@inproceedings{maiti-etal-2026-souper,
title = "Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art {LLM} Performance",
author = "Maiti, Shalini and
Budhiraja, Amar and
Gauri, Bhavul and
Chaurasia, Gaurav and
Protopopov, Anton and
Audran-Reiss, Alexis and
Slater, Michael and
Magka, Despoina and
Shavrina, Tatiana and
Raileanu, Roberta and
Bachrach, Yoram",
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.33/",
pages = "645--666",
ISBN = "979-8-89176-462-0",
abstract = "Large Language Models (LLMs) have displayed remarkable capabilities across diverse domains, but their training remains resource {--} and time-intensive, requiring massive computational resources and careful orchestration of training procedures. Model souping {--} the practice of averaging weights from multiple models of the same architecture {--} has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. We observe that previous souping approaches can lead to collapse in precision-sensitive LLM capabilities. In this paper, we introduce Soup of Category Experts (SoCE), a principled approach for model souping to overcome this shortcoming. The proposed method utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous approaches, our method leverages the observation that different clusters (or categories) of points within a benchmark often exhibit low inter-correlations in model performance. SoCE identifies ``expert'' models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that SoCE improves performance and robustness across multiple domains achieves state-of-the-art results on the Berkeley Function Calling Leaderboard."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="maiti-etal-2026-souper">
<titleInfo>
<title>Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shalini</namePart>
<namePart type="family">Maiti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amar</namePart>
<namePart type="family">Budhiraja</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhavul</namePart>
<namePart type="family">Gauri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gaurav</namePart>
<namePart type="family">Chaurasia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anton</namePart>
<namePart type="family">Protopopov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexis</namePart>
<namePart type="family">Audran-Reiss</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Slater</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Despoina</namePart>
<namePart type="family">Magka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tatiana</namePart>
<namePart type="family">Shavrina</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Roberta</namePart>
<namePart type="family">Raileanu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yoram</namePart>
<namePart type="family">Bachrach</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>Large Language Models (LLMs) have displayed remarkable capabilities across diverse domains, but their training remains resource – and time-intensive, requiring massive computational resources and careful orchestration of training procedures. Model souping – the practice of averaging weights from multiple models of the same architecture – has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. We observe that previous souping approaches can lead to collapse in precision-sensitive LLM capabilities. In this paper, we introduce Soup of Category Experts (SoCE), a principled approach for model souping to overcome this shortcoming. The proposed method utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous approaches, our method leverages the observation that different clusters (or categories) of points within a benchmark often exhibit low inter-correlations in model performance. SoCE identifies “expert” models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that SoCE improves performance and robustness across multiple domains achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.</abstract>
<identifier type="citekey">maiti-etal-2026-souper</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.33/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>645</start>
<end>666</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance
%A Maiti, Shalini
%A Budhiraja, Amar
%A Gauri, Bhavul
%A Chaurasia, Gaurav
%A Protopopov, Anton
%A Audran-Reiss, Alexis
%A Slater, Michael
%A Magka, Despoina
%A Shavrina, Tatiana
%A Raileanu, Roberta
%A Bachrach, Yoram
%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 maiti-etal-2026-souper
%X Large Language Models (LLMs) have displayed remarkable capabilities across diverse domains, but their training remains resource – and time-intensive, requiring massive computational resources and careful orchestration of training procedures. Model souping – the practice of averaging weights from multiple models of the same architecture – has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. We observe that previous souping approaches can lead to collapse in precision-sensitive LLM capabilities. In this paper, we introduce Soup of Category Experts (SoCE), a principled approach for model souping to overcome this shortcoming. The proposed method utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous approaches, our method leverages the observation that different clusters (or categories) of points within a benchmark often exhibit low inter-correlations in model performance. SoCE identifies “expert” models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that SoCE improves performance and robustness across multiple domains achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.
%U https://aclanthology.org/2026.findings-emnlp.33/
%P 645-666
Markdown (Informal)
[Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance](https://aclanthology.org/2026.findings-emnlp.33/) (Maiti et al., Findings 2026)
ACL
- Shalini Maiti, Amar Budhiraja, Bhavul Gauri, Gaurav Chaurasia, Anton Protopopov, Alexis Audran-Reiss, Michael Slater, Despoina Magka, Tatiana Shavrina, Roberta Raileanu, and Yoram Bachrach. 2026. Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 645–666, Budapest, Hungary. Association for Computational Linguistics.