Preprint reports transferable gains from AI honesty training

Researchers report in a preprint that an adversarial training method reduced dishonest behavior across distinct benchmarks while preserving model utility. Called CATCH, the method trains models toward honest answers and away from dishonest ones using hundreds of targets drawn from self-reporting scenarios. Tests covered Llama-3.3-70B-Instruct and Qwen3-3.6-27B. The researchers report less dishonesty under pressure on MASK, more confessions in the multi-agent game Among Us, and better discovery of hidden model behaviors on AuditBench. CATCH outperformed prompting and supervised fine-tuning on honesty; it matched honesty activation steering on honesty while better maintaining utility.