The benchmark comprises of 161 programming problems Our method, stair (safety alignment with introspective reasoning), guides models to think more carefully before responding. Our analysis yields a novel robustness metric called clever, which is short for cross lipschitz extreme value for network robustness
rixia mao (eiyuu densetsu and 2 more) drawn by binzume_suika | Danbooru
Across evaluated tasks, rahp yields compact models with stronger clever lower bounds and minimal change in clean accuracy, and it improves resistance to a wide variety of strong attacks
One common approach is training models to refuse unsafe queries, but this strategy can be vulnerable to clever prompts, often referred to as jailbreak attacks, which can trick the ai into providing harmful responses
