@inproceedings{cao-etal-2026-cmie,
title = "{CMIE}: Cross-Modal Interaction Erasure for {MLLM} Safety Unlearning via Counterfactual Localization and Subspace Nullification",
author = "Cao, Jiatian and
Cai, Linyi and
Wang, Mengqi and
Liu, Jinduo and
Ji, Junzhong and
Du, Yongping and
Han, Honggui",
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.2220/",
pages = "50518--50541",
ISBN = "979-8-89176-459-0",
abstract = "Safety unlearning for multimodal large language models (MLLMs) fails when it treats danger as an object, topic, or prompt template. A useful MLLM must still recognize risky visual evidence and answer benign educational or preventive questions about the same content. The safety-critical target is narrower: the post-fusion interaction that converts risky visual evidence and unsafe textual intent into actionable harmful assistance. Existing refusal-oriented and unlearning objectives reach this mechanism only indirectly, widening refusal boundaries or leaving recoverable multimodal harmful pathways. We propose CMIE, Cross-Modal Interaction Erasure, a mechanism-level method that directly treats the harmful image-text interaction as the object of erasure. CMIE first estimates harmful interaction directions with intent-matched counterfactual contrasts that cancel unimodal main effects. It then localizes behaviorally active post-fusion layers and estimates a safe-whitened low-rank harmful subspace. A closed-form minimum-distortion weight edit erases this subspace, and projection-constrained recovery restores benign utility without rewriting erased directions. Across SafeEraser, VLGuard, and MSSBench, CMIE achieves the best overall mean rank and outperforms strong baselines on multiple safety, utility, visual selectivity, and paired-context metrics. Warning: This paper contains examples of harmful language and images, and reader discretion is recommended."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cao-etal-2026-cmie">
<titleInfo>
<title>CMIE: Cross-Modal Interaction Erasure for MLLM Safety Unlearning via Counterfactual Localization and Subspace Nullification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jiatian</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linyi</namePart>
<namePart type="family">Cai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mengqi</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinduo</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junzhong</namePart>
<namePart type="family">Ji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yongping</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Honggui</namePart>
<namePart type="family">Han</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>Safety unlearning for multimodal large language models (MLLMs) fails when it treats danger as an object, topic, or prompt template. A useful MLLM must still recognize risky visual evidence and answer benign educational or preventive questions about the same content. The safety-critical target is narrower: the post-fusion interaction that converts risky visual evidence and unsafe textual intent into actionable harmful assistance. Existing refusal-oriented and unlearning objectives reach this mechanism only indirectly, widening refusal boundaries or leaving recoverable multimodal harmful pathways. We propose CMIE, Cross-Modal Interaction Erasure, a mechanism-level method that directly treats the harmful image-text interaction as the object of erasure. CMIE first estimates harmful interaction directions with intent-matched counterfactual contrasts that cancel unimodal main effects. It then localizes behaviorally active post-fusion layers and estimates a safe-whitened low-rank harmful subspace. A closed-form minimum-distortion weight edit erases this subspace, and projection-constrained recovery restores benign utility without rewriting erased directions. Across SafeEraser, VLGuard, and MSSBench, CMIE achieves the best overall mean rank and outperforms strong baselines on multiple safety, utility, visual selectivity, and paired-context metrics. Warning: This paper contains examples of harmful language and images, and reader discretion is recommended.</abstract>
<identifier type="citekey">cao-etal-2026-cmie</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2220/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>50518</start>
<end>50541</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CMIE: Cross-Modal Interaction Erasure for MLLM Safety Unlearning via Counterfactual Localization and Subspace Nullification
%A Cao, Jiatian
%A Cai, Linyi
%A Wang, Mengqi
%A Liu, Jinduo
%A Ji, Junzhong
%A Du, Yongping
%A Han, Honggui
%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 cao-etal-2026-cmie
%X Safety unlearning for multimodal large language models (MLLMs) fails when it treats danger as an object, topic, or prompt template. A useful MLLM must still recognize risky visual evidence and answer benign educational or preventive questions about the same content. The safety-critical target is narrower: the post-fusion interaction that converts risky visual evidence and unsafe textual intent into actionable harmful assistance. Existing refusal-oriented and unlearning objectives reach this mechanism only indirectly, widening refusal boundaries or leaving recoverable multimodal harmful pathways. We propose CMIE, Cross-Modal Interaction Erasure, a mechanism-level method that directly treats the harmful image-text interaction as the object of erasure. CMIE first estimates harmful interaction directions with intent-matched counterfactual contrasts that cancel unimodal main effects. It then localizes behaviorally active post-fusion layers and estimates a safe-whitened low-rank harmful subspace. A closed-form minimum-distortion weight edit erases this subspace, and projection-constrained recovery restores benign utility without rewriting erased directions. Across SafeEraser, VLGuard, and MSSBench, CMIE achieves the best overall mean rank and outperforms strong baselines on multiple safety, utility, visual selectivity, and paired-context metrics. Warning: This paper contains examples of harmful language and images, and reader discretion is recommended.
%U https://aclanthology.org/2026.emnlp-main.2220/
%P 50518-50541
Markdown (Informal)
[CMIE: Cross-Modal Interaction Erasure for MLLM Safety Unlearning via Counterfactual Localization and Subspace Nullification](https://aclanthology.org/2026.emnlp-main.2220/) (Cao et al., EMNLP 2026)
ACL