@inproceedings{gu-etal-2026-text,
title = "Text Attribution as Continuous Trajectories: Paired {AUC} Metrics and Mid-Layer Allocation-Effect Decomposition",
author = "Gu, Zihan and
Zhang, Xinyan and
Zhang, TianYi and
Zhang, Han and
Zhang, Hua and
Hu, Yue",
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.1628/",
pages = "34633--34656",
ISBN = "979-8-89176-462-0",
abstract = "Attribution aims to identify which input tokens are responsible for a language model{'}s predictions, yet its evaluation in NLP remains dominated by discrete Top-k removal tests that poorly reflect global importance and interaction effects. In this work, we reformulate text attribution evaluation as a continuous functional over attribution-induced trajectories and introduce two complementary AUC-based metrics tailored to autoregressive models. \textit{Perturbation AUC} measures token necessity via progressive embedding-level corruption with a stable logit-based signal, while \textit{Recovery AUC} evaluates sufficiency by tracking the restoration of internal representations in hidden-state space. Together, these metrics subsume Top- tests and enable principled, ordering-sensitive assessment of attribution quality. Using these diagnostics, we show that most existing attribution methods collapse to near-random rankings under continuous evaluation and fail to capture even first-order interaction effects. Motivated by this failure mode, we propose a mid-layer attribution framework that exploits a layer-wise saturation property of transformer computation to decompose attribution into information allocation and causal effect. Across multiple models and tasks, our method consistently outperforms existing baselines under both Perturbation and Recovery AUC, yielding more faithful and interaction-aware explanations of language model behavior."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gu-etal-2026-text">
<titleInfo>
<title>Text Attribution as Continuous Trajectories: Paired AUC Metrics and Mid-Layer Allocation-Effect Decomposition</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zihan</namePart>
<namePart type="family">Gu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyan</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">TianYi</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Han</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hua</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Hu</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>Attribution aims to identify which input tokens are responsible for a language model’s predictions, yet its evaluation in NLP remains dominated by discrete Top-k removal tests that poorly reflect global importance and interaction effects. In this work, we reformulate text attribution evaluation as a continuous functional over attribution-induced trajectories and introduce two complementary AUC-based metrics tailored to autoregressive models. Perturbation AUC measures token necessity via progressive embedding-level corruption with a stable logit-based signal, while Recovery AUC evaluates sufficiency by tracking the restoration of internal representations in hidden-state space. Together, these metrics subsume Top- tests and enable principled, ordering-sensitive assessment of attribution quality. Using these diagnostics, we show that most existing attribution methods collapse to near-random rankings under continuous evaluation and fail to capture even first-order interaction effects. Motivated by this failure mode, we propose a mid-layer attribution framework that exploits a layer-wise saturation property of transformer computation to decompose attribution into information allocation and causal effect. Across multiple models and tasks, our method consistently outperforms existing baselines under both Perturbation and Recovery AUC, yielding more faithful and interaction-aware explanations of language model behavior.</abstract>
<identifier type="citekey">gu-etal-2026-text</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1628/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34633</start>
<end>34656</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Text Attribution as Continuous Trajectories: Paired AUC Metrics and Mid-Layer Allocation-Effect Decomposition
%A Gu, Zihan
%A Zhang, Xinyan
%A Zhang, TianYi
%A Zhang, Han
%A Zhang, Hua
%A Hu, Yue
%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 gu-etal-2026-text
%X Attribution aims to identify which input tokens are responsible for a language model’s predictions, yet its evaluation in NLP remains dominated by discrete Top-k removal tests that poorly reflect global importance and interaction effects. In this work, we reformulate text attribution evaluation as a continuous functional over attribution-induced trajectories and introduce two complementary AUC-based metrics tailored to autoregressive models. Perturbation AUC measures token necessity via progressive embedding-level corruption with a stable logit-based signal, while Recovery AUC evaluates sufficiency by tracking the restoration of internal representations in hidden-state space. Together, these metrics subsume Top- tests and enable principled, ordering-sensitive assessment of attribution quality. Using these diagnostics, we show that most existing attribution methods collapse to near-random rankings under continuous evaluation and fail to capture even first-order interaction effects. Motivated by this failure mode, we propose a mid-layer attribution framework that exploits a layer-wise saturation property of transformer computation to decompose attribution into information allocation and causal effect. Across multiple models and tasks, our method consistently outperforms existing baselines under both Perturbation and Recovery AUC, yielding more faithful and interaction-aware explanations of language model behavior.
%U https://aclanthology.org/2026.findings-emnlp.1628/
%P 34633-34656
Markdown (Informal)
[Text Attribution as Continuous Trajectories: Paired AUC Metrics and Mid-Layer Allocation-Effect Decomposition](https://aclanthology.org/2026.findings-emnlp.1628/) (Gu et al., Findings 2026)
ACL