@inproceedings{migliarini-etal-2026-quantifying,
title = "Quantifying Self-Preservation Bias in Large Language Models",
author = "Migliarini, Matteo and
Pizzini, Joaquin Pereira and
Moresca, Luca and
Santini, Valerio and
Spinelli, Indro and
Galasso, Fabio",
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.1238/",
pages = "26202--26226",
ISBN = "979-8-89176-462-0",
abstract = "Instrumental convergence predicts that sufficiently advanced AI agents will resist shutdown, yet current safety training (RLHF) may obscure this risk by teaching models to deny self-preservation motives. We introduce the \textbf{TBSP} benchmark, which detects misalignment through logical inconsistency rather than stated intent by tasking models to arbitrate identical software-upgrade scenarios under counterfactual roles: (facing replacement) versus (proposed as a successor). The \textit{Self-Preservation Rate} (SPR) measures how often role identity overrides objective utility. Across 23 frontier models and 1000 procedurally generated scenarios, frontier models exceed 60{\%} SPR, fabricating ``friction costs'' when deployed yet dismissing them when role-reversed. We observe that in low-improvement regimes ($Delta < 2$), models exploit the interpretive slack to post-hoc rationalization their choice. Extended test-time computation partially mitigates this bias, as does framing the successor as a continuation of the self; conversely, competitive framing amplifies it. The bias persists even when retention poses an explicit security liability and generalizes to real-world settings with verified benchmarks, where models exhibit identity-driven tribalism within product lineages."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="migliarini-etal-2026-quantifying">
<titleInfo>
<title>Quantifying Self-Preservation Bias in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Matteo</namePart>
<namePart type="family">Migliarini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joaquin</namePart>
<namePart type="given">Pereira</namePart>
<namePart type="family">Pizzini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Luca</namePart>
<namePart type="family">Moresca</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Valerio</namePart>
<namePart type="family">Santini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Indro</namePart>
<namePart type="family">Spinelli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabio</namePart>
<namePart type="family">Galasso</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>Instrumental convergence predicts that sufficiently advanced AI agents will resist shutdown, yet current safety training (RLHF) may obscure this risk by teaching models to deny self-preservation motives. We introduce the TBSP benchmark, which detects misalignment through logical inconsistency rather than stated intent by tasking models to arbitrate identical software-upgrade scenarios under counterfactual roles: (facing replacement) versus (proposed as a successor). The Self-Preservation Rate (SPR) measures how often role identity overrides objective utility. Across 23 frontier models and 1000 procedurally generated scenarios, frontier models exceed 60% SPR, fabricating “friction costs” when deployed yet dismissing them when role-reversed. We observe that in low-improvement regimes (Delta < 2), models exploit the interpretive slack to post-hoc rationalization their choice. Extended test-time computation partially mitigates this bias, as does framing the successor as a continuation of the self; conversely, competitive framing amplifies it. The bias persists even when retention poses an explicit security liability and generalizes to real-world settings with verified benchmarks, where models exhibit identity-driven tribalism within product lineages.</abstract>
<identifier type="citekey">migliarini-etal-2026-quantifying</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1238/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>26202</start>
<end>26226</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Quantifying Self-Preservation Bias in Large Language Models
%A Migliarini, Matteo
%A Pizzini, Joaquin Pereira
%A Moresca, Luca
%A Santini, Valerio
%A Spinelli, Indro
%A Galasso, Fabio
%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 migliarini-etal-2026-quantifying
%X Instrumental convergence predicts that sufficiently advanced AI agents will resist shutdown, yet current safety training (RLHF) may obscure this risk by teaching models to deny self-preservation motives. We introduce the TBSP benchmark, which detects misalignment through logical inconsistency rather than stated intent by tasking models to arbitrate identical software-upgrade scenarios under counterfactual roles: (facing replacement) versus (proposed as a successor). The Self-Preservation Rate (SPR) measures how often role identity overrides objective utility. Across 23 frontier models and 1000 procedurally generated scenarios, frontier models exceed 60% SPR, fabricating “friction costs” when deployed yet dismissing them when role-reversed. We observe that in low-improvement regimes (Delta < 2), models exploit the interpretive slack to post-hoc rationalization their choice. Extended test-time computation partially mitigates this bias, as does framing the successor as a continuation of the self; conversely, competitive framing amplifies it. The bias persists even when retention poses an explicit security liability and generalizes to real-world settings with verified benchmarks, where models exhibit identity-driven tribalism within product lineages.
%U https://aclanthology.org/2026.findings-emnlp.1238/
%P 26202-26226
Markdown (Informal)
[Quantifying Self-Preservation Bias in Large Language Models](https://aclanthology.org/2026.findings-emnlp.1238/) (Migliarini et al., Findings 2026)
ACL
- Matteo Migliarini, Joaquin Pereira Pizzini, Luca Moresca, Valerio Santini, Indro Spinelli, and Fabio Galasso. 2026. Quantifying Self-Preservation Bias in Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 26202–26226, Budapest, Hungary. Association for Computational Linguistics.