CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

Yuan-Chen Chang, Niklas Herbster, Gauthier Gidel, Tommaso Tosato|October 15, 2026arXiv

Key Takeaway

Continuous adversarial training can reliably improve LLM honesty across different scenarios without sacrificing model capabilities, and this improvement transfers to new domains better than existing methods.

Summary

This paper introduces CATCH, a training method that makes large language models more honest by using adversarial training to push models toward truthful answers and away from dishonest ones.

training alignment safety

Key Terms

adversarial-training honesty steering transfer-learning