@inproceedings{chae-etal-2026-exponential,
title = "Exponential-Tilted Unlearning",
author = "Chae, Sungwon and
Kim, Keonwoo and
Kim, Hoki",
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.2431/",
pages = "52173--52199",
ISBN = "979-8-89176-462-0",
abstract = "Large Language Models (LLMs) often memorize and regenerate sensitive or protected content that requires inference-time suppression. While machine unlearning has emerged as a highly active area of research, existing unlearning methods often degrade performance on retain dataset or require costly retraining. In this paper, we introduce Exponential-Tilted Unlearning (ETU), an inference-time method that formulates unlearning as the KL I-projection of the base next-token distribution under a user-specified upper bound $\varepsilon$ on the forget-token mass, with a closed-form exponential tilt whose scalar multiplier $\lambda$ is set analytically by $\varepsilon$. Under a single shared protocol, (i) ETU satisfies the $\varepsilon$ constraint across nine instruction-tuned dense and Mixture-of-Experts backbones on TOFU, RWKU, and WPU while preserving retain utility; (ii) the same calibrated family covers the full $\varepsilon$ spectrum with the strict zero-mass case as a limiting operating point; and (iii) representative weight-update methods expose no such $\varepsilon$-calibrated residual-mass control."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chae-etal-2026-exponential">
<titleInfo>
<title>Exponential-Tilted Unlearning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sungwon</namePart>
<namePart type="family">Chae</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Keonwoo</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hoki</namePart>
<namePart type="family">Kim</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>Large Language Models (LLMs) often memorize and regenerate sensitive or protected content that requires inference-time suppression. While machine unlearning has emerged as a highly active area of research, existing unlearning methods often degrade performance on retain dataset or require costly retraining. In this paper, we introduce Exponential-Tilted Unlearning (ETU), an inference-time method that formulates unlearning as the KL I-projection of the base next-token distribution under a user-specified upper bound ǎrepsilon on the forget-token mass, with a closed-form exponential tilt whose scalar multiplier łambda is set analytically by ǎrepsilon. Under a single shared protocol, (i) ETU satisfies the ǎrepsilon constraint across nine instruction-tuned dense and Mixture-of-Experts backbones on TOFU, RWKU, and WPU while preserving retain utility; (ii) the same calibrated family covers the full ǎrepsilon spectrum with the strict zero-mass case as a limiting operating point; and (iii) representative weight-update methods expose no such ǎrepsilon-calibrated residual-mass control.</abstract>
<identifier type="citekey">chae-etal-2026-exponential</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2431/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>52173</start>
<end>52199</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Exponential-Tilted Unlearning
%A Chae, Sungwon
%A Kim, Keonwoo
%A Kim, Hoki
%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 chae-etal-2026-exponential
%X Large Language Models (LLMs) often memorize and regenerate sensitive or protected content that requires inference-time suppression. While machine unlearning has emerged as a highly active area of research, existing unlearning methods often degrade performance on retain dataset or require costly retraining. In this paper, we introduce Exponential-Tilted Unlearning (ETU), an inference-time method that formulates unlearning as the KL I-projection of the base next-token distribution under a user-specified upper bound ǎrepsilon on the forget-token mass, with a closed-form exponential tilt whose scalar multiplier łambda is set analytically by ǎrepsilon. Under a single shared protocol, (i) ETU satisfies the ǎrepsilon constraint across nine instruction-tuned dense and Mixture-of-Experts backbones on TOFU, RWKU, and WPU while preserving retain utility; (ii) the same calibrated family covers the full ǎrepsilon spectrum with the strict zero-mass case as a limiting operating point; and (iii) representative weight-update methods expose no such ǎrepsilon-calibrated residual-mass control.
%U https://aclanthology.org/2026.findings-emnlp.2431/
%P 52173-52199
Markdown (Informal)
[Exponential-Tilted Unlearning](https://aclanthology.org/2026.findings-emnlp.2431/) (Chae et al., Findings 2026)
ACL
- Sungwon Chae, Keonwoo Kim, and Hoki Kim. 2026. Exponential-Tilted Unlearning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 52173–52199, Budapest, Hungary. Association for Computational Linguistics.