"RAJ-PGA: Reasoning-Activated Jailbreak and Principle-Guided Alignment Framework for Large Reasoning Models"

"The PGA dataset is a safety alignment dataset specifically designed for Large Reasoning Models (LRMs). It contains 3,989 rigorously verified high-quality samples, aimed at addressing the unique Reasoning-Activated Jailbreak problem in LRMs. Unlike traditional Simple Refusal datasets, PGA strives to shift the alignment paradigm toward Intrinsic Safe Reasoning, thereby enhancing model safety while preserving deep reasoning capabilities."