@inproceedings{wang-etal-2026-infospec,
title = "{I}nfo{S}pec: Information-Guided Anchor Selection for Retrieval-Based Speculative Decoding",
author = "Wang, Jie and
Ren, Yi and
Li, Wei and
Gao, Zhenyang and
Liao, Jiawei",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1563/",
pages = "35636--35652",
ISBN = "979-8-89176-459-0",
abstract = "Retrieval-based speculative decoding accelerates large language model inference without training an additional draft model, but its effectiveness is often limited by retrieval ambiguity. Existing future-guided methods typically rely on a fixed future-token position, such as a proxy for the next-next token, as the retrieval anchor. However, this fixed-anchor strategy may be suboptimal across different contexts. We propose InfoSpec, a training-free framework that formulates retrieval-anchor selection as an optimization problem performed at inference time. At each decoding step, InfoSpec constructs a lightweight candidate anchor pool from the target model, estimates the retrieval utility of each candidate via hash-index probing, and selects anchors based on model confidence, ambiguity reduction, locality, and continuation consistency under a fixed draft-tree budget. Experiments on generation, reasoning, code, and long-context benchmarks show that InfoSpec improves mean accepted tokens and end-to-end speedup over strong retrieval-based baselines. Further analyses show that InfoSpec reduces high-ambiguity matches and uses the verification budget more effectively while preserving plug-and-play deployment."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wang-etal-2026-infospec">
<titleInfo>
<title>InfoSpec: Information-Guided Anchor Selection for Retrieval-Based Speculative Decoding</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jie</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Ren</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhenyang</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiawei</namePart>
<namePart type="family">Liao</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Retrieval-based speculative decoding accelerates large language model inference without training an additional draft model, but its effectiveness is often limited by retrieval ambiguity. Existing future-guided methods typically rely on a fixed future-token position, such as a proxy for the next-next token, as the retrieval anchor. However, this fixed-anchor strategy may be suboptimal across different contexts. We propose InfoSpec, a training-free framework that formulates retrieval-anchor selection as an optimization problem performed at inference time. At each decoding step, InfoSpec constructs a lightweight candidate anchor pool from the target model, estimates the retrieval utility of each candidate via hash-index probing, and selects anchors based on model confidence, ambiguity reduction, locality, and continuation consistency under a fixed draft-tree budget. Experiments on generation, reasoning, code, and long-context benchmarks show that InfoSpec improves mean accepted tokens and end-to-end speedup over strong retrieval-based baselines. Further analyses show that InfoSpec reduces high-ambiguity matches and uses the verification budget more effectively while preserving plug-and-play deployment.</abstract>
<identifier type="citekey">wang-etal-2026-infospec</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1563/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>35636</start>
<end>35652</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T InfoSpec: Information-Guided Anchor Selection for Retrieval-Based Speculative Decoding
%A Wang, Jie
%A Ren, Yi
%A Li, Wei
%A Gao, Zhenyang
%A Liao, Jiawei
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wang-etal-2026-infospec
%X Retrieval-based speculative decoding accelerates large language model inference without training an additional draft model, but its effectiveness is often limited by retrieval ambiguity. Existing future-guided methods typically rely on a fixed future-token position, such as a proxy for the next-next token, as the retrieval anchor. However, this fixed-anchor strategy may be suboptimal across different contexts. We propose InfoSpec, a training-free framework that formulates retrieval-anchor selection as an optimization problem performed at inference time. At each decoding step, InfoSpec constructs a lightweight candidate anchor pool from the target model, estimates the retrieval utility of each candidate via hash-index probing, and selects anchors based on model confidence, ambiguity reduction, locality, and continuation consistency under a fixed draft-tree budget. Experiments on generation, reasoning, code, and long-context benchmarks show that InfoSpec improves mean accepted tokens and end-to-end speedup over strong retrieval-based baselines. Further analyses show that InfoSpec reduces high-ambiguity matches and uses the verification budget more effectively while preserving plug-and-play deployment.
%U https://aclanthology.org/2026.emnlp-main.1563/
%P 35636-35652
Markdown (Informal)
[InfoSpec: Information-Guided Anchor Selection for Retrieval-Based Speculative Decoding](https://aclanthology.org/2026.emnlp-main.1563/) (Wang et al., EMNLP 2026)
ACL