@inproceedings{bukkapatnam-2026-compositional,
title = "The Compositional Grounding Gap: Why Vision-Language Models Fail at Relational Reasoning and How to Fix It",
author = "Bukkapatnam, Kaustubh S.",
editor = "Yan, Qianqi and
Montariol, Syrielle and
Fan, Yue and
Gu, Jing and
Pan, Jiayi and
Li, Manling and
Kordjamshidi, Parisa and
Suhr, Alane and
Wang, Xin Eric",
booktitle = "Proceedings of the 4th Workshop on Advances in Language and Vision Research ({ALVR})",
month = jul,
year = "2026",
address = "San Diego, California, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.alvr-main.28/",
pages = "287--293",
ISBN = "979-8-89176-398-2",
abstract = "Large vision-language models (LVLMs) achieve strong performance on many multimodal tasks, yet consistently fail at compositional relational reasoning{---}distinguishing ``the cat on the mat'' from ``the mat on the cat.'' We provide a formal explanation for this failure. We prove that any vision-language alignment operating on pooled (order-invariant) visual features contains compositional blind spots: semantically distinct scenes that map to identical representations. We show that the number of blind spots grows factorially with scene complexity, establishing a fundamental limit on pooled-feature architectures. Motivated by this analysis, we propose REGROUND, a training-free, test-time method that re-introduces spatial structure into alignment by performing relation-guided cross-attention over spatial visual tokens, directed by a lightweight parse of the text query. Without any fine-tuning, REGROUND improves compositional accuracy by +8.6 points on Winoground, +8.4 on ARO-Relation, +6.4 on SugarCrepe, and +8.4 on VSR when applied to LLaVA-1.5, and provides consistent gains across other LVLMs. Ablation studies confirm that each component{---}parse guidance, token-level attention, and relation masking{---}contributes significantly."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bukkapatnam-2026-compositional">
<titleInfo>
<title>The Compositional Grounding Gap: Why Vision-Language Models Fail at Relational Reasoning and How to Fix It</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kaustubh</namePart>
<namePart type="given">S</namePart>
<namePart type="family">Bukkapatnam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qianqi</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Syrielle</namePart>
<namePart type="family">Montariol</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Fan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jing</namePart>
<namePart type="family">Gu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiayi</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Manling</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Parisa</namePart>
<namePart type="family">Kordjamshidi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alane</namePart>
<namePart type="family">Suhr</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="given">Eric</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-398-2</identifier>
</relatedItem>
<abstract>Large vision-language models (LVLMs) achieve strong performance on many multimodal tasks, yet consistently fail at compositional relational reasoning—distinguishing “the cat on the mat” from “the mat on the cat.” We provide a formal explanation for this failure. We prove that any vision-language alignment operating on pooled (order-invariant) visual features contains compositional blind spots: semantically distinct scenes that map to identical representations. We show that the number of blind spots grows factorially with scene complexity, establishing a fundamental limit on pooled-feature architectures. Motivated by this analysis, we propose REGROUND, a training-free, test-time method that re-introduces spatial structure into alignment by performing relation-guided cross-attention over spatial visual tokens, directed by a lightweight parse of the text query. Without any fine-tuning, REGROUND improves compositional accuracy by +8.6 points on Winoground, +8.4 on ARO-Relation, +6.4 on SugarCrepe, and +8.4 on VSR when applied to LLaVA-1.5, and provides consistent gains across other LVLMs. Ablation studies confirm that each component—parse guidance, token-level attention, and relation masking—contributes significantly.</abstract>
<identifier type="citekey">bukkapatnam-2026-compositional</identifier>
<location>
<url>https://aclanthology.org/2026.alvr-main.28/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>287</start>
<end>293</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Compositional Grounding Gap: Why Vision-Language Models Fail at Relational Reasoning and How to Fix It
%A Bukkapatnam, Kaustubh S.
%Y Yan, Qianqi
%Y Montariol, Syrielle
%Y Fan, Yue
%Y Gu, Jing
%Y Pan, Jiayi
%Y Li, Manling
%Y Kordjamshidi, Parisa
%Y Suhr, Alane
%Y Wang, Xin Eric
%S Proceedings of the 4th Workshop on Advances in Language and Vision Research (ALVR)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, USA
%@ 979-8-89176-398-2
%F bukkapatnam-2026-compositional
%X Large vision-language models (LVLMs) achieve strong performance on many multimodal tasks, yet consistently fail at compositional relational reasoning—distinguishing “the cat on the mat” from “the mat on the cat.” We provide a formal explanation for this failure. We prove that any vision-language alignment operating on pooled (order-invariant) visual features contains compositional blind spots: semantically distinct scenes that map to identical representations. We show that the number of blind spots grows factorially with scene complexity, establishing a fundamental limit on pooled-feature architectures. Motivated by this analysis, we propose REGROUND, a training-free, test-time method that re-introduces spatial structure into alignment by performing relation-guided cross-attention over spatial visual tokens, directed by a lightweight parse of the text query. Without any fine-tuning, REGROUND improves compositional accuracy by +8.6 points on Winoground, +8.4 on ARO-Relation, +6.4 on SugarCrepe, and +8.4 on VSR when applied to LLaVA-1.5, and provides consistent gains across other LVLMs. Ablation studies confirm that each component—parse guidance, token-level attention, and relation masking—contributes significantly.
%U https://aclanthology.org/2026.alvr-main.28/
%P 287-293
Markdown (Informal)
[The Compositional Grounding Gap: Why Vision-Language Models Fail at Relational Reasoning and How to Fix It](https://aclanthology.org/2026.alvr-main.28/) (Bukkapatnam, ALVR 2026)
ACL