@inproceedings{le-etal-2026-distribution,
title = "When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift",
author = "Le, Khoi M. and
Cao, Tri and
Nguyen, Phong Nam Huu and
Nguyen, Cong-Duy T and
Luu, Anh Tuan and
Miao, Chunyan and
Ng, See-Kiong and
Nguyen, Thong Thanh",
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.1240/",
pages = "28270--28288",
ISBN = "979-8-89176-459-0",
abstract = "Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence consistent with a representational failure mode in which weak-supervised fine-tuning may move the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose *Representation Anchoring* (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model{'}s representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="le-etal-2026-distribution">
<titleInfo>
<title>When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift</title>
</titleInfo>
<name type="personal">
<namePart type="given">Khoi</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tri</namePart>
<namePart type="family">Cao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Phong</namePart>
<namePart type="given">Nam</namePart>
<namePart type="given">Huu</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cong-Duy</namePart>
<namePart type="given">T</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anh</namePart>
<namePart type="given">Tuan</namePart>
<namePart type="family">Luu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chunyan</namePart>
<namePart type="family">Miao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">See-Kiong</namePart>
<namePart type="family">Ng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thong</namePart>
<namePart type="given">Thanh</namePart>
<namePart type="family">Nguyen</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>Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence consistent with a representational failure mode in which weak-supervised fine-tuning may move the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose *Representation Anchoring* (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model’s representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer.</abstract>
<identifier type="citekey">le-etal-2026-distribution</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1240/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>28270</start>
<end>28288</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift
%A Le, Khoi M.
%A Cao, Tri
%A Nguyen, Phong Nam Huu
%A Nguyen, Cong-Duy T.
%A Luu, Anh Tuan
%A Miao, Chunyan
%A Ng, See-Kiong
%A Nguyen, Thong Thanh
%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 le-etal-2026-distribution
%X Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence consistent with a representational failure mode in which weak-supervised fine-tuning may move the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose *Representation Anchoring* (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model’s representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer.
%U https://aclanthology.org/2026.emnlp-main.1240/
%P 28270-28288
Markdown (Informal)
[When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift](https://aclanthology.org/2026.emnlp-main.1240/) (Le et al., EMNLP 2026)
ACL
- Khoi M. Le, Tri Cao, Phong Nam Huu Nguyen, Cong-Duy T Nguyen, Anh Tuan Luu, Chunyan Miao, See-Kiong Ng, and Thong Thanh Nguyen. 2026. When In-Distribution Gains Fail: Evaluating Weak-to-Strong Reward Models under Preference Shift. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 28270–28288, Budapest, Hungary. Association for Computational Linguistics.