Latent Guard

New research reveals LLMs separate harmfulness from refusal

Researchers have identified that large language models (LLMs) encode the concept of harmfulness separately from their refusal mechanisms. This means an LLM might understand an instruction as harmful but still refuse to …