@inproceedings{quoc-etal-2026-steering,
title = "Steering from Random Quotients: A Sample-Complexity Theory of Polysemantic Contrast Pairs",
author = "Quoc, Khanh Nguyen and
Doan, Luong and
Nguyen, Duc Hai and
Hung, Mai Phan Quoc and
Vu, Ngoc Mai and
Ho, Phong and
Duong, Nhung and
Do, Tuan",
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.2190/",
pages = "49800--49813",
ISBN = "979-8-89176-459-0",
abstract = "Activation steering often averages activation differences between positive and negative prompts to construct a steering vector. This difference-in-means (DiM) estimator is simple, but a contrast pair can coactivate the intended concept together with topic, style, length, refusal, or truthfulness-related nuisance features. We model each pair as revealing the target vector through a quotient that removes its active nuisance span. The resulting polysemantic information operator determines which target components are identifiable and the number of contrast pairs needed to recover them. We establish identifiability, sample-complexity bounds, a coactivation phase diagram, and sparse-dictionary recovery guarantees. Synthetic and controlled experiments support the predicted quotient-recovery behavior. Public-data and Pythia-410M pilots instead evaluate recovery of the quotient estimand induced by chosen nuisance supports, not an externally validated semantic target. A controlled TruthfulQA counter-confounder illustrates the main NLP lesson: raw behavioral margin can reward a style-entangled DiM vector even when quotienting removes the injected style component."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="quoc-etal-2026-steering">
<titleInfo>
<title>Steering from Random Quotients: A Sample-Complexity Theory of Polysemantic Contrast Pairs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Khanh</namePart>
<namePart type="given">Nguyen</namePart>
<namePart type="family">Quoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Luong</namePart>
<namePart type="family">Doan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Duc</namePart>
<namePart type="given">Hai</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mai</namePart>
<namePart type="given">Phan</namePart>
<namePart type="given">Quoc</namePart>
<namePart type="family">Hung</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ngoc</namePart>
<namePart type="given">Mai</namePart>
<namePart type="family">Vu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Phong</namePart>
<namePart type="family">Ho</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nhung</namePart>
<namePart type="family">Duong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tuan</namePart>
<namePart type="family">Do</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>Activation steering often averages activation differences between positive and negative prompts to construct a steering vector. This difference-in-means (DiM) estimator is simple, but a contrast pair can coactivate the intended concept together with topic, style, length, refusal, or truthfulness-related nuisance features. We model each pair as revealing the target vector through a quotient that removes its active nuisance span. The resulting polysemantic information operator determines which target components are identifiable and the number of contrast pairs needed to recover them. We establish identifiability, sample-complexity bounds, a coactivation phase diagram, and sparse-dictionary recovery guarantees. Synthetic and controlled experiments support the predicted quotient-recovery behavior. Public-data and Pythia-410M pilots instead evaluate recovery of the quotient estimand induced by chosen nuisance supports, not an externally validated semantic target. A controlled TruthfulQA counter-confounder illustrates the main NLP lesson: raw behavioral margin can reward a style-entangled DiM vector even when quotienting removes the injected style component.</abstract>
<identifier type="citekey">quoc-etal-2026-steering</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2190/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>49800</start>
<end>49813</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Steering from Random Quotients: A Sample-Complexity Theory of Polysemantic Contrast Pairs
%A Quoc, Khanh Nguyen
%A Doan, Luong
%A Nguyen, Duc Hai
%A Hung, Mai Phan Quoc
%A Vu, Ngoc Mai
%A Ho, Phong
%A Duong, Nhung
%A Do, Tuan
%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 quoc-etal-2026-steering
%X Activation steering often averages activation differences between positive and negative prompts to construct a steering vector. This difference-in-means (DiM) estimator is simple, but a contrast pair can coactivate the intended concept together with topic, style, length, refusal, or truthfulness-related nuisance features. We model each pair as revealing the target vector through a quotient that removes its active nuisance span. The resulting polysemantic information operator determines which target components are identifiable and the number of contrast pairs needed to recover them. We establish identifiability, sample-complexity bounds, a coactivation phase diagram, and sparse-dictionary recovery guarantees. Synthetic and controlled experiments support the predicted quotient-recovery behavior. Public-data and Pythia-410M pilots instead evaluate recovery of the quotient estimand induced by chosen nuisance supports, not an externally validated semantic target. A controlled TruthfulQA counter-confounder illustrates the main NLP lesson: raw behavioral margin can reward a style-entangled DiM vector even when quotienting removes the injected style component.
%U https://aclanthology.org/2026.emnlp-main.2190/
%P 49800-49813
Markdown (Informal)
[Steering from Random Quotients: A Sample-Complexity Theory of Polysemantic Contrast Pairs](https://aclanthology.org/2026.emnlp-main.2190/) (Quoc et al., EMNLP 2026)
ACL
- Khanh Nguyen Quoc, Luong Doan, Duc Hai Nguyen, Mai Phan Quoc Hung, Ngoc Mai Vu, Phong Ho, Nhung Duong, and Tuan Do. 2026. Steering from Random Quotients: A Sample-Complexity Theory of Polysemantic Contrast Pairs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 49800–49813, Budapest, Hungary. Association for Computational Linguistics.