@inproceedings{duszenko-etal-2026-sycophantic,
title = "Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models",
author = "Duszenko, Jacek and
Kazienko, Przemyslaw and
Kocon, Jan",
editor = "T.Y.S.S., Santosh and
Rodriguez, Juan Diego and
de Gibert, Ona",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics ({ACL} 2026)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-srw.20/",
pages = "225--239",
ISBN = "979-8-89176-393-7",
abstract = "Reasoning models frequently agree with incorrect user suggestions - a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. We introduce sycophantic anchors - sentences identified via counterfactual analysis that commit models to user agreement. Across four reasoning models spanning three architecture families (Llama, Qwen, Falcon-hybrid) and 1.5B - 8B parameters, we analyze over 200,000 counterfactual rollouts and show that linear probes reliably detect sycophantic anchors (74 - 85{\%} balanced accuracy), outperforming text-only baselines at high commitment levels -confirming they capture internal states beyond surface vocabulary. Regressors further predict commitment strength from activations ($R^2$ up to 0.74). We observe a consistent asymmetry: sycophancy leaves a stronger mechanistic footprint than correct reasoning. We also find that sycophancy builds gradually during generation rather than being determined by the prompt. These findings enable sentence-level detection and quantification of model misalignment mid-inference."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="duszenko-etal-2026-sycophantic">
<titleInfo>
<title>Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jacek</namePart>
<namePart type="family">Duszenko</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Przemyslaw</namePart>
<namePart type="family">Kazienko</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jan</namePart>
<namePart type="family">Kocon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Santosh</namePart>
<namePart type="family">T.Y.S.S.</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Juan</namePart>
<namePart type="given">Diego</namePart>
<namePart type="family">Rodriguez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ona</namePart>
<namePart type="family">de Gibert</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-393-7</identifier>
</relatedItem>
<abstract>Reasoning models frequently agree with incorrect user suggestions - a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. We introduce sycophantic anchors - sentences identified via counterfactual analysis that commit models to user agreement. Across four reasoning models spanning three architecture families (Llama, Qwen, Falcon-hybrid) and 1.5B - 8B parameters, we analyze over 200,000 counterfactual rollouts and show that linear probes reliably detect sycophantic anchors (74 - 85% balanced accuracy), outperforming text-only baselines at high commitment levels -confirming they capture internal states beyond surface vocabulary. Regressors further predict commitment strength from activations (R² up to 0.74). We observe a consistent asymmetry: sycophancy leaves a stronger mechanistic footprint than correct reasoning. We also find that sycophancy builds gradually during generation rather than being determined by the prompt. These findings enable sentence-level detection and quantification of model misalignment mid-inference.</abstract>
<identifier type="citekey">duszenko-etal-2026-sycophantic</identifier>
<location>
<url>https://aclanthology.org/2026.acl-srw.20/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>225</start>
<end>239</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models
%A Duszenko, Jacek
%A Kazienko, Przemyslaw
%A Kocon, Jan
%Y T.Y.S.S., Santosh
%Y Rodriguez, Juan Diego
%Y de Gibert, Ona
%S Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-393-7
%F duszenko-etal-2026-sycophantic
%X Reasoning models frequently agree with incorrect user suggestions - a behavior known as sycophancy. However, it is unclear where in the reasoning trace this agreement originates and how strong the commitment is. We introduce sycophantic anchors - sentences identified via counterfactual analysis that commit models to user agreement. Across four reasoning models spanning three architecture families (Llama, Qwen, Falcon-hybrid) and 1.5B - 8B parameters, we analyze over 200,000 counterfactual rollouts and show that linear probes reliably detect sycophantic anchors (74 - 85% balanced accuracy), outperforming text-only baselines at high commitment levels -confirming they capture internal states beyond surface vocabulary. Regressors further predict commitment strength from activations (R² up to 0.74). We observe a consistent asymmetry: sycophancy leaves a stronger mechanistic footprint than correct reasoning. We also find that sycophancy builds gradually during generation rather than being determined by the prompt. These findings enable sentence-level detection and quantification of model misalignment mid-inference.
%U https://aclanthology.org/2026.acl-srw.20/
%P 225-239
Markdown (Informal)
[Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models](https://aclanthology.org/2026.acl-srw.20/) (Duszenko et al., ACL 2026)
ACL