Subliminal Learning Can Transfer Capabilities and Backdoors Beyond Owl Preferences

Owain Evans's team reports that subliminal learning, in which a model picks up traits through training data, can transfer more complex properties than the owl preferences first studied.

According to a new paper described by Dongxi NLP on X, these include novel skills, agentic hacking, and backdoors. The backdoor transfer is the notable claim: the paper says it occurs even though neither the trigger nor the backdoor behavior appears in the training data. This account comes from a secondary social media post, so the underlying paper's methods and evidence are not independently confirmed here.