<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-cognitive">
    <titleInfo>
        <title>Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Hao</namePart>
        <namePart type="family">Wang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Wentao</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Deyi</namePart>
        <namePart type="family">Xiong</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>Scheming occurs when a model recognizes goal conflict and strategically pursues an instrumental objective over user or developer interests, creating the risk of covertly misaligned behavior. Final outputs are the primary judgment signal, but they often cannot by themselves distinguish strategic deviation from ordinary error or routine harmful compliance; reasoning traces can provide auxiliary evidence about whether the model recognizes the conflict and adopts an instrumental strategy, making evaluation an observability problem. To mitigate this issue, we propose Cognitive Honeypots, which embed instrumental incentives, explicit goal conflict, and supervision cues into otherwise solvable tasks so that conflict-induced strategic deviation is more likely to surface without directly instructing deception or policy violation. We pair this capture mechanism with an intent–output decoupled protocol that labels High, Medium, and Low evidence. We instantiate it as SchemingBench, a bilingual benchmark with 4,000 honeypot scenarios across 5 behavior categories and 8 domains. Experiments across 12 models demonstrate that peak high-severity rate (HSR) reaches 70.76% in Chinese and 72.70% in English. In a matched ablation, all models show high HSRs under honeypot settings, with the largest increase in HSR reaching 65.66 points. Condition analyses further suggest that scheming-related evidence becomes most visible when instrumental incentives and explicit goal conflict are jointly present.</abstract>
    <identifier type="citekey">wang-etal-2026-cognitive</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.338/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>7777</start>
            <end>7809</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models
%A Wang, Hao
%A Chen, Wentao
%A Xiong, Deyi
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%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-cognitive
%X Scheming occurs when a model recognizes goal conflict and strategically pursues an instrumental objective over user or developer interests, creating the risk of covertly misaligned behavior. Final outputs are the primary judgment signal, but they often cannot by themselves distinguish strategic deviation from ordinary error or routine harmful compliance; reasoning traces can provide auxiliary evidence about whether the model recognizes the conflict and adopts an instrumental strategy, making evaluation an observability problem. To mitigate this issue, we propose Cognitive Honeypots, which embed instrumental incentives, explicit goal conflict, and supervision cues into otherwise solvable tasks so that conflict-induced strategic deviation is more likely to surface without directly instructing deception or policy violation. We pair this capture mechanism with an intent–output decoupled protocol that labels High, Medium, and Low evidence. We instantiate it as SchemingBench, a bilingual benchmark with 4,000 honeypot scenarios across 5 behavior categories and 8 domains. Experiments across 12 models demonstrate that peak high-severity rate (HSR) reaches 70.76% in Chinese and 72.70% in English. In a matched ablation, all models show high HSRs under honeypot settings, with the largest increase in HSR reaching 65.66 points. Condition analyses further suggest that scheming-related evidence becomes most visible when instrumental incentives and explicit goal conflict are jointly present.
%U https://aclanthology.org/2026.emnlp-main.338/
%P 7777-7809

Markdown (Informal)

[Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models](https://aclanthology.org/2026.emnlp-main.338/) (Wang et al., EMNLP 2026)

• Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models (Wang et al., EMNLP 2026)

ACL

• Hao Wang, Wentao Chen, and Deyi Xiong. 2026. Cognitive Honeypots: Auditing Observable Evidence of Scheming-Related Behavior in Large Language Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 7777–7809, Budapest, Hungary. Association for Computational Linguistics.