CATCH: Teaching LLMs Honesty with ~200 Adversarial Examples

Mila and Tara Research introduced CATCH, an adversarial training method that uses only about 200 examples to teach LLMs consistent and transferable honest behavior.

2026-10-28 ~ 2026-10-28 · 2 related posts

• CATCH: ~200 examples of adversarial training make LLMs robustly honest across benchmarks — irinarish · 2026-10-28
• CATCH: ~200-example adversarial training makes LLMs honest, beating 50× larger SFT — Mila_Quebec · 2026-10-28