@inproceedings{ferrao-etal-2026-paved,
title = "Paved with True Intents: Intent-Aware Training Improves {LLM} Safety Classification Across Training Regimes",
author = {Ferrao, Jeremias Lino and
M{\"u}ller-Hof, Niclas and
Sirbu, Iustin and
Rebedea, Traian and
Ziser, Yftah},
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.1833/",
pages = "39128--39155",
ISBN = "979-8-89176-462-0",
abstract = "We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce **AIMS**, a human-annotated dataset containing 1,724 intent-label annotations over 1,275 unique difficult safety prompts, each paired with an intent description and harm label. We use **AIMS** to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, **AIMS** enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher{--}student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency{--}F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ferrao-etal-2026-paved">
<titleInfo>
<title>Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jeremias</namePart>
<namePart type="given">Lino</namePart>
<namePart type="family">Ferrao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Niclas</namePart>
<namePart type="family">Müller-Hof</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Iustin</namePart>
<namePart type="family">Sirbu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Traian</namePart>
<namePart type="family">Rebedea</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yftah</namePart>
<namePart type="family">Ziser</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>We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce **AIMS**, a human-annotated dataset containing 1,724 intent-label annotations over 1,275 unique difficult safety prompts, each paired with an intent description and harm label. We use **AIMS** to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, **AIMS** enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher–student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency–F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.</abstract>
<identifier type="citekey">ferrao-etal-2026-paved</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1833/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>39128</start>
<end>39155</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes
%A Ferrao, Jeremias Lino
%A Müller-Hof, Niclas
%A Sirbu, Iustin
%A Rebedea, Traian
%A Ziser, Yftah
%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 ferrao-etal-2026-paved
%X We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce **AIMS**, a human-annotated dataset containing 1,724 intent-label annotations over 1,275 unique difficult safety prompts, each paired with an intent description and harm label. We use **AIMS** to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, **AIMS** enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher–student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency–F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.
%U https://aclanthology.org/2026.findings-emnlp.1833/
%P 39128-39155
Markdown (Informal)
[Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes](https://aclanthology.org/2026.findings-emnlp.1833/) (Ferrao et al., Findings 2026)
ACL