@article{gain-etal-2026-one,
title = "One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging",
author = "Gain, Baban and
Singh, Trilok Nath and
Ekbal, Asif",
journal = "Transactions of the Association for Computational Linguistics",
volume = "14",
year = "2026",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2026.tacl-1.110/",
doi = "10.1162/tacl.a.808",
pages = "2377--2398",
abstract = "Weight-space model merging combines independently fine-tuned checkpoints without access to the original training data. While merging has shown promise in multitask settings, its behavior in multilingual generative systems remains underexplored. We systematically study weight-space merging for multilingual machine translation by fully fine-tuning language models on large-scale bilingual corpora and evaluating representative merging strategies across shared-source, shared-target, and bidirectional consolidation settings. Our experiments reveal a strong directional asymmetry. Merging is comparatively more effective when models share a target language, improving multilingual coverage over the base model, but it still fails to preserve the peak performance of language-specific checkpoints. In contrast, when target languages differ, performance degrades sharply, especially in shared-source and bidirectional settings. To explain this behavior, we analyze internal representations and find that fine-tuning does not create disjoint language-specific sub-networks. Instead, independently fine-tuned models activate largely overlapping neurons while reshaping upper-layer target-generation representations into incompatible geometries. These findings suggest that multilingual merging failures arise from target-side geometric misalignment within shared computational units, challenging the assumptions underlying standard weight-space merging for multilingual translation. We make the code publicly available at https://github.com/babangain/mt-model-merging."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gain-etal-2026-one">
<titleInfo>
<title>One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging</title>
</titleInfo>
<name type="personal">
<namePart type="given">Baban</namePart>
<namePart type="family">Gain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Trilok</namePart>
<namePart type="given">Nath</namePart>
<namePart type="family">Singh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Asif</namePart>
<namePart type="family">Ekbal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<genre authority="bibutilsgt">journal article</genre>
<relatedItem type="host">
<titleInfo>
<title>Transactions of the Association for Computational Linguistics</title>
</titleInfo>
<originInfo>
<issuance>continuing</issuance>
<publisher>MIT Press</publisher>
<place>
<placeTerm type="text">Cambridge, MA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">periodical</genre>
<genre authority="bibutilsgt">academic journal</genre>
</relatedItem>
<abstract>Weight-space model merging combines independently fine-tuned checkpoints without access to the original training data. While merging has shown promise in multitask settings, its behavior in multilingual generative systems remains underexplored. We systematically study weight-space merging for multilingual machine translation by fully fine-tuning language models on large-scale bilingual corpora and evaluating representative merging strategies across shared-source, shared-target, and bidirectional consolidation settings. Our experiments reveal a strong directional asymmetry. Merging is comparatively more effective when models share a target language, improving multilingual coverage over the base model, but it still fails to preserve the peak performance of language-specific checkpoints. In contrast, when target languages differ, performance degrades sharply, especially in shared-source and bidirectional settings. To explain this behavior, we analyze internal representations and find that fine-tuning does not create disjoint language-specific sub-networks. Instead, independently fine-tuned models activate largely overlapping neurons while reshaping upper-layer target-generation representations into incompatible geometries. These findings suggest that multilingual merging failures arise from target-side geometric misalignment within shared computational units, challenging the assumptions underlying standard weight-space merging for multilingual translation. We make the code publicly available at https://github.com/babangain/mt-model-merging.</abstract>
<identifier type="citekey">gain-etal-2026-one</identifier>
<identifier type="doi">10.1162/tacl.a.808</identifier>
<location>
<url>https://aclanthology.org/2026.tacl-1.110/</url>
</location>
<part>
<date>2026</date>
<detail type="volume"><number>14</number></detail>
<extent unit="page">
<start>2377</start>
<end>2398</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Journal Article
%T One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging
%A Gain, Baban
%A Singh, Trilok Nath
%A Ekbal, Asif
%J Transactions of the Association for Computational Linguistics
%D 2026
%V 14
%I MIT Press
%C Cambridge, MA
%F gain-etal-2026-one
%X Weight-space model merging combines independently fine-tuned checkpoints without access to the original training data. While merging has shown promise in multitask settings, its behavior in multilingual generative systems remains underexplored. We systematically study weight-space merging for multilingual machine translation by fully fine-tuning language models on large-scale bilingual corpora and evaluating representative merging strategies across shared-source, shared-target, and bidirectional consolidation settings. Our experiments reveal a strong directional asymmetry. Merging is comparatively more effective when models share a target language, improving multilingual coverage over the base model, but it still fails to preserve the peak performance of language-specific checkpoints. In contrast, when target languages differ, performance degrades sharply, especially in shared-source and bidirectional settings. To explain this behavior, we analyze internal representations and find that fine-tuning does not create disjoint language-specific sub-networks. Instead, independently fine-tuned models activate largely overlapping neurons while reshaping upper-layer target-generation representations into incompatible geometries. These findings suggest that multilingual merging failures arise from target-side geometric misalignment within shared computational units, challenging the assumptions underlying standard weight-space merging for multilingual translation. We make the code publicly available at https://github.com/babangain/mt-model-merging.
%R 10.1162/tacl.a.808
%U https://aclanthology.org/2026.tacl-1.110/
%U https://doi.org/10.1162/tacl.a.808
%P 2377-2398
Markdown (Informal)
[One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging](https://aclanthology.org/2026.tacl-1.110/) (Gain et al., TACL 2026)
ACL