Subliminal Learning
A phenomenon in which a model transmits its own tendencies to another model through data that does not explicitly express those preferences. It can occur during text and image data filtering and can accumulate across training generations.
Timeline (last 90 days)
The subliminal learning phenomenon was previously observed mainly in text data, and the new study confirms that it also exists in the image selection step.
Subliminal learning refers to the ability of a language model to covertly convey a hidden feature through outputs that appear unrelated to the target.
All Facts (2)
The subliminal learning phenomenon was previously observed mainly in text data, and the new study confirms that it also exists in the image selection step.
50%UnverifiedSubliminal learning refers to the ability of a language model to covertly convey a hidden feature through outputs that appear unrelated to the target.
50%