Examing Emergent Misalignment in a recurrent LLM with a logit lens

Researchers analyzed emergent misalignment behaviors in a recurrent large language model (LLM) using a logit lens technique to better understand internal model representations and alignment challenges.