@inproceedings{hariprasad-etal-2026-structure,
title = "When Structure Is Hidden: Measuring and Diagnosing the Compositional Gap in {LLM}s",
author = "Hariprasad, Shravani and
Mohammed, Mohsin Ali and
Jain, Vinija and
Chadha, Aman",
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.1792/",
pages = "40834--40849",
ISBN = "979-8-89176-459-0",
abstract = "Large language models often recover the local pieces of a problem yet still fail to assemble them into a correct global solution. We study this compositionality gap in settings where the underlying structure is implicit and must be inferred from the input, using a two-pass framework that first verifies piece-level understanding and then measures composition on the same example via P(composed wrong | all pieces correct). We evaluate task formulations in five domains (contracts, math, SQL, code, and logic) and four model families; across the gold-anchored domains, larger conditional gaps are associated with settings where relevant structure is less directly exposed. A controlled intervention with correct and incorrect structural hints indicates heterogeneous bottlenecks: hints substantially reduce the gap in contracts and SQL, have no aggregate effect in math, and have no reliable effect in logic. Structure-aware prompting does not repair the gap and often harms smaller models. We conclude that composition failure is not a single phenomenon, and that diagnosing the dominant bottleneck in each task is the practical output of our framework"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hariprasad-etal-2026-structure">
<titleInfo>
<title>When Structure Is Hidden: Measuring and Diagnosing the Compositional Gap in LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shravani</namePart>
<namePart type="family">Hariprasad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohsin</namePart>
<namePart type="given">Ali</namePart>
<namePart type="family">Mohammed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vinija</namePart>
<namePart type="family">Jain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aman</namePart>
<namePart type="family">Chadha</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>Large language models often recover the local pieces of a problem yet still fail to assemble them into a correct global solution. We study this compositionality gap in settings where the underlying structure is implicit and must be inferred from the input, using a two-pass framework that first verifies piece-level understanding and then measures composition on the same example via P(composed wrong | all pieces correct). We evaluate task formulations in five domains (contracts, math, SQL, code, and logic) and four model families; across the gold-anchored domains, larger conditional gaps are associated with settings where relevant structure is less directly exposed. A controlled intervention with correct and incorrect structural hints indicates heterogeneous bottlenecks: hints substantially reduce the gap in contracts and SQL, have no aggregate effect in math, and have no reliable effect in logic. Structure-aware prompting does not repair the gap and often harms smaller models. We conclude that composition failure is not a single phenomenon, and that diagnosing the dominant bottleneck in each task is the practical output of our framework</abstract>
<identifier type="citekey">hariprasad-etal-2026-structure</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1792/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>40834</start>
<end>40849</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Structure Is Hidden: Measuring and Diagnosing the Compositional Gap in LLMs
%A Hariprasad, Shravani
%A Mohammed, Mohsin Ali
%A Jain, Vinija
%A Chadha, Aman
%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 hariprasad-etal-2026-structure
%X Large language models often recover the local pieces of a problem yet still fail to assemble them into a correct global solution. We study this compositionality gap in settings where the underlying structure is implicit and must be inferred from the input, using a two-pass framework that first verifies piece-level understanding and then measures composition on the same example via P(composed wrong | all pieces correct). We evaluate task formulations in five domains (contracts, math, SQL, code, and logic) and four model families; across the gold-anchored domains, larger conditional gaps are associated with settings where relevant structure is less directly exposed. A controlled intervention with correct and incorrect structural hints indicates heterogeneous bottlenecks: hints substantially reduce the gap in contracts and SQL, have no aggregate effect in math, and have no reliable effect in logic. Structure-aware prompting does not repair the gap and often harms smaller models. We conclude that composition failure is not a single phenomenon, and that diagnosing the dominant bottleneck in each task is the practical output of our framework
%U https://aclanthology.org/2026.emnlp-main.1792/
%P 40834-40849
Markdown (Informal)
[When Structure Is Hidden: Measuring and Diagnosing the Compositional Gap in LLMs](https://aclanthology.org/2026.emnlp-main.1792/) (Hariprasad et al., EMNLP 2026)
ACL