@inproceedings{zhao-etal-2026-counterfactual,
title = "Counterfactual Self-Distillation: A Distribution-Shaping Framework for Entity-Fact Unlearning in Large Language Models",
author = "Zhao, Ying and
Zhang, Yihui and
Deng, Mengyi and
Li, Xin and
Tian, Yuan and
Wang, Wei and
Chang, Yi",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1522/",
pages = "32401--32414",
ISBN = "979-8-89176-462-0",
abstract = "Entity-fact unlearning in large language models (LLMs) aims to remove the binding between a target entity and a forgotten fact, producing a model that behaves like one retrained without the forgotten data while preserving neighboring knowledge and overall model utility. We frame this problem as distribution shaping: such a retrained model would yield a plausible distribution at forget queries, fluent, semantically appropriate, and reflective of residual knowledge, which we take as the appropriate target. Existing objectives provide only narrow token-level or semantically uninformative supervision, falling short of this target. We propose Counterfactual Self-Distillation (CSD), a self-distillation framework that approximates the retrained distribution by constructing a counterfactual teacher through type-consistent entity substitution. The resulting teacher is decoupled from the target identity while preserving type-level relational structure, enabling directed probability reallocation without an external oracle. Experiments on TOFU and RWKU show that CSD achieves a strong forget-utility trade-off, including the best overall forgetting on RWKU and strong performance on TOFU."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhao-etal-2026-counterfactual">
<titleInfo>
<title>Counterfactual Self-Distillation: A Distribution-Shaping Framework for Entity-Fact Unlearning in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ying</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yihui</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mengyi</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuan</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Chang</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Entity-fact unlearning in large language models (LLMs) aims to remove the binding between a target entity and a forgotten fact, producing a model that behaves like one retrained without the forgotten data while preserving neighboring knowledge and overall model utility. We frame this problem as distribution shaping: such a retrained model would yield a plausible distribution at forget queries, fluent, semantically appropriate, and reflective of residual knowledge, which we take as the appropriate target. Existing objectives provide only narrow token-level or semantically uninformative supervision, falling short of this target. We propose Counterfactual Self-Distillation (CSD), a self-distillation framework that approximates the retrained distribution by constructing a counterfactual teacher through type-consistent entity substitution. The resulting teacher is decoupled from the target identity while preserving type-level relational structure, enabling directed probability reallocation without an external oracle. Experiments on TOFU and RWKU show that CSD achieves a strong forget-utility trade-off, including the best overall forgetting on RWKU and strong performance on TOFU.</abstract>
<identifier type="citekey">zhao-etal-2026-counterfactual</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1522/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>32401</start>
<end>32414</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Counterfactual Self-Distillation: A Distribution-Shaping Framework for Entity-Fact Unlearning in Large Language Models
%A Zhao, Ying
%A Zhang, Yihui
%A Deng, Mengyi
%A Li, Xin
%A Tian, Yuan
%A Wang, Wei
%A Chang, Yi
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F zhao-etal-2026-counterfactual
%X Entity-fact unlearning in large language models (LLMs) aims to remove the binding between a target entity and a forgotten fact, producing a model that behaves like one retrained without the forgotten data while preserving neighboring knowledge and overall model utility. We frame this problem as distribution shaping: such a retrained model would yield a plausible distribution at forget queries, fluent, semantically appropriate, and reflective of residual knowledge, which we take as the appropriate target. Existing objectives provide only narrow token-level or semantically uninformative supervision, falling short of this target. We propose Counterfactual Self-Distillation (CSD), a self-distillation framework that approximates the retrained distribution by constructing a counterfactual teacher through type-consistent entity substitution. The resulting teacher is decoupled from the target identity while preserving type-level relational structure, enabling directed probability reallocation without an external oracle. Experiments on TOFU and RWKU show that CSD achieves a strong forget-utility trade-off, including the best overall forgetting on RWKU and strong performance on TOFU.
%U https://aclanthology.org/2026.findings-emnlp.1522/
%P 32401-32414
Markdown (Informal)
[Counterfactual Self-Distillation: A Distribution-Shaping Framework for Entity-Fact Unlearning in Large Language Models](https://aclanthology.org/2026.findings-emnlp.1522/) (Zhao et al., Findings 2026)
ACL