@inproceedings{wang-etal-2026-know-know,
title = "Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing",
author = "Wang, Zongsheng and
Wu, Bowen and
yu, Qun and
Dengzhengwei and
Lin, Qin and
Chen, Xu and
Wang, Baoxun",
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.545/",
pages = "12530--12548",
ISBN = "979-8-89176-459-0",
abstract = "Role-playing conversational agents (RPCAs) built on large language models often leak out-of-boundary knowledge: when asked about topics outside a character{'}s limits, the model{'}s Chain-of-Thought reasoning reveals knowledge that the character should not possess, undermining immersion. To locate the source of this leakage, we conduct heuristic experiments at three levels of granularity, and find that intervening on individual Feed-Forward Network (FFN) neurons yields the strongest and most effective suppression. Motivated by this finding, we propose the Dynamic Attention Gating (DAG) Router, a lightweight trainable network that encodes the system prompt and dynamically generates per-layer FFN gating signals at inference time. Evaluated on multiple models and benchmarks, the DAG Router improves out-of-boundary scores by up to +162{\%} compared with the original baseline, along with better overall role-playing quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-know-know">
<titleInfo>
<title>Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zongsheng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bowen</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qun</namePart>
<namePart type="family">yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name>
<namePart>Dengzhengwei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qin</namePart>
<namePart type="family">Lin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Baoxun</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>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>Role-playing conversational agents (RPCAs) built on large language models often leak out-of-boundary knowledge: when asked about topics outside a character’s limits, the model’s Chain-of-Thought reasoning reveals knowledge that the character should not possess, undermining immersion. To locate the source of this leakage, we conduct heuristic experiments at three levels of granularity, and find that intervening on individual Feed-Forward Network (FFN) neurons yields the strongest and most effective suppression. Motivated by this finding, we propose the Dynamic Attention Gating (DAG) Router, a lightweight trainable network that encodes the system prompt and dynamically generates per-layer FFN gating signals at inference time. Evaluated on multiple models and benchmarks, the DAG Router improves out-of-boundary scores by up to +162% compared with the original baseline, along with better overall role-playing quality.</abstract>
<identifier type="citekey">wang-etal-2026-know-know</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.545/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>12530</start>
<end>12548</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing
%A Wang, Zongsheng
%A Wu, Bowen
%A yu, Qun
%A Lin, Qin
%A Chen, Xu
%A Wang, Baoxun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%A Dengzhengwei
%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 wang-etal-2026-know-know
%X Role-playing conversational agents (RPCAs) built on large language models often leak out-of-boundary knowledge: when asked about topics outside a character’s limits, the model’s Chain-of-Thought reasoning reveals knowledge that the character should not possess, undermining immersion. To locate the source of this leakage, we conduct heuristic experiments at three levels of granularity, and find that intervening on individual Feed-Forward Network (FFN) neurons yields the strongest and most effective suppression. Motivated by this finding, we propose the Dynamic Attention Gating (DAG) Router, a lightweight trainable network that encodes the system prompt and dynamically generates per-layer FFN gating signals at inference time. Evaluated on multiple models and benchmarks, the DAG Router improves out-of-boundary scores by up to +162% compared with the original baseline, along with better overall role-playing quality.
%U https://aclanthology.org/2026.emnlp-main.545/
%P 12530-12548
Markdown (Informal)
[Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing](https://aclanthology.org/2026.emnlp-main.545/) (Wang et al., EMNLP 2026)
ACL