Constitutional adapters: Inference-time interventions for misalignment and misuse
It is shown that constitution-consistent behavior can be distilled from synthetic corpora into lightweight objects (low-rank adapters and steering vectors) that increase jailbreak defense success and measured alignment and recommend CAs as a lightweight, portable, and tunable lever for mitigating misalignment and misus...
Adam S. Lowet, Mark Kurzeja
· 0 citations