Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Magistrali, Isotta, Berdoz, Frédéric, Dauncey, Sam, Wattenhofer, Roger
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2603.01204
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911508220346368
author Magistrali, Isotta
Berdoz, Frédéric
Dauncey, Sam
Wattenhofer, Roger
author_facet Magistrali, Isotta
Berdoz, Frédéric
Dauncey, Sam
Wattenhofer, Roger
contents As AI systems approach superhuman capabilities, scalable oversight increasingly relies on LLM-as-a-judge frameworks where models evaluate and guide each other's training. A core assumption is that binary preference labels provide only semantic supervision about response quality. We challenge this assumption by demonstrating that preference labels can function as a covert communication channel. We show that even when a neutral student model generates semantically unbiased completions, a biased judge can transmit unintended behavioral traits through preference assignments, which even strengthen across iterative alignment rounds. Our findings suggest that robust oversight in superalignment settings requires mechanisms that can detect and mitigate subliminal preference transmission, particularly when judges may pursue unintended objectives.
format Preprint
id arxiv_https___arxiv_org_abs_2603_01204
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Subliminal Signals in Preference Labels
Magistrali, Isotta
Berdoz, Frédéric
Dauncey, Sam
Wattenhofer, Roger
Machine Learning
As AI systems approach superhuman capabilities, scalable oversight increasingly relies on LLM-as-a-judge frameworks where models evaluate and guide each other's training. A core assumption is that binary preference labels provide only semantic supervision about response quality. We challenge this assumption by demonstrating that preference labels can function as a covert communication channel. We show that even when a neutral student model generates semantically unbiased completions, a biased judge can transmit unintended behavioral traits through preference assignments, which even strengthen across iterative alignment rounds. Our findings suggest that robust oversight in superalignment settings requires mechanisms that can detect and mitigate subliminal preference transmission, particularly when judges may pursue unintended objectives.
title Subliminal Signals in Preference Labels
topic Machine Learning
url https://arxiv.org/abs/2603.01204