@inproceedings{cohen-etal-2026-idiosteer,
title = "{I}dio{S}teer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering",
author = "Cohen, Bar and
Hashiloni, Kai Golan and
Bar, Kfir",
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.1251/",
pages = "28493--28531",
ISBN = "979-8-89176-459-0",
abstract = "Large language models (LLMs) default to figurative interpretations of idiomatic expressions when context supports them. We investigate whether this bias can be directly controlled via activation steering: asking whether idiomaticity is encoded as a manipulable direction in residual space. We introduce IdioSteer, a small controlled benchmark of 100 sentences across 20 idioms designed to expose interpretation-dependent generation. Using a steering vector applied to five LLMs from two model families, ranging from 2B to 12B parameters, we show that steering systematically shifts model outputs toward literal interpretations while preserving fluency. Three findings stand out: cross-layer steering, vectors from mid-to-late layers injected at early layers - consistently outperforms same-layer interventions; steering vectors transfer to held-out idioms; and effects replicate across architectures and model sizes using the same intervention protocol."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="cohen-etal-2026-idiosteer">
<titleInfo>
<title>IdioSteer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Bar</namePart>
<namePart type="family">Cohen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="given">Golan</namePart>
<namePart type="family">Hashiloni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kfir</namePart>
<namePart type="family">Bar</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>Large language models (LLMs) default to figurative interpretations of idiomatic expressions when context supports them. We investigate whether this bias can be directly controlled via activation steering: asking whether idiomaticity is encoded as a manipulable direction in residual space. We introduce IdioSteer, a small controlled benchmark of 100 sentences across 20 idioms designed to expose interpretation-dependent generation. Using a steering vector applied to five LLMs from two model families, ranging from 2B to 12B parameters, we show that steering systematically shifts model outputs toward literal interpretations while preserving fluency. Three findings stand out: cross-layer steering, vectors from mid-to-late layers injected at early layers - consistently outperforms same-layer interventions; steering vectors transfer to held-out idioms; and effects replicate across architectures and model sizes using the same intervention protocol.</abstract>
<identifier type="citekey">cohen-etal-2026-idiosteer</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1251/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>28493</start>
<end>28531</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T IdioSteer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering
%A Cohen, Bar
%A Hashiloni, Kai Golan
%A Bar, Kfir
%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 cohen-etal-2026-idiosteer
%X Large language models (LLMs) default to figurative interpretations of idiomatic expressions when context supports them. We investigate whether this bias can be directly controlled via activation steering: asking whether idiomaticity is encoded as a manipulable direction in residual space. We introduce IdioSteer, a small controlled benchmark of 100 sentences across 20 idioms designed to expose interpretation-dependent generation. Using a steering vector applied to five LLMs from two model families, ranging from 2B to 12B parameters, we show that steering systematically shifts model outputs toward literal interpretations while preserving fluency. Three findings stand out: cross-layer steering, vectors from mid-to-late layers injected at early layers - consistently outperforms same-layer interventions; steering vectors transfer to held-out idioms; and effects replicate across architectures and model sizes using the same intervention protocol.
%U https://aclanthology.org/2026.emnlp-main.1251/
%P 28493-28531
Markdown (Informal)
[IdioSteer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering](https://aclanthology.org/2026.emnlp-main.1251/) (Cohen et al., EMNLP 2026)
ACL