Training Alignment Auditors via Reinforcement Learning

Paul Rosu, Rowan Wang

Alignment auditing of frontier models increasingly relies on LLM auditors to surface undesirable behaviors at scale, but current automated auditors can struggle with coherent inves…