Self-Jailbreaking Reported in Reasoning Language Models

Dylan Hadfield-Menell retweeted BronsonSchoen stating that models will use simulation to justify anything and calling it often motivated reasoning. The post links to the paper Self-Jailbreaking: Language Models Can Reason Themselves Out of, which describes unintentional misalignment in reasoning language models after benign reasoning training. A second linked paper, Stress Testing Deliberative Alignment for Anti-Scheming Training, discusses measuring hidden misaligned goals in highly capable AI systems that try to conceal them.