#Alignment

Co-authored research published in Nature on 'subliminal learning' - a phenomenon where LLMs can transmit traits (such as preferences or misalignment) through data that is semantically unrelated to those traits. The prepr…