@inproceedings{xiao-etal-2026-learnable,
title = "Learnable Termination Safeguard: Token-Level Early Termination for Efficient {LLM} Safety Monitoring",
author = "Xiao, Yulong and
Lin, Shanshan and
Chen, Chao and
Yang, Yizhe and
Zhang, Xi and
Liao, Xiangwen",
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.383/",
pages = "8084--8098",
ISBN = "979-8-89176-462-0",
abstract = "Large language models exhibit non-negligible risks of generating policy-violating content, underscoring the need for effective and timely guardrail mechanisms. Existing streaming safety monitoring approaches enable detection during decoding, but still lack a reliable mechanism for deciding when to terminate generation, limiting their ability to act on early risk signals that are often sparse and unstable under partial observability. In response, we propose LTS, a token-level framework that explicitly learns a stopping criterion for preemptive early termination. LTS introduces a risk-aware STOP head for token-level risk assessment and leverages temporally structured supervision with multiscale evidence to model the evolution of safety risks during generation. By casting streaming safety monitoring as a constrained online stopping task and incorporating exponential moving average temporal aggregation, LTS enables reliable translation of emerging risk evidence into stable and timely termination decisions. LTS achieves an F1 score of \textbf{95.5{\%}} and is able to terminate harmful content at \textbf{13{\%}} of the generated sequence, substantially outperforming existing methods during early generation. In realistic streaming settings, it achieves around \textbf{90{\%}} token reduction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xiao-etal-2026-learnable">
<titleInfo>
<title>Learnable Termination Safeguard: Token-Level Early Termination for Efficient LLM Safety Monitoring</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yulong</namePart>
<namePart type="family">Xiao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shanshan</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chao</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yizhe</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xi</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiangwen</namePart>
<namePart type="family">Liao</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 exhibit non-negligible risks of generating policy-violating content, underscoring the need for effective and timely guardrail mechanisms. Existing streaming safety monitoring approaches enable detection during decoding, but still lack a reliable mechanism for deciding when to terminate generation, limiting their ability to act on early risk signals that are often sparse and unstable under partial observability. In response, we propose LTS, a token-level framework that explicitly learns a stopping criterion for preemptive early termination. LTS introduces a risk-aware STOP head for token-level risk assessment and leverages temporally structured supervision with multiscale evidence to model the evolution of safety risks during generation. By casting streaming safety monitoring as a constrained online stopping task and incorporating exponential moving average temporal aggregation, LTS enables reliable translation of emerging risk evidence into stable and timely termination decisions. LTS achieves an F1 score of 95.5% and is able to terminate harmful content at 13% of the generated sequence, substantially outperforming existing methods during early generation. In realistic streaming settings, it achieves around 90% token reduction.</abstract>
<identifier type="citekey">xiao-etal-2026-learnable</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.383/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>8084</start>
<end>8098</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Learnable Termination Safeguard: Token-Level Early Termination for Efficient LLM Safety Monitoring
%A Xiao, Yulong
%A Lin, Shanshan
%A Chen, Chao
%A Yang, Yizhe
%A Zhang, Xi
%A Liao, Xiangwen
%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 xiao-etal-2026-learnable
%X Large language models exhibit non-negligible risks of generating policy-violating content, underscoring the need for effective and timely guardrail mechanisms. Existing streaming safety monitoring approaches enable detection during decoding, but still lack a reliable mechanism for deciding when to terminate generation, limiting their ability to act on early risk signals that are often sparse and unstable under partial observability. In response, we propose LTS, a token-level framework that explicitly learns a stopping criterion for preemptive early termination. LTS introduces a risk-aware STOP head for token-level risk assessment and leverages temporally structured supervision with multiscale evidence to model the evolution of safety risks during generation. By casting streaming safety monitoring as a constrained online stopping task and incorporating exponential moving average temporal aggregation, LTS enables reliable translation of emerging risk evidence into stable and timely termination decisions. LTS achieves an F1 score of 95.5% and is able to terminate harmful content at 13% of the generated sequence, substantially outperforming existing methods during early generation. In realistic streaming settings, it achieves around 90% token reduction.
%U https://aclanthology.org/2026.findings-emnlp.383/
%P 8084-8098
Markdown (Informal)
[Learnable Termination Safeguard: Token-Level Early Termination for Efficient LLM Safety Monitoring](https://aclanthology.org/2026.findings-emnlp.383/) (Xiao et al., Findings 2026)
ACL