@inproceedings{wang-etal-2026-memescope,
title = "{M}eme{S}cope: Probing and Causally Localizing Pragmatic Inference in {LVLM} Meme Understanding",
author = "Wang, Tongguan and
Wu, You and
Zhou, Yuling and
Chen, Siyuan and
Li, Junkai and
Xue, Feiyue and
Chen, Wei and
Liu, Hui and
Ba, Xiqiao and
Sha, Ying",
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.951/",
pages = "21923--21945",
ISBN = "979-8-89176-459-0",
abstract = "Memes convey implicit pragmatic meanings through tightly coupled visual and textual cues, making their interpretation more challenging than that of conventional image-text pairs. Although large vision-language models (LVLMs) perform well on multimodal benchmarks, it remains unclear whether they infer non-literal mechanisms underlying memes, such as irony, wordplay, or rely mainly on superficial visual-textual correlations. To study this problem, we introduce MemeScope, a fine-grained benchmark of 1,781 memes annotated with eight pragmatic mechanisms and stage-wise explanations covering visual grounding, cross-modal alignment, and pragmatic inference. Building on MemeScope, we propose a unified interpretability framework that combines layer-wise probing with hidden-state patching to examine both the emergence and the causal contributions of meme-level pragmatic representations within LVLMs. Our analysis shows that literal visual and textual cues are often captured in earlier layers, whereas pragmatic mechanism information emerges more strongly in higher layers and shows stronger patching effects in late-layer multimodal representations. We further identify a recurring pragmatic drift, where models ground meme content correctly but move toward plausible yet incorrect mechanisms during deeper decoding."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-memescope">
<titleInfo>
<title>MemeScope: Probing and Causally Localizing Pragmatic Inference in LVLM Meme Understanding</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tongguan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">You</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuling</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyuan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junkai</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feiyue</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiqiao</namePart>
<namePart type="family">Ba</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ying</namePart>
<namePart type="family">Sha</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>Memes convey implicit pragmatic meanings through tightly coupled visual and textual cues, making their interpretation more challenging than that of conventional image-text pairs. Although large vision-language models (LVLMs) perform well on multimodal benchmarks, it remains unclear whether they infer non-literal mechanisms underlying memes, such as irony, wordplay, or rely mainly on superficial visual-textual correlations. To study this problem, we introduce MemeScope, a fine-grained benchmark of 1,781 memes annotated with eight pragmatic mechanisms and stage-wise explanations covering visual grounding, cross-modal alignment, and pragmatic inference. Building on MemeScope, we propose a unified interpretability framework that combines layer-wise probing with hidden-state patching to examine both the emergence and the causal contributions of meme-level pragmatic representations within LVLMs. Our analysis shows that literal visual and textual cues are often captured in earlier layers, whereas pragmatic mechanism information emerges more strongly in higher layers and shows stronger patching effects in late-layer multimodal representations. We further identify a recurring pragmatic drift, where models ground meme content correctly but move toward plausible yet incorrect mechanisms during deeper decoding.</abstract>
<identifier type="citekey">wang-etal-2026-memescope</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.951/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>21923</start>
<end>21945</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MemeScope: Probing and Causally Localizing Pragmatic Inference in LVLM Meme Understanding
%A Wang, Tongguan
%A Wu, You
%A Zhou, Yuling
%A Chen, Siyuan
%A Li, Junkai
%A Xue, Feiyue
%A Chen, Wei
%A Liu, Hui
%A Ba, Xiqiao
%A Sha, Ying
%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 wang-etal-2026-memescope
%X Memes convey implicit pragmatic meanings through tightly coupled visual and textual cues, making their interpretation more challenging than that of conventional image-text pairs. Although large vision-language models (LVLMs) perform well on multimodal benchmarks, it remains unclear whether they infer non-literal mechanisms underlying memes, such as irony, wordplay, or rely mainly on superficial visual-textual correlations. To study this problem, we introduce MemeScope, a fine-grained benchmark of 1,781 memes annotated with eight pragmatic mechanisms and stage-wise explanations covering visual grounding, cross-modal alignment, and pragmatic inference. Building on MemeScope, we propose a unified interpretability framework that combines layer-wise probing with hidden-state patching to examine both the emergence and the causal contributions of meme-level pragmatic representations within LVLMs. Our analysis shows that literal visual and textual cues are often captured in earlier layers, whereas pragmatic mechanism information emerges more strongly in higher layers and shows stronger patching effects in late-layer multimodal representations. We further identify a recurring pragmatic drift, where models ground meme content correctly but move toward plausible yet incorrect mechanisms during deeper decoding.
%U https://aclanthology.org/2026.emnlp-main.951/
%P 21923-21945
Markdown (Informal)
[MemeScope: Probing and Causally Localizing Pragmatic Inference in LVLM Meme Understanding](https://aclanthology.org/2026.emnlp-main.951/) (Wang et al., EMNLP 2026)
ACL
- Tongguan Wang, You Wu, Yuling Zhou, Siyuan Chen, Junkai Li, Feiyue Xue, Wei Chen, Hui Liu, Xiqiao Ba, and Ying Sha. 2026. MemeScope: Probing and Causally Localizing Pragmatic Inference in LVLM Meme Understanding. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 21923–21945, Budapest, Hungary. Association for Computational Linguistics.