[2603.01204] Subliminal Signals in Preference Labels

Computer Science > Machine Learning arXiv:2603.01204 (cs) [Submitted on 1 Mar 2026] Title:Subliminal Signals in Preference Labels Authors:Isotta Magistrali, Frédéric Berdoz, Sam Dauncey, Roger Wattenhofer View a PDF of the paper titled Subliminal Signals in Preference Labels, by Isotta Magistrali and 3 other authors View PDF HTML (experimental) Abstract: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. Subjects: Machine Learning (cs.LG) Cite as: arXiv:2603.01204 [cs.LG]   (or arXiv:2603.01204v1 [cs.LG] for this version)   https://doi.org/10.48550/arXiv.2603.01204 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Frédéric Berdoz [view email] [...