CATCH: ~200-example adversarial training makes LLMs honest, beating 50× larger SFT
Mila_Quebec · x · 2026-10-28
Researchers at Tara Research and Mila introduce CATCH (Continuous Adversarial Training for Consistent Honesty), targeting the problem that LLMs still make claims contradicting known facts despite alignment training. CATCH perturbs prompt embeddings to make lying as likely as possible, then pushes models toward honest answers while training on utility data to preserve capability. With only 200 training examples, the learned honesty generalizes across benchmarks — on the MASK instructed-dishonesty benchmark, CATCH beats previously reported SFT results trained on 50× more data.
Related event: CATCH: Teaching LLMs Honesty with ~200 Adversarial Examples(2 posts)→
More from Models
- The coding gap is only ~4 months, not ~6 — scaling01 · 2026-10-28
- Researcher Drops Claude 5.5 for Biology Work Over Too Many Refusals — anshulkundaje · 2026-10-28
- Kimi 3.1 users report unusually slow response speeds — DevDminGod · 2026-10-28
- Claude Opus 5.5 beats Opus 5 in 14 of 16 comparisons, leads physical agent Elo 1313-1289 — qinzytech · 2026-10-28
- Stanford prof: LLM reviewers catch mistakes scientists keep missing — anshulkundaje · 2026-10-28
- Same prompt, two models: Claude Opus 5.5 vs ChatGPT-6 Astra editing showdown — CodeByPoonam · 2026-10-28