@inproceedings{kim-etal-2026-plasticity,
title = "Plasticity without Collapse: Plasticity-Inducing Control Optimizer for Cross-Lingual Continual Pre-Training",
author = "Kim, Sumin and
Song, Minjun and
Lee, Surin and
Rahman, S M Wahidur and
Lee, Yongtae and
Choi, Haeung and
Lee, Heung-No",
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.528/",
pages = "12174--12203",
ISBN = "979-8-89176-459-0",
abstract = "Language models (LMs) operate in evolving environments where adapting to new domains and linguistic shifts is critical. Continual pre-training (CPT) addresses this but introduces a strict stability-plasticity dilemma. While current approaches manage this via data replay or architectural changes, they largely treat the optimizer as a static component. We propose the Plasticity-Inducing Control Optimizer (PICO) to tackle this dilemma directly at the update level. The optimizer defines a utility-gated update together with a stochastic perturbation path controlled by periodic spectral sensitivity checks. Structural flags suppress the perturbation path while leaving the protected-gradient update active. Across the evaluated legal and medical streams, PICO exhibits a favorable stability{--}plasticity trade-off through optimizer-level control, while incurring a measurable spectral-monitoring cost. Code is available at \url{https://github.com/infonet-gist/infonet-ai-pico}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-etal-2026-plasticity">
<titleInfo>
<title>Plasticity without Collapse: Plasticity-Inducing Control Optimizer for Cross-Lingual Continual Pre-Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sumin</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minjun</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Surin</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">S</namePart>
<namePart type="given">M</namePart>
<namePart type="given">Wahidur</namePart>
<namePart type="family">Rahman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yongtae</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haeung</namePart>
<namePart type="family">Choi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Heung-No</namePart>
<namePart type="family">Lee</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>Language models (LMs) operate in evolving environments where adapting to new domains and linguistic shifts is critical. Continual pre-training (CPT) addresses this but introduces a strict stability-plasticity dilemma. While current approaches manage this via data replay or architectural changes, they largely treat the optimizer as a static component. We propose the Plasticity-Inducing Control Optimizer (PICO) to tackle this dilemma directly at the update level. The optimizer defines a utility-gated update together with a stochastic perturbation path controlled by periodic spectral sensitivity checks. Structural flags suppress the perturbation path while leaving the protected-gradient update active. Across the evaluated legal and medical streams, PICO exhibits a favorable stability–plasticity trade-off through optimizer-level control, while incurring a measurable spectral-monitoring cost. Code is available at https://github.com/infonet-gist/infonet-ai-pico.</abstract>
<identifier type="citekey">kim-etal-2026-plasticity</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.528/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12174</start>
<end>12203</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Plasticity without Collapse: Plasticity-Inducing Control Optimizer for Cross-Lingual Continual Pre-Training
%A Kim, Sumin
%A Song, Minjun
%A Lee, Surin
%A Rahman, S. M. Wahidur
%A Lee, Yongtae
%A Choi, Haeung
%A Lee, Heung-No
%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 kim-etal-2026-plasticity
%X Language models (LMs) operate in evolving environments where adapting to new domains and linguistic shifts is critical. Continual pre-training (CPT) addresses this but introduces a strict stability-plasticity dilemma. While current approaches manage this via data replay or architectural changes, they largely treat the optimizer as a static component. We propose the Plasticity-Inducing Control Optimizer (PICO) to tackle this dilemma directly at the update level. The optimizer defines a utility-gated update together with a stochastic perturbation path controlled by periodic spectral sensitivity checks. Structural flags suppress the perturbation path while leaving the protected-gradient update active. Across the evaluated legal and medical streams, PICO exhibits a favorable stability–plasticity trade-off through optimizer-level control, while incurring a measurable spectral-monitoring cost. Code is available at https://github.com/infonet-gist/infonet-ai-pico.
%U https://aclanthology.org/2026.emnlp-main.528/
%P 12174-12203
Markdown (Informal)
[Plasticity without Collapse: Plasticity-Inducing Control Optimizer for Cross-Lingual Continual Pre-Training](https://aclanthology.org/2026.emnlp-main.528/) (Kim et al., EMNLP 2026)
ACL