@inproceedings{sun-etal-2026-prism,
title = "{PRISM}: Multidimensional Safety Risk Detection and Measurement Beyond Superficial Compliance for {LLM}s",
author = "Sun, YIfan and
Zhao, Xiang and
Lin, Beining and
Huang, Feng and
Song, You",
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.1627/",
pages = "34612--34632",
ISBN = "979-8-89176-462-0",
abstract = "Existing LLM safety measures operate as black boxes, lacking categorical precision and calibrated severity assessment. We propose **PRISM** (**P**rototypical **R**epresentation for **I**nternal **S**afety **M**apping), a fine-grained two-stage safety auditing framework grounded in internal mechanistic representations. Stage I localizes functional safety zones via spectral analysis and maps internal activations onto concept prototypes for camouflage-resilient categorical risk identification. Stage II decomposes violation severity across orthogonal dimensions through structured CoT reasoning, producing calibrated human-auditable ordinal verdicts. PRISM achieves a new best F1 of 77.60{\%} on the adversarial ToxicChat benchmark and approaches 671B frontier judges within a 2.3{\%} margin on severity calibration, demonstrating that mechanistic grounding and principled task formulation can substitute for scale in fine-grained safety governance."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sun-etal-2026-prism">
<titleInfo>
<title>PRISM: Multidimensional Safety Risk Detection and Measurement Beyond Superficial Compliance for LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">YIfan</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiang</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Beining</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feng</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">You</namePart>
<namePart type="family">Song</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>Existing LLM safety measures operate as black boxes, lacking categorical precision and calibrated severity assessment. We propose **PRISM** (**P**rototypical **R**epresentation for **I**nternal **S**afety **M**apping), a fine-grained two-stage safety auditing framework grounded in internal mechanistic representations. Stage I localizes functional safety zones via spectral analysis and maps internal activations onto concept prototypes for camouflage-resilient categorical risk identification. Stage II decomposes violation severity across orthogonal dimensions through structured CoT reasoning, producing calibrated human-auditable ordinal verdicts. PRISM achieves a new best F1 of 77.60% on the adversarial ToxicChat benchmark and approaches 671B frontier judges within a 2.3% margin on severity calibration, demonstrating that mechanistic grounding and principled task formulation can substitute for scale in fine-grained safety governance.</abstract>
<identifier type="citekey">sun-etal-2026-prism</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1627/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>34612</start>
<end>34632</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T PRISM: Multidimensional Safety Risk Detection and Measurement Beyond Superficial Compliance for LLMs
%A Sun, YIfan
%A Zhao, Xiang
%A Lin, Beining
%A Huang, Feng
%A Song, You
%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 sun-etal-2026-prism
%X Existing LLM safety measures operate as black boxes, lacking categorical precision and calibrated severity assessment. We propose **PRISM** (**P**rototypical **R**epresentation for **I**nternal **S**afety **M**apping), a fine-grained two-stage safety auditing framework grounded in internal mechanistic representations. Stage I localizes functional safety zones via spectral analysis and maps internal activations onto concept prototypes for camouflage-resilient categorical risk identification. Stage II decomposes violation severity across orthogonal dimensions through structured CoT reasoning, producing calibrated human-auditable ordinal verdicts. PRISM achieves a new best F1 of 77.60% on the adversarial ToxicChat benchmark and approaches 671B frontier judges within a 2.3% margin on severity calibration, demonstrating that mechanistic grounding and principled task formulation can substitute for scale in fine-grained safety governance.
%U https://aclanthology.org/2026.findings-emnlp.1627/
%P 34612-34632
Markdown (Informal)
[PRISM: Multidimensional Safety Risk Detection and Measurement Beyond Superficial Compliance for LLMs](https://aclanthology.org/2026.findings-emnlp.1627/) (Sun et al., Findings 2026)
ACL