them difficult to detect with surface-, consistency-, or confidence-based signals. We propose STAR, a lightweight detection framework that compares input-conditioned token probabilities with prior-only reasoning probabilities. STAR measures token-level prior-posterior amplification and applies CUSUM to identify persistent anomalies in the generated reasoning process. Experiments across four LLMs (8B-70B), five reasoning benchmarks, and three inference-time backdoor attacks show that STAR achieves strong detection performance, with an average AUROC of 0.980, and substantially lower TFLOPs than high-overhead baselines such as Onion. STAR also remains effective under top-k logit access and the evaluated burn-in-aware adaptive attacks, while exposing a trade-off between evasion attempts and attack success.
%U https://aclanthology.org/2026.findings-emnlp.2010/
%P 42780-42797
Markdown (Informal)
[STAR: Detecting Inference-time Backdoors in LLM Reasoning via Prior-Posterior Amplification](https://aclanthology.org/2026.findings-emnlp.2010/) (Park et al., Findings 2026)
ACL