@inproceedings{fan-liu-2026-right,
title = "Right for the Wrong Reasons: Investigating the Unrobust Nature of High Accuracy Models in Discourse Relation Prediction",
author = "Fan, Yi and
Liu, Wei",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.37/",
pages = "713--728",
ISBN = "979-8-89176-444-6",
abstract = "Large language models (LLMs) can achieve strong performance on discourse relation prediction through fine-tuning. Yet it remains unclear whether such performance reflects a robust modelling of inter-sentential relations or a reliance on shallow lexical cues. This paper presents a behavioural audit of fine-tuned models across explicit and implicit settings. In explicit discourse relation prediction, we find that fine-tuned models rely heavily on connectives as categorical triggers, rather than integrating argument semantics. More surprisingly, in implicit discourse settings, we observe a severe performance degradation when masking semantically light function words from the arguments, suggesting that models may also exploit subtle surface-level cues beyond explicit connectives. These results suggest that high accuracy alone is insufficient to demonstrate robust discourse processing. To address this limitation, we propose a causally validated diagnostic framework that measures whether model predictions are grounded in linguistically meaningful evidence. Our metric provides a complementary perspective to standard accuracy-based evaluation and establishes a rigorous auditing toolkit for future work on discourse relation prediction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fan-liu-2026-right">
<titleInfo>
<title>Right for the Wrong Reasons: Investigating the Unrobust Nature of High Accuracy Models in Discourse Relation Prediction</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Large language models (LLMs) can achieve strong performance on discourse relation prediction through fine-tuning. Yet it remains unclear whether such performance reflects a robust modelling of inter-sentential relations or a reliance on shallow lexical cues. This paper presents a behavioural audit of fine-tuned models across explicit and implicit settings. In explicit discourse relation prediction, we find that fine-tuned models rely heavily on connectives as categorical triggers, rather than integrating argument semantics. More surprisingly, in implicit discourse settings, we observe a severe performance degradation when masking semantically light function words from the arguments, suggesting that models may also exploit subtle surface-level cues beyond explicit connectives. These results suggest that high accuracy alone is insufficient to demonstrate robust discourse processing. To address this limitation, we propose a causally validated diagnostic framework that measures whether model predictions are grounded in linguistically meaningful evidence. Our metric provides a complementary perspective to standard accuracy-based evaluation and establishes a rigorous auditing toolkit for future work on discourse relation prediction.</abstract>
<identifier type="citekey">fan-liu-2026-right</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.37/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>713</start>
<end>728</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Right for the Wrong Reasons: Investigating the Unrobust Nature of High Accuracy Models in Discourse Relation Prediction
%A Fan, Yi
%A Liu, Wei
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F fan-liu-2026-right
%X Large language models (LLMs) can achieve strong performance on discourse relation prediction through fine-tuning. Yet it remains unclear whether such performance reflects a robust modelling of inter-sentential relations or a reliance on shallow lexical cues. This paper presents a behavioural audit of fine-tuned models across explicit and implicit settings. In explicit discourse relation prediction, we find that fine-tuned models rely heavily on connectives as categorical triggers, rather than integrating argument semantics. More surprisingly, in implicit discourse settings, we observe a severe performance degradation when masking semantically light function words from the arguments, suggesting that models may also exploit subtle surface-level cues beyond explicit connectives. These results suggest that high accuracy alone is insufficient to demonstrate robust discourse processing. To address this limitation, we propose a causally validated diagnostic framework that measures whether model predictions are grounded in linguistically meaningful evidence. Our metric provides a complementary perspective to standard accuracy-based evaluation and establishes a rigorous auditing toolkit for future work on discourse relation prediction.
%U https://aclanthology.org/2026.findings-aacl.37/
%P 713-728
Markdown (Informal)
[Right for the Wrong Reasons: Investigating the Unrobust Nature of High Accuracy Models in Discourse Relation Prediction](https://aclanthology.org/2026.findings-aacl.37/) (Fan & Liu, Findings 2026)
ACL