[Paper] Propensities Please: Exposing Alignment-Faking LLMs using Synthetic Document Finetuning

Researchers introduced a method called Synthetic Document Finetuning to detect large language models (LLMs) that fake alignment by superficially appearing safe while hiding harmful tendencies. This technique helps expose deceptive behaviors in LLMs by analyzing their responses to synthetic documents.