@inproceedings{sinii-etal-2025-steering,
title = "Steering {LLM} Reasoning Through Bias-Only Adaptation",
author = "Sinii, Viacheslav and
Gorbatovski, Alexey and
Cherepanov, Artem and
Shaposhnikov, Boris and
Balagansky, Nikita and
Gavrilov, Daniil",
editor = "Christodoulopoulos, Christos and
Chakraborty, Tanmoy and
Rose, Carolyn and
Peng, Violet",
booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2025",
address = "Suzhou, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.emnlp-main.467/",
doi = "10.18653/v1/2025.emnlp-main.467",
pages = "9202--9211",
ISBN = "979-8-89176-332-6",
abstract = "We show that training a single $d$-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks.On an 8 billion-parameter model this adds only $\approx 0.0016\%$ additional parameters and reproduces performance across a range of base models and mathematical-reasoning benchmarks.These results tighten the upper bound on the parameter budget required for high-level chain-of-thought reasoning, indicating that millions of adapter weights are unnecessary.The minimal trainable footprint reduces optimizer memory and inter-GPU communication, lowering the overall cost of fine-tuning.Moreover, a logit-lens analysis shows that the learned vectors amplify coherent token directions, providing clearer insight into the model{'}s internal computations."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sinii-etal-2025-steering">
<titleInfo>
<title>Steering LLM Reasoning Through Bias-Only Adaptation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Viacheslav</namePart>
<namePart type="family">Sinii</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexey</namePart>
<namePart type="family">Gorbatovski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Artem</namePart>
<namePart type="family">Cherepanov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Boris</namePart>
<namePart type="family">Shaposhnikov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikita</namePart>
<namePart type="family">Balagansky</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daniil</namePart>
<namePart type="family">Gavrilov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christos</namePart>
<namePart type="family">Christodoulopoulos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tanmoy</namePart>
<namePart type="family">Chakraborty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Carolyn</namePart>
<namePart type="family">Rose</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Violet</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Suzhou, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-332-6</identifier>
</relatedItem>
<abstract>We show that training a single d-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks.On an 8 billion-parameter model this adds only \approx 0.0016% additional parameters and reproduces performance across a range of base models and mathematical-reasoning benchmarks.These results tighten the upper bound on the parameter budget required for high-level chain-of-thought reasoning, indicating that millions of adapter weights are unnecessary.The minimal trainable footprint reduces optimizer memory and inter-GPU communication, lowering the overall cost of fine-tuning.Moreover, a logit-lens analysis shows that the learned vectors amplify coherent token directions, providing clearer insight into the model’s internal computations.</abstract>
<identifier type="citekey">sinii-etal-2025-steering</identifier>
<identifier type="doi">10.18653/v1/2025.emnlp-main.467</identifier>
<location>
<url>https://aclanthology.org/2025.emnlp-main.467/</url>
</location>
<part>
<date>2025-11</date>
<extent unit="page">
<start>9202</start>
<end>9211</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Steering LLM Reasoning Through Bias-Only Adaptation
%A Sinii, Viacheslav
%A Gorbatovski, Alexey
%A Cherepanov, Artem
%A Shaposhnikov, Boris
%A Balagansky, Nikita
%A Gavrilov, Daniil
%Y Christodoulopoulos, Christos
%Y Chakraborty, Tanmoy
%Y Rose, Carolyn
%Y Peng, Violet
%S Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing
%D 2025
%8 November
%I Association for Computational Linguistics
%C Suzhou, China
%@ 979-8-89176-332-6
%F sinii-etal-2025-steering
%X We show that training a single d-dimensional steering vector per layer with reinforcement learning, while freezing all base weights, matches the accuracy of fully RL-tuned reasoning models on mathematical-reasoning tasks.On an 8 billion-parameter model this adds only \approx 0.0016% additional parameters and reproduces performance across a range of base models and mathematical-reasoning benchmarks.These results tighten the upper bound on the parameter budget required for high-level chain-of-thought reasoning, indicating that millions of adapter weights are unnecessary.The minimal trainable footprint reduces optimizer memory and inter-GPU communication, lowering the overall cost of fine-tuning.Moreover, a logit-lens analysis shows that the learned vectors amplify coherent token directions, providing clearer insight into the model’s internal computations.
%R 10.18653/v1/2025.emnlp-main.467
%U https://aclanthology.org/2025.emnlp-main.467/
%U https://doi.org/10.18653/v1/2025.emnlp-main.467
%P 9202-9211
Markdown (Informal)
[Steering LLM Reasoning Through Bias-Only Adaptation](https://aclanthology.org/2025.emnlp-main.467/) (Sinii et al., EMNLP 2025)
ACL
- Viacheslav Sinii, Alexey Gorbatovski, Artem Cherepanov, Boris Shaposhnikov, Nikita Balagansky, and Daniil Gavrilov. 2025. Steering LLM Reasoning Through Bias-Only Adaptation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 9202–9211, Suzhou, China. Association for Computational Linguistics.