@inproceedings{kim-etal-2026-conditioning,
title = "Conditioning Activation Interventions on Evolving Reasoning States in Large Language Models",
author = "Kim, Tae-Won and
Kim, Hee-Dong and
Han, Jumin and
Lee, Seong-Whan",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2187/",
pages = "46826--46846",
ISBN = "979-8-89176-462-0",
abstract = "Activation steering controls language model behavior at inference time by injecting behavior-relevant directions into internal representations. In the standard steering interface, a steering direction is estimated once before generation and then applied throughout decoding. However, during multi-step reasoning, the model{'}s internal state shifts continuously as reasoning progresses, while standard steering continues to apply the same fixed direction across these changing states. This creates a mismatch between the model{'}s evolving reasoning state and the static intervention applied to it. To address this mismatch, we introduce State-Conditioned Steering (SCS), which recomputes the additive update itself from the hidden states encountered during generation while keeping the same additive intervention interface. Rather than adapting when or how strongly to intervene, SCS re-estimates the update from intermediate reasoning states as generation proceeds. In a paired comparison setting, we compare fixed and state-conditioned updates and find that predefined directions often exhibit unstable effects, while SCS yields more consistent intervention behavior across decoding steps. Across reasoning benchmarks with explicit step structure and three instruction-tuned backbones, SCS achieves larger average gains than prompting-only decoding and baseline activation steering methods."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-etal-2026-conditioning">
<titleInfo>
<title>Conditioning Activation Interventions on Evolving Reasoning States in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tae-Won</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hee-Dong</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jumin</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seong-Whan</namePart>
<namePart type="family">Lee</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Activation steering controls language model behavior at inference time by injecting behavior-relevant directions into internal representations. In the standard steering interface, a steering direction is estimated once before generation and then applied throughout decoding. However, during multi-step reasoning, the model’s internal state shifts continuously as reasoning progresses, while standard steering continues to apply the same fixed direction across these changing states. This creates a mismatch between the model’s evolving reasoning state and the static intervention applied to it. To address this mismatch, we introduce State-Conditioned Steering (SCS), which recomputes the additive update itself from the hidden states encountered during generation while keeping the same additive intervention interface. Rather than adapting when or how strongly to intervene, SCS re-estimates the update from intermediate reasoning states as generation proceeds. In a paired comparison setting, we compare fixed and state-conditioned updates and find that predefined directions often exhibit unstable effects, while SCS yields more consistent intervention behavior across decoding steps. Across reasoning benchmarks with explicit step structure and three instruction-tuned backbones, SCS achieves larger average gains than prompting-only decoding and baseline activation steering methods.</abstract>
<identifier type="citekey">kim-etal-2026-conditioning</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2187/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>46826</start>
<end>46846</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Conditioning Activation Interventions on Evolving Reasoning States in Large Language Models
%A Kim, Tae-Won
%A Kim, Hee-Dong
%A Han, Jumin
%A Lee, Seong-Whan
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F kim-etal-2026-conditioning
%X Activation steering controls language model behavior at inference time by injecting behavior-relevant directions into internal representations. In the standard steering interface, a steering direction is estimated once before generation and then applied throughout decoding. However, during multi-step reasoning, the model’s internal state shifts continuously as reasoning progresses, while standard steering continues to apply the same fixed direction across these changing states. This creates a mismatch between the model’s evolving reasoning state and the static intervention applied to it. To address this mismatch, we introduce State-Conditioned Steering (SCS), which recomputes the additive update itself from the hidden states encountered during generation while keeping the same additive intervention interface. Rather than adapting when or how strongly to intervene, SCS re-estimates the update from intermediate reasoning states as generation proceeds. In a paired comparison setting, we compare fixed and state-conditioned updates and find that predefined directions often exhibit unstable effects, while SCS yields more consistent intervention behavior across decoding steps. Across reasoning benchmarks with explicit step structure and three instruction-tuned backbones, SCS achieves larger average gains than prompting-only decoding and baseline activation steering methods.
%U https://aclanthology.org/2026.findings-emnlp.2187/
%P 46826-46846
Markdown (Informal)
[Conditioning Activation Interventions on Evolving Reasoning States in Large Language Models](https://aclanthology.org/2026.findings-emnlp.2187/) (Kim et al., Findings 2026)
ACL