@inproceedings{ma-wang-2026-tcd,
title = "{TCD}: Empowering Zero-Shot Detection of {LLM}-Generated Text using Token Contextual Dependency",
author = "Ma, Shihan and
Wang, Quan",
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.1708/",
pages = "38911--38926",
ISBN = "979-8-89176-459-0",
abstract = "Zero-shot detectors for LLM-generated text aggregate token-level statistics uniformly, implicitly assuming that every position contributes equally to detection. We challenge this assumption and show that detection signal is highly uneven across token positions: positions of high contextual dependency, where the surrogate model relies on concentrated context, carry far more discriminative power than others. Using attention entropy to quantify this dependency, we empirically validate that aggregating only low-entropy tokens yields clearly separated human and LLM score distributions, whereas high-entropy tokens do not. Based on this finding, we propose Token Contextual Dependency (TCD), a plug-and-play framework that reweights token contributions in existing zero-shot detectors via attention entropy-based weights computed from dynamically selected layers. Experiments on state-of-the-art detectors across multiple benchmarks demonstrate consistent and meaningful improvements in detection performance. Our code is available at: https://github.com/RecursionM/TCD."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ma-wang-2026-tcd">
<titleInfo>
<title>TCD: Empowering Zero-Shot Detection of LLM-Generated Text using Token Contextual Dependency</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shihan</namePart>
<namePart type="family">Ma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Quan</namePart>
<namePart type="family">Wang</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>Zero-shot detectors for LLM-generated text aggregate token-level statistics uniformly, implicitly assuming that every position contributes equally to detection. We challenge this assumption and show that detection signal is highly uneven across token positions: positions of high contextual dependency, where the surrogate model relies on concentrated context, carry far more discriminative power than others. Using attention entropy to quantify this dependency, we empirically validate that aggregating only low-entropy tokens yields clearly separated human and LLM score distributions, whereas high-entropy tokens do not. Based on this finding, we propose Token Contextual Dependency (TCD), a plug-and-play framework that reweights token contributions in existing zero-shot detectors via attention entropy-based weights computed from dynamically selected layers. Experiments on state-of-the-art detectors across multiple benchmarks demonstrate consistent and meaningful improvements in detection performance. Our code is available at: https://github.com/RecursionM/TCD.</abstract>
<identifier type="citekey">ma-wang-2026-tcd</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1708/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38911</start>
<end>38926</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T TCD: Empowering Zero-Shot Detection of LLM-Generated Text using Token Contextual Dependency
%A Ma, Shihan
%A Wang, Quan
%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 ma-wang-2026-tcd
%X Zero-shot detectors for LLM-generated text aggregate token-level statistics uniformly, implicitly assuming that every position contributes equally to detection. We challenge this assumption and show that detection signal is highly uneven across token positions: positions of high contextual dependency, where the surrogate model relies on concentrated context, carry far more discriminative power than others. Using attention entropy to quantify this dependency, we empirically validate that aggregating only low-entropy tokens yields clearly separated human and LLM score distributions, whereas high-entropy tokens do not. Based on this finding, we propose Token Contextual Dependency (TCD), a plug-and-play framework that reweights token contributions in existing zero-shot detectors via attention entropy-based weights computed from dynamically selected layers. Experiments on state-of-the-art detectors across multiple benchmarks demonstrate consistent and meaningful improvements in detection performance. Our code is available at: https://github.com/RecursionM/TCD.
%U https://aclanthology.org/2026.emnlp-main.1708/
%P 38911-38926
Markdown (Informal)
[TCD: Empowering Zero-Shot Detection of LLM-Generated Text using Token Contextual Dependency](https://aclanthology.org/2026.emnlp-main.1708/) (Ma & Wang, EMNLP 2026)
ACL