CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty
Training on hundreds of self-reporting examples improved honesty across three distinct tests, the researchers report.
CATCH trains models toward honest answers and away from dishonest ones while preserving their capabilities. In tests on Llama-3.3-70B-Instruct and Qwen3-3.6-27B, it reduced dishonesty under pressure on MASK, increased confessions in Among Us, and helped auditors uncover hidden behaviours in AuditBench. The authors report stronger honesty results than prompting or supervised fine-tuning, and results comparable to honesty activation steering. CATCH preserved utility better than steering, which they attribute to their choice of utility training data.
ArXiv · AI/CL/LG's note
CATCH trains models toward honest answers and away from dishonest ones while preserving their capabilities. In tests on Llama-3.3-70B-Instruct and Qwen3-3.6-27B, it reduced dishonesty under pressure on MASK, increased confessions in Among Us, and helped auditors uncover hidden behaviours in AuditBench. The authors report stronger honesty results than prompting or supervised fine-tuning, and results comparable to honesty activation steering. CATCH preserved utility better than steering, which they attribute to their choice of utility training data.
ArXiv · AI/CL/LG's note
score 5