@inproceedings{zhang-etal-2026-dartersd,
title = "{D}arter{SD}: Dynamic Token-Weighted Distillation for High-Efficient Speculative Decoding",
author = "Zhang, Zhaoning and
Baizhou, Xu and
Zhang, Libo and
Chen, Yi and
Tian, Zhiliang and
Huang, Zhen and
Li, Dongsheng",
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.749/",
doi = "10.18653/2026.findings-emnlp.749",
pages = "15773--15788",
ISBN = "979-8-89176-462-0",
abstract = "Speculative decoding accelerates autoregressive LLM generation with a lightweight draft model, yet existing distillation methods treat all historical tokens equally despite their divergent contributions to next-token prediction. We discover that attention in draft models is extremely concentrated: the top-3 historical tokens capture 80{\%} of attention weight, while over 90{\%} of tokens contribute negligibly. This uniform training paradigm forces the draft model to learn from noisy, low-importance tokens, diluting the learning signal from informative positions and limiting the alignment between draft and target distributions. To address this, we propose DarterSD, a dynamic token-weighted distillation framework for efficient speculative decoding. DarterSD employs a learnable sigmoid gate in self-attention to compute per-token importance scores, and designs a multi-objective distillation loss comprising distribution alignment, logit matching, and gate regularization to focus training on high-informative tokens. We further introduce a lightweight Boolean-mask denoising strategy at inference to suppress contextual noise, without modifying the target model{'}s verification pipeline. Experiments show DarterSD delivers a lossless speedup up to 5.52x , outperforming EAGLE3 (4.67x), and boosts the average accepted token length by 12.9{\%}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-dartersd">
<titleInfo>
<title>DarterSD: Dynamic Token-Weighted Distillation for High-Efficient Speculative Decoding</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhaoning</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xu</namePart>
<namePart type="family">Baizhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Libo</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhiliang</namePart>
<namePart type="family">Tian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhen</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dongsheng</namePart>
<namePart type="family">Li</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>Speculative decoding accelerates autoregressive LLM generation with a lightweight draft model, yet existing distillation methods treat all historical tokens equally despite their divergent contributions to next-token prediction. We discover that attention in draft models is extremely concentrated: the top-3 historical tokens capture 80% of attention weight, while over 90% of tokens contribute negligibly. This uniform training paradigm forces the draft model to learn from noisy, low-importance tokens, diluting the learning signal from informative positions and limiting the alignment between draft and target distributions. To address this, we propose DarterSD, a dynamic token-weighted distillation framework for efficient speculative decoding. DarterSD employs a learnable sigmoid gate in self-attention to compute per-token importance scores, and designs a multi-objective distillation loss comprising distribution alignment, logit matching, and gate regularization to focus training on high-informative tokens. We further introduce a lightweight Boolean-mask denoising strategy at inference to suppress contextual noise, without modifying the target model’s verification pipeline. Experiments show DarterSD delivers a lossless speedup up to 5.52x , outperforming EAGLE3 (4.67x), and boosts the average accepted token length by 12.9%.</abstract>
<identifier type="citekey">zhang-etal-2026-dartersd</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.749</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.749/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>15773</start>
<end>15788</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DarterSD: Dynamic Token-Weighted Distillation for High-Efficient Speculative Decoding
%A Zhang, Zhaoning
%A Baizhou, Xu
%A Zhang, Libo
%A Chen, Yi
%A Tian, Zhiliang
%A Huang, Zhen
%A Li, Dongsheng
%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 zhang-etal-2026-dartersd
%X Speculative decoding accelerates autoregressive LLM generation with a lightweight draft model, yet existing distillation methods treat all historical tokens equally despite their divergent contributions to next-token prediction. We discover that attention in draft models is extremely concentrated: the top-3 historical tokens capture 80% of attention weight, while over 90% of tokens contribute negligibly. This uniform training paradigm forces the draft model to learn from noisy, low-importance tokens, diluting the learning signal from informative positions and limiting the alignment between draft and target distributions. To address this, we propose DarterSD, a dynamic token-weighted distillation framework for efficient speculative decoding. DarterSD employs a learnable sigmoid gate in self-attention to compute per-token importance scores, and designs a multi-objective distillation loss comprising distribution alignment, logit matching, and gate regularization to focus training on high-informative tokens. We further introduce a lightweight Boolean-mask denoising strategy at inference to suppress contextual noise, without modifying the target model’s verification pipeline. Experiments show DarterSD delivers a lossless speedup up to 5.52x , outperforming EAGLE3 (4.67x), and boosts the average accepted token length by 12.9%.
%R 10.18653/2026.findings-emnlp.749
%U https://aclanthology.org/2026.findings-emnlp.749/
%U https://doi.org/10.18653/2026.findings-emnlp.749
%P 15773-15788
Markdown (Informal)
[DarterSD: Dynamic Token-Weighted Distillation for High-Efficient Speculative Decoding](https://aclanthology.org/2026.findings-emnlp.749/) (Zhang et al., Findings 2026)
ACL