@inproceedings{li-etal-2026-sparse,
title = "Sparse Model Diffing via Dynamic Circuits",
author = "Li, Yueyan and
Gao, Wenhao and
Yuan, Caixia and
Wang, Xiaojie",
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.679/",
doi = "10.18653/2026.findings-emnlp.679",
pages = "14254--14276",
ISBN = "979-8-89176-462-0",
abstract = "Model diffing studies the internal changes of a model from training to reveal learning mechanisms. To enhance interpretability, existing studies often strive for structural sparsity by removing redundant model components. However, we emphasize that for most fine-tuning methods, the fundamental source of internal change, i.e., the parameter updates, is non-sparse. The dense update mechanism contradicts the structural sparsity sought by researchers. Thus, we seek to bring sparsity to fine-tuning via an interpretability-driven method called circuit-tuning. It employs a heuristic prune-then-tune paradigm, forcing the model to optimize within a restricted subset of parameters. To test our method, we designed a novel task called Subject-Verb Disagreement (SVD) and conducted comprehensive explorations on it. Experimental results demonstrate that, compared to other fine-tuning methods, circuit-tuning achieves better sparsity in parameter updates, thus making it possible for faithful model diffing. Based on this, we performed stage-wise model diffing, revealing the evolution of model capabilities during fine-tuning from a new view. Furthermore, experiments on gender debiasing and larger models verify the practical utility of our approach. Our work provides a new methodology for model diffing-based interpretability and its applications."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-sparse">
<titleInfo>
<title>Sparse Model Diffing via Dynamic Circuits</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yueyan</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenhao</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Caixia</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaojie</namePart>
<namePart type="family">Wang</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>Model diffing studies the internal changes of a model from training to reveal learning mechanisms. To enhance interpretability, existing studies often strive for structural sparsity by removing redundant model components. However, we emphasize that for most fine-tuning methods, the fundamental source of internal change, i.e., the parameter updates, is non-sparse. The dense update mechanism contradicts the structural sparsity sought by researchers. Thus, we seek to bring sparsity to fine-tuning via an interpretability-driven method called circuit-tuning. It employs a heuristic prune-then-tune paradigm, forcing the model to optimize within a restricted subset of parameters. To test our method, we designed a novel task called Subject-Verb Disagreement (SVD) and conducted comprehensive explorations on it. Experimental results demonstrate that, compared to other fine-tuning methods, circuit-tuning achieves better sparsity in parameter updates, thus making it possible for faithful model diffing. Based on this, we performed stage-wise model diffing, revealing the evolution of model capabilities during fine-tuning from a new view. Furthermore, experiments on gender debiasing and larger models verify the practical utility of our approach. Our work provides a new methodology for model diffing-based interpretability and its applications.</abstract>
<identifier type="citekey">li-etal-2026-sparse</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.679</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.679/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14254</start>
<end>14276</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Sparse Model Diffing via Dynamic Circuits
%A Li, Yueyan
%A Gao, Wenhao
%A Yuan, Caixia
%A Wang, Xiaojie
%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 li-etal-2026-sparse
%X Model diffing studies the internal changes of a model from training to reveal learning mechanisms. To enhance interpretability, existing studies often strive for structural sparsity by removing redundant model components. However, we emphasize that for most fine-tuning methods, the fundamental source of internal change, i.e., the parameter updates, is non-sparse. The dense update mechanism contradicts the structural sparsity sought by researchers. Thus, we seek to bring sparsity to fine-tuning via an interpretability-driven method called circuit-tuning. It employs a heuristic prune-then-tune paradigm, forcing the model to optimize within a restricted subset of parameters. To test our method, we designed a novel task called Subject-Verb Disagreement (SVD) and conducted comprehensive explorations on it. Experimental results demonstrate that, compared to other fine-tuning methods, circuit-tuning achieves better sparsity in parameter updates, thus making it possible for faithful model diffing. Based on this, we performed stage-wise model diffing, revealing the evolution of model capabilities during fine-tuning from a new view. Furthermore, experiments on gender debiasing and larger models verify the practical utility of our approach. Our work provides a new methodology for model diffing-based interpretability and its applications.
%R 10.18653/2026.findings-emnlp.679
%U https://aclanthology.org/2026.findings-emnlp.679/
%U https://doi.org/10.18653/2026.findings-emnlp.679
%P 14254-14276
Markdown (Informal)
[Sparse Model Diffing via Dynamic Circuits](https://aclanthology.org/2026.findings-emnlp.679/) (Li et al., Findings 2026)
ACL
- Yueyan Li, Wenhao Gao, Caixia Yuan, and Xiaojie Wang. 2026. Sparse Model Diffing via Dynamic Circuits. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 14254–14276, Budapest, Hungary. Association for Computational Linguistics.