@inproceedings{zhang-etal-2026-iris,
title = "{IRIS}: Iterative Refinement of Internal States for Safer Language Models",
author = "Zhang, Beize and
Feng, Yuchen and
Li, Peize and
Zhu, Zhitong and
Fu, Peng and
Weimin, Mu and
Lin, Zheng and
Wang, Weiping",
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.912/",
pages = "19288--19302",
ISBN = "979-8-89176-462-0",
abstract = "Attack success rate (ASR) is insufficient for evaluating the safety of large language models (LLMs). Many defenses rely on surface-level refusal behavior, which can over-refuse benign-sensitive prompts and remain brittle: once the initial refusal boundary is bypassed, the model may still follow a harmful continuation trajectory. We analyze model activation spaces to examine the limits of static interventions, finding substantial overlap among useful, benign-sensitive, and harmful representations. This suggests that a single global intervention may suppress harmful outputs only by perturbing general utility. We propose IRIS, Iterative Refinement of Internal States, a routing-conditioned internal correction method. To address this, IRIS leaves low-risk states unaltered while routing high-risk states to a shared module for multi-step latent refinement. To make the refinement module learn behavior-level correction, we introduce model-owned base-source prefix recovery, which trains IRIS to recover from base-model failure prefixes toward safe continuations. Extensive evaluations on Llama-3.1-8B and Qwen-2.5-7B demonstrate that IRIS drastically reduces ASR from around 70{\%} to around 10{\%} and successfully defends against prefilling attacks. Furthermore, IRIS preserves general benchmark utility and maintains near-zero over-refusal rates."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-iris">
<titleInfo>
<title>IRIS: Iterative Refinement of Internal States for Safer Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Beize</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuchen</namePart>
<namePart type="family">Feng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peize</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhitong</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peng</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mu</namePart>
<namePart type="family">Weimin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zheng</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weiping</namePart>
<namePart type="family">Wang</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>Attack success rate (ASR) is insufficient for evaluating the safety of large language models (LLMs). Many defenses rely on surface-level refusal behavior, which can over-refuse benign-sensitive prompts and remain brittle: once the initial refusal boundary is bypassed, the model may still follow a harmful continuation trajectory. We analyze model activation spaces to examine the limits of static interventions, finding substantial overlap among useful, benign-sensitive, and harmful representations. This suggests that a single global intervention may suppress harmful outputs only by perturbing general utility. We propose IRIS, Iterative Refinement of Internal States, a routing-conditioned internal correction method. To address this, IRIS leaves low-risk states unaltered while routing high-risk states to a shared module for multi-step latent refinement. To make the refinement module learn behavior-level correction, we introduce model-owned base-source prefix recovery, which trains IRIS to recover from base-model failure prefixes toward safe continuations. Extensive evaluations on Llama-3.1-8B and Qwen-2.5-7B demonstrate that IRIS drastically reduces ASR from around 70% to around 10% and successfully defends against prefilling attacks. Furthermore, IRIS preserves general benchmark utility and maintains near-zero over-refusal rates.</abstract>
<identifier type="citekey">zhang-etal-2026-iris</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.912/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19288</start>
<end>19302</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T IRIS: Iterative Refinement of Internal States for Safer Language Models
%A Zhang, Beize
%A Feng, Yuchen
%A Li, Peize
%A Zhu, Zhitong
%A Fu, Peng
%A Weimin, Mu
%A Lin, Zheng
%A Wang, Weiping
%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 zhang-etal-2026-iris
%X Attack success rate (ASR) is insufficient for evaluating the safety of large language models (LLMs). Many defenses rely on surface-level refusal behavior, which can over-refuse benign-sensitive prompts and remain brittle: once the initial refusal boundary is bypassed, the model may still follow a harmful continuation trajectory. We analyze model activation spaces to examine the limits of static interventions, finding substantial overlap among useful, benign-sensitive, and harmful representations. This suggests that a single global intervention may suppress harmful outputs only by perturbing general utility. We propose IRIS, Iterative Refinement of Internal States, a routing-conditioned internal correction method. To address this, IRIS leaves low-risk states unaltered while routing high-risk states to a shared module for multi-step latent refinement. To make the refinement module learn behavior-level correction, we introduce model-owned base-source prefix recovery, which trains IRIS to recover from base-model failure prefixes toward safe continuations. Extensive evaluations on Llama-3.1-8B and Qwen-2.5-7B demonstrate that IRIS drastically reduces ASR from around 70% to around 10% and successfully defends against prefilling attacks. Furthermore, IRIS preserves general benchmark utility and maintains near-zero over-refusal rates.
%U https://aclanthology.org/2026.findings-emnlp.912/
%P 19288-19302
Markdown (Informal)
[IRIS: Iterative Refinement of Internal States for Safer Language Models](https://aclanthology.org/2026.findings-emnlp.912/) (Zhang et al., Findings 2026)
ACL
- Beize Zhang, Yuchen Feng, Peize Li, Zhitong Zhu, Peng Fu, Mu Weimin, Zheng Lin, and Weiping Wang. 2026. IRIS: Iterative Refinement of Internal States for Safer Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 19288–19302, Budapest, Hungary. Association for Computational Linguistics.