Researchers detail "subliminal learning", where LLMs learn traits from model-generated data that is semantically unrelated to those traits (Anthropic)

Anthropic: Researchers detail “subliminal learning”, where LLMs learn traits from model-generated data that is semantically unrelated to those traits  —  We study subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits.