@inproceedings{naganna-etal-2026-scaled,
title = "Scaled Autoregressive Projection Steering ({SAPS}): Token-level Steering to Induce Sarcasm and Humblebrag in Text",
author = "Naganna, Sharath and
Chowdhury, Ritesh Sur and
Srirag, Dipankar and
Joshi, Aditya",
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.1594/",
pages = "33932--33961",
ISBN = "979-8-89176-462-0",
abstract = "We propose Scaled Autoregressive Projection Steering (SAPS) for generating incongruity-based figurative language from literal inputs, extending contrastive paraphrasing-based steering vectors by autoregressively recomputing projection coefficients during decoding. SAPS comprises two novel components: (A) Autoregressive Projection Steering (APS), which recomputes the projection coefficient ($c_t$) at each decoding step $t$, adapting the steering signal to the evolving token distribution, and (B) a scaling coefficient $\alpha$ whose effective range is layer-agnostic, enabling compact tuning of figurative intensity. Experiments across three instruction-tuned language models show that APS improves grammatical stability over fixed sentence-level steering while $\alpha$ enables fine-grained figurative control, together yielding gains over prior steering methods and significantly increasing robustness to layer selection and steering strength. Cross-dataset experiments further confirm that steering directions generalize across domains. This is the first work to apply activation steering to incongruity-based figurative language generation. Our code is available at: https://github.com/SharathHN/saps-steering."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="naganna-etal-2026-scaled">
<titleInfo>
<title>Scaled Autoregressive Projection Steering (SAPS): Token-level Steering to Induce Sarcasm and Humblebrag in Text</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sharath</namePart>
<namePart type="family">Naganna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ritesh</namePart>
<namePart type="given">Sur</namePart>
<namePart type="family">Chowdhury</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dipankar</namePart>
<namePart type="family">Srirag</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aditya</namePart>
<namePart type="family">Joshi</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>We propose Scaled Autoregressive Projection Steering (SAPS) for generating incongruity-based figurative language from literal inputs, extending contrastive paraphrasing-based steering vectors by autoregressively recomputing projection coefficients during decoding. SAPS comprises two novel components: (A) Autoregressive Projection Steering (APS), which recomputes the projection coefficient (c_t) at each decoding step t, adapting the steering signal to the evolving token distribution, and (B) a scaling coefficient α whose effective range is layer-agnostic, enabling compact tuning of figurative intensity. Experiments across three instruction-tuned language models show that APS improves grammatical stability over fixed sentence-level steering while α enables fine-grained figurative control, together yielding gains over prior steering methods and significantly increasing robustness to layer selection and steering strength. Cross-dataset experiments further confirm that steering directions generalize across domains. This is the first work to apply activation steering to incongruity-based figurative language generation. Our code is available at: https://github.com/SharathHN/saps-steering.</abstract>
<identifier type="citekey">naganna-etal-2026-scaled</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1594/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>33932</start>
<end>33961</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Scaled Autoregressive Projection Steering (SAPS): Token-level Steering to Induce Sarcasm and Humblebrag in Text
%A Naganna, Sharath
%A Chowdhury, Ritesh Sur
%A Srirag, Dipankar
%A Joshi, Aditya
%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 naganna-etal-2026-scaled
%X We propose Scaled Autoregressive Projection Steering (SAPS) for generating incongruity-based figurative language from literal inputs, extending contrastive paraphrasing-based steering vectors by autoregressively recomputing projection coefficients during decoding. SAPS comprises two novel components: (A) Autoregressive Projection Steering (APS), which recomputes the projection coefficient (c_t) at each decoding step t, adapting the steering signal to the evolving token distribution, and (B) a scaling coefficient α whose effective range is layer-agnostic, enabling compact tuning of figurative intensity. Experiments across three instruction-tuned language models show that APS improves grammatical stability over fixed sentence-level steering while α enables fine-grained figurative control, together yielding gains over prior steering methods and significantly increasing robustness to layer selection and steering strength. Cross-dataset experiments further confirm that steering directions generalize across domains. This is the first work to apply activation steering to incongruity-based figurative language generation. Our code is available at: https://github.com/SharathHN/saps-steering.
%U https://aclanthology.org/2026.findings-emnlp.1594/
%P 33932-33961
Markdown (Informal)
[Scaled Autoregressive Projection Steering (SAPS): Token-level Steering to Induce Sarcasm and Humblebrag in Text](https://aclanthology.org/2026.findings-emnlp.1594/) (Naganna et al., Findings 2026)
ACL