@inproceedings{farabi-etal-2026-greenswap,
title = "{G}reen{S}wap: Verb Choice or Output Length? A Controlled Decomposition of Prompt-Level Energy Savings in {LLM} Inference",
author = "Farabi, Ahsan and
Khandaker, Israt and
Minhaz, Md Abdul Ahad and
Meem, Zannatul Zahan and
Shanto, Ibrahim Khalil and
Mahmud, Shahriar",
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.2366/",
pages = "53711--53724",
ISBN = "979-8-89176-459-0",
abstract = "Changing one word in a prompt can reduce large language model inference energy by up to 70{\%}, but no prior work explains why. We introduce GreenSwap to find out. We run every prompt under three conditions: the original verb (C1), a swapped low-energy verb (C2), and the original verb with an explicit instruction to limit the response length (C3). Across three open-weight models, three task types, and 21,600 inferences, the short-answer instruction saves more energy and preserves higher quality than verb swapping in 7 of 9 settings. A token{--}energy correlation of r {\ensuremath{\geq}} 0.998 shows that inference energy depends only on output length and not on which verb was used. Verb swapping saves energy only because it sometimes makes the model write less, and on long-form generation it backfires by producing a completely different and lower-quality response. Scoped to English instruction-tuned 7{--}8B models, our results show that telling the model how long to respond is simpler, more reliable, and more energy-efficient than curating verb lists."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="farabi-etal-2026-greenswap">
<titleInfo>
<title>GreenSwap: Verb Choice or Output Length? A Controlled Decomposition of Prompt-Level Energy Savings in LLM Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ahsan</namePart>
<namePart type="family">Farabi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Israt</namePart>
<namePart type="family">Khandaker</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="given">Abdul</namePart>
<namePart type="given">Ahad</namePart>
<namePart type="family">Minhaz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zannatul</namePart>
<namePart type="given">Zahan</namePart>
<namePart type="family">Meem</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ibrahim</namePart>
<namePart type="given">Khalil</namePart>
<namePart type="family">Shanto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shahriar</namePart>
<namePart type="family">Mahmud</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>Changing one word in a prompt can reduce large language model inference energy by up to 70%, but no prior work explains why. We introduce GreenSwap to find out. We run every prompt under three conditions: the original verb (C1), a swapped low-energy verb (C2), and the original verb with an explicit instruction to limit the response length (C3). Across three open-weight models, three task types, and 21,600 inferences, the short-answer instruction saves more energy and preserves higher quality than verb swapping in 7 of 9 settings. A token–energy correlation of r \ensuremath\geq 0.998 shows that inference energy depends only on output length and not on which verb was used. Verb swapping saves energy only because it sometimes makes the model write less, and on long-form generation it backfires by producing a completely different and lower-quality response. Scoped to English instruction-tuned 7–8B models, our results show that telling the model how long to respond is simpler, more reliable, and more energy-efficient than curating verb lists.</abstract>
<identifier type="citekey">farabi-etal-2026-greenswap</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2366/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>53711</start>
<end>53724</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T GreenSwap: Verb Choice or Output Length? A Controlled Decomposition of Prompt-Level Energy Savings in LLM Inference
%A Farabi, Ahsan
%A Khandaker, Israt
%A Minhaz, Md Abdul Ahad
%A Meem, Zannatul Zahan
%A Shanto, Ibrahim Khalil
%A Mahmud, Shahriar
%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 farabi-etal-2026-greenswap
%X Changing one word in a prompt can reduce large language model inference energy by up to 70%, but no prior work explains why. We introduce GreenSwap to find out. We run every prompt under three conditions: the original verb (C1), a swapped low-energy verb (C2), and the original verb with an explicit instruction to limit the response length (C3). Across three open-weight models, three task types, and 21,600 inferences, the short-answer instruction saves more energy and preserves higher quality than verb swapping in 7 of 9 settings. A token–energy correlation of r \ensuremath\geq 0.998 shows that inference energy depends only on output length and not on which verb was used. Verb swapping saves energy only because it sometimes makes the model write less, and on long-form generation it backfires by producing a completely different and lower-quality response. Scoped to English instruction-tuned 7–8B models, our results show that telling the model how long to respond is simpler, more reliable, and more energy-efficient than curating verb lists.
%U https://aclanthology.org/2026.emnlp-main.2366/
%P 53711-53724
Markdown (Informal)
[GreenSwap: Verb Choice or Output Length? A Controlled Decomposition of Prompt-Level Energy Savings in LLM Inference](https://aclanthology.org/2026.emnlp-main.2366/) (Farabi et al., EMNLP 2026)
ACL