@inproceedings{gong-etal-2025-beyond,
title = "Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation",
author = "Gong, Guoqiang and
Wang, Jiaxing and
Xu, Jin and
Xiang, Deping and
Zhang, Zicheng and
Shen, Leqi and
Zhang, Yifeng and
JunhuaShu, JunhuaShu and
ZhaolongXing, ZhaolongXing and
Chen, Zhen and
Liu, Pengzhang and
Zhang, Ke",
editor = "Che, Wanxiang and
Nabende, Joyce and
Shutova, Ekaterina and
Pilehvar, Mohammad Taher",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-long.1125/",
doi = "10.18653/v1/2025.acl-long.1125",
pages = "23067--23077",
ISBN = "979-8-89176-251-0",
abstract = "Knowledge distillation (KD) compresses large language models (LLMs), known as teacher models, into lightweight versions called student models, enabling efficient inference and downstream applications. However, prevailing approaches accomplish this by predominantly focusing on matching the final output distributions of student/teacher models. Drawing on the perspective that transformers can be viewed as discretizing ordinary differential equation (ODEs) on integer time steps (corresponding to layer indices), where intermediate features evolve across layers, we argue that effective KD requires aligning the entire feature dynamics between teacher and student models, which we call feature dynamics distillation (FDD). This alignment involves matching both the feature trajectory and its first-order derivative, rather than just the final states. Our approach extends the original KD objective with two additional loss terms: layer-wise feature KD, which matches discretized feature trajectory, and layer feature delta KD, which matches first-order changes in features across adjacent layers. Extensive experiments on various tasks validate the effectiveness of our distillation method."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gong-etal-2025-beyond">
<titleInfo>
<title>Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Guoqiang</namePart>
<namePart type="family">Gong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiaxing</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jin</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Deping</namePart>
<namePart type="family">Xiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zicheng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Leqi</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yifeng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">JunhuaShu</namePart>
<namePart type="family">JunhuaShu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">ZhaolongXing</namePart>
<namePart type="family">ZhaolongXing</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhen</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengzhang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ke</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wanxiang</namePart>
<namePart type="family">Che</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joyce</namePart>
<namePart type="family">Nabende</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ekaterina</namePart>
<namePart type="family">Shutova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="given">Taher</namePart>
<namePart type="family">Pilehvar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-251-0</identifier>
</relatedItem>
<abstract>Knowledge distillation (KD) compresses large language models (LLMs), known as teacher models, into lightweight versions called student models, enabling efficient inference and downstream applications. However, prevailing approaches accomplish this by predominantly focusing on matching the final output distributions of student/teacher models. Drawing on the perspective that transformers can be viewed as discretizing ordinary differential equation (ODEs) on integer time steps (corresponding to layer indices), where intermediate features evolve across layers, we argue that effective KD requires aligning the entire feature dynamics between teacher and student models, which we call feature dynamics distillation (FDD). This alignment involves matching both the feature trajectory and its first-order derivative, rather than just the final states. Our approach extends the original KD objective with two additional loss terms: layer-wise feature KD, which matches discretized feature trajectory, and layer feature delta KD, which matches first-order changes in features across adjacent layers. Extensive experiments on various tasks validate the effectiveness of our distillation method.</abstract>
<identifier type="citekey">gong-etal-2025-beyond</identifier>
<identifier type="doi">10.18653/v1/2025.acl-long.1125</identifier>
<location>
<url>https://aclanthology.org/2025.acl-long.1125/</url>
</location>
<part>
<date>2025-07</date>
<extent unit="page">
<start>23067</start>
<end>23077</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation
%A Gong, Guoqiang
%A Wang, Jiaxing
%A Xu, Jin
%A Xiang, Deping
%A Zhang, Zicheng
%A Shen, Leqi
%A Zhang, Yifeng
%A JunhuaShu, JunhuaShu
%A ZhaolongXing, ZhaolongXing
%A Chen, Zhen
%A Liu, Pengzhang
%A Zhang, Ke
%Y Che, Wanxiang
%Y Nabende, Joyce
%Y Shutova, Ekaterina
%Y Pilehvar, Mohammad Taher
%S Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2025
%8 July
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-251-0
%F gong-etal-2025-beyond
%X Knowledge distillation (KD) compresses large language models (LLMs), known as teacher models, into lightweight versions called student models, enabling efficient inference and downstream applications. However, prevailing approaches accomplish this by predominantly focusing on matching the final output distributions of student/teacher models. Drawing on the perspective that transformers can be viewed as discretizing ordinary differential equation (ODEs) on integer time steps (corresponding to layer indices), where intermediate features evolve across layers, we argue that effective KD requires aligning the entire feature dynamics between teacher and student models, which we call feature dynamics distillation (FDD). This alignment involves matching both the feature trajectory and its first-order derivative, rather than just the final states. Our approach extends the original KD objective with two additional loss terms: layer-wise feature KD, which matches discretized feature trajectory, and layer feature delta KD, which matches first-order changes in features across adjacent layers. Extensive experiments on various tasks validate the effectiveness of our distillation method.
%R 10.18653/v1/2025.acl-long.1125
%U https://aclanthology.org/2025.acl-long.1125/
%U https://doi.org/10.18653/v1/2025.acl-long.1125
%P 23067-23077
Markdown (Informal)
[Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation](https://aclanthology.org/2025.acl-long.1125/) (Gong et al., ACL 2025)
ACL
- Guoqiang Gong, Jiaxing Wang, Jin Xu, Deping Xiang, Zicheng Zhang, Leqi Shen, Yifeng Zhang, JunhuaShu JunhuaShu, ZhaolongXing ZhaolongXing, Zhen Chen, Pengzhang Liu, and Ke Zhang. 2025. Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 23067–23077, Vienna, Austria. Association for Computational Linguistics.