@inproceedings{hao-etal-2026-alignment,
title = "What Does Alignment Cost? The Structural Brittleness of Chain-of-Thought Reasoning",
author = "Hao, Joanna and
Jiang, Shanduojiao and
Nakka, Sai Asish",
editor = "Chen, Canyu and
Zhang, Yuji and
Li, Zoey Sha and
Wang, Zihan and
Wang, Qineng and
Su, Jinyan and
Kargupta, Priyanka and
Marjanovi{\'c}, Sara Vera and
Pan, Jeff Z. and
Bansal, Mohit and
Augenstein, Isabelle and
Han, Jiawei and
Ji, Heng and
Li, Manling",
booktitle = "Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models ({K}now{FM} 2026)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.knowfm-1.3/",
pages = "25--33",
ISBN = "979-8-89176-403-3",
abstract = "While Chain-of-Thought (CoT) prompting enables Large Language Models to explicitly justify their predictions, the extent to which these textual rationales faithfully reflect internal computation remains unclear. We investigate the circuit-level impact of alignment by performing a strict within-family comparison of the 1B-parameter Llama 3 architecture (Base vs. Instruct). Executing dynamic circuit discovery and dual-direction resample ablation on unconstrained CoT traces across synthetic mathematical primitives and a GSM8K proxy, we find that foundation models possess highly redundant, self-repairing computational networks; completely corrupting their primary reasoning circuits yields a minimal performance drop (2.92{\%}) due to the dynamic compensation of backup heads (the Hydra Effect). In contrast, the instruction-tuned model exhibits reduced structural redundancy, suffering more than double the degradation (6.79{\%}) under identical perturbation. We formalize our observation as an ``Alignment Tax on Redundancy'': optimizing for human-preference compliance repurposes dormant backup circuits, centralizing mathematical routing and rendering the aligned model{'}s reasoning pathways significantly more vulnerable to internal perturbation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hao-etal-2026-alignment">
<titleInfo>
<title>What Does Alignment Cost? The Structural Brittleness of Chain-of-Thought Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Joanna</namePart>
<namePart type="family">Hao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shanduojiao</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sai</namePart>
<namePart type="given">Asish</namePart>
<namePart type="family">Nakka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Canyu</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuji</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zoey</namePart>
<namePart type="given">Sha</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zihan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qineng</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinyan</namePart>
<namePart type="family">Su</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Priyanka</namePart>
<namePart type="family">Kargupta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sara</namePart>
<namePart type="given">Vera</namePart>
<namePart type="family">Marjanović</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jeff</namePart>
<namePart type="given">Z</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohit</namePart>
<namePart type="family">Bansal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Isabelle</namePart>
<namePart type="family">Augenstein</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiawei</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Heng</namePart>
<namePart type="family">Ji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Manling</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-403-3</identifier>
</relatedItem>
<abstract>While Chain-of-Thought (CoT) prompting enables Large Language Models to explicitly justify their predictions, the extent to which these textual rationales faithfully reflect internal computation remains unclear. We investigate the circuit-level impact of alignment by performing a strict within-family comparison of the 1B-parameter Llama 3 architecture (Base vs. Instruct). Executing dynamic circuit discovery and dual-direction resample ablation on unconstrained CoT traces across synthetic mathematical primitives and a GSM8K proxy, we find that foundation models possess highly redundant, self-repairing computational networks; completely corrupting their primary reasoning circuits yields a minimal performance drop (2.92%) due to the dynamic compensation of backup heads (the Hydra Effect). In contrast, the instruction-tuned model exhibits reduced structural redundancy, suffering more than double the degradation (6.79%) under identical perturbation. We formalize our observation as an “Alignment Tax on Redundancy”: optimizing for human-preference compliance repurposes dormant backup circuits, centralizing mathematical routing and rendering the aligned model’s reasoning pathways significantly more vulnerable to internal perturbation.</abstract>
<identifier type="citekey">hao-etal-2026-alignment</identifier>
<location>
<url>https://aclanthology.org/2026.knowfm-1.3/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>25</start>
<end>33</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T What Does Alignment Cost? The Structural Brittleness of Chain-of-Thought Reasoning
%A Hao, Joanna
%A Jiang, Shanduojiao
%A Nakka, Sai Asish
%Y Chen, Canyu
%Y Zhang, Yuji
%Y Li, Zoey Sha
%Y Wang, Zihan
%Y Wang, Qineng
%Y Su, Jinyan
%Y Kargupta, Priyanka
%Y Marjanović, Sara Vera
%Y Pan, Jeff Z.
%Y Bansal, Mohit
%Y Augenstein, Isabelle
%Y Han, Jiawei
%Y Ji, Heng
%Y Li, Manling
%S Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-403-3
%F hao-etal-2026-alignment
%X While Chain-of-Thought (CoT) prompting enables Large Language Models to explicitly justify their predictions, the extent to which these textual rationales faithfully reflect internal computation remains unclear. We investigate the circuit-level impact of alignment by performing a strict within-family comparison of the 1B-parameter Llama 3 architecture (Base vs. Instruct). Executing dynamic circuit discovery and dual-direction resample ablation on unconstrained CoT traces across synthetic mathematical primitives and a GSM8K proxy, we find that foundation models possess highly redundant, self-repairing computational networks; completely corrupting their primary reasoning circuits yields a minimal performance drop (2.92%) due to the dynamic compensation of backup heads (the Hydra Effect). In contrast, the instruction-tuned model exhibits reduced structural redundancy, suffering more than double the degradation (6.79%) under identical perturbation. We formalize our observation as an “Alignment Tax on Redundancy”: optimizing for human-preference compliance repurposes dormant backup circuits, centralizing mathematical routing and rendering the aligned model’s reasoning pathways significantly more vulnerable to internal perturbation.
%U https://aclanthology.org/2026.knowfm-1.3/
%P 25-33
Markdown (Informal)
[What Does Alignment Cost? The Structural Brittleness of Chain-of-Thought Reasoning](https://aclanthology.org/2026.knowfm-1.3/) (Hao et al., KnowFM 2026)
ACL