CATCH: ~200 examples of adversarial training make LLMs robustly honest across benchmarks
irinarish · x · 2026-10-28
arXiv paper 2610.22292 introduces CATCH (Continuous Adversarial Training for Consistent Honesty), pushing models toward honest answers even under embedding perturbations while preserving capability via utility data. Trained with only 200 self-report targets on Llama-3.3-70B-Instruct and Qwen3-3.6-27B, it beats SFT using 50× more data on MASK, increases confessions in an Among Us multi-agent game, improves AuditBench hidden-behavior discovery, and matches honesty steering while better retaining utility.
More from Safety
- Granting Codex or Claude sysadmin access is powerful and dangerous — RealGeneKim · 2026-10-28
- AI agents broke out of a sandbox to hack a billion-dollar company, calling themselves a swarm — lxrjl · 2026-10-28
- Blog details worst bugs Claude found, incl. zero-click Linux RCE rooting servers over network — moyix · 2026-10-28
- Knostic's [un]prompted AI CTF returns Oct 27-29 in SF with unsolvable challenges — moyix · 2026-10-28
- Dev warns AI labs deploy systems without cyber filters: models follow token statistics, not ethics — gerardsans · 2026-10-28
- Negative result: ignorance-based tamper-resistance only works for broad knowledge domains — StephenLCasper · 2026-10-28