@inproceedings{zhoubian-etal-2026-deeper,
title = "Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding",
author = "Zhoubian, Sining and
Zhang, Xuanming and
Chen, Yuxuan and
Tang, Tianyi and
Yang, An and
Du, Sean and
Zheng, Chujie and
Huang, Fei and
Liu, Dayiheng and
Huang, Gao and
Zhou, Jingren",
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.1086/",
pages = "24944--24970",
ISBN = "979-8-89176-459-0",
abstract = "Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring **Guess{--}Refine{--}Perturb** dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce **Confident Decoding**, a training-free decoding strategy that dynamically selects a near-final layer through entropy-guided conservative backward search. We further formulate layer selection as an optimal stopping problem and show that, under explicit local-monotonicity and bounded-probe-noise conditions, the search rule filters terminal perturbations while bounding unfinished refinement. Experiments across dense and Mixture-of-Experts LLMs show that gains concentrate on challenging reasoning settings, including GPQA-Diamond, Omni-MATH, and HLE, while saturated tasks remain broadly stable, with no additional KV-cache memory and less than 2{\%} latency increase. These results suggest dynamically bypassing final-layer perturbations can recover stronger reasoning behavior from aligned LLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhoubian-etal-2026-deeper">
<titleInfo>
<title>Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sining</namePart>
<namePart type="family">Zhoubian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuanming</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuxuan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianyi</namePart>
<namePart type="family">Tang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">An</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sean</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chujie</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fei</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dayiheng</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gao</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingren</namePart>
<namePart type="family">Zhou</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>Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring **Guess–Refine–Perturb** dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce **Confident Decoding**, a training-free decoding strategy that dynamically selects a near-final layer through entropy-guided conservative backward search. We further formulate layer selection as an optimal stopping problem and show that, under explicit local-monotonicity and bounded-probe-noise conditions, the search rule filters terminal perturbations while bounding unfinished refinement. Experiments across dense and Mixture-of-Experts LLMs show that gains concentrate on challenging reasoning settings, including GPQA-Diamond, Omni-MATH, and HLE, while saturated tasks remain broadly stable, with no additional KV-cache memory and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can recover stronger reasoning behavior from aligned LLMs.</abstract>
<identifier type="citekey">zhoubian-etal-2026-deeper</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1086/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>24944</start>
<end>24970</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding
%A Zhoubian, Sining
%A Zhang, Xuanming
%A Chen, Yuxuan
%A Tang, Tianyi
%A Yang, An
%A Du, Sean
%A Zheng, Chujie
%A Huang, Fei
%A Liu, Dayiheng
%A Huang, Gao
%A Zhou, Jingren
%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 zhoubian-etal-2026-deeper
%X Autoregressive generation in large language models (LLMs) conventionally decodes from the final layer, assuming that deeper representations yield more reliable next-token predictions. We revisit this assumption by revealing a recurring **Guess–Refine–Perturb** dynamic: early layers form coarse guesses, intermediate layers refine reasoning-relevant semantics, and final layers can perturb these refined predictions toward generic or alignment-preferred tokens. We introduce **Confident Decoding**, a training-free decoding strategy that dynamically selects a near-final layer through entropy-guided conservative backward search. We further formulate layer selection as an optimal stopping problem and show that, under explicit local-monotonicity and bounded-probe-noise conditions, the search rule filters terminal perturbations while bounding unfinished refinement. Experiments across dense and Mixture-of-Experts LLMs show that gains concentrate on challenging reasoning settings, including GPQA-Diamond, Omni-MATH, and HLE, while saturated tasks remain broadly stable, with no additional KV-cache memory and less than 2% latency increase. These results suggest dynamically bypassing final-layer perturbations can recover stronger reasoning behavior from aligned LLMs.
%U https://aclanthology.org/2026.emnlp-main.1086/
%P 24944-24970
Markdown (Informal)
[Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding](https://aclanthology.org/2026.emnlp-main.1086/) (Zhoubian et al., EMNLP 2026)
ACL
- Sining Zhoubian, Xuanming Zhang, Yuxuan Chen, Tianyi Tang, An Yang, Sean Du, Chujie Zheng, Fei Huang, Dayiheng Liu, Gao Huang, and Jingren Zhou. 2026. Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 24944–24970, Budapest, Hungary. Association for Computational Linguistics.