<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-harve">
    <titleInfo>
        <title>HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shuang</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yuxuan</namePart>
        <namePart type="family">Bo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Qiuyang</namePart>
        <namePart type="family">Zhao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Caiyue</namePart>
        <namePart type="family">Huang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xiaorong</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yanguang</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Mengnan</namePart>
        <namePart type="family">Du</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>Reward models are central to large language model (LLM) alignment, but they remain vulnerable to reward hacking. To evaluate reward-model robustness, we introduce RewardHackBench containing 13 reward-hacking patterns covering real life high-stakes domains and general settings, and we find severe failures on specific subcategories across eight reward models. To mitigate these failures, we propose HARVE, a training-free reward-head editing method for scalar reward models. Instead of fine-tuning the reward model, HARVE identifies a multi-directional hacking subspace from residual-stream directions associated with selected hacking subcategories, and removes the component of the reward-head vector aligned with that subspace. This directly reduces the reward head’s sensitivity to hacking-related features using only a small set of contrastive gold-hacked examples, without gradient updates or fine-tuning. Comprehensive experiments across eight reward models indicates that improves hacking robustness, outperforms fine-tuning baselines, and preserves reward-models’ general capability. Further analyses show that combining multiple target-specific residual-stream directions is more effective than relying on a single surface-level feature.</abstract>
    <identifier type="citekey">liu-etal-2026-harve</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1095/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>25127</start>
            <end>25150</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models
%A Liu, Shuang
%A Bo, Yuxuan
%A Zhao, Qiuyang
%A Huang, Caiyue
%A Chen, Xiaorong
%A Liu, Yanguang
%A Du, Mengnan
%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 liu-etal-2026-harve
%X Reward models are central to large language model (LLM) alignment, but they remain vulnerable to reward hacking. To evaluate reward-model robustness, we introduce RewardHackBench containing 13 reward-hacking patterns covering real life high-stakes domains and general settings, and we find severe failures on specific subcategories across eight reward models. To mitigate these failures, we propose HARVE, a training-free reward-head editing method for scalar reward models. Instead of fine-tuning the reward model, HARVE identifies a multi-directional hacking subspace from residual-stream directions associated with selected hacking subcategories, and removes the component of the reward-head vector aligned with that subspace. This directly reduces the reward head’s sensitivity to hacking-related features using only a small set of contrastive gold-hacked examples, without gradient updates or fine-tuning. Comprehensive experiments across eight reward models indicates that improves hacking robustness, outperforms fine-tuning baselines, and preserves reward-models’ general capability. Further analyses show that combining multiple target-specific residual-stream directions is more effective than relying on a single surface-level feature.
%U https://aclanthology.org/2026.emnlp-main.1095/
%P 25127-25150

Markdown (Informal)

[HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models](https://aclanthology.org/2026.emnlp-main.1095/) (Liu et al., EMNLP 2026)

• HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models (Liu et al., EMNLP 2026)

ACL

• Shuang Liu, Yuxuan Bo, Qiuyang Zhao, Caiyue Huang, Xiaorong Chen, Yanguang Liu, and Mengnan Du. 2026. HARVE: Hacking-Aware Reward-Head Vector Editing for Robust Reward Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 25127–25150, Budapest, Hungary. Association for Computational Linguistics.