Briefing: Contrastive Reasoning Alignment: Reinforcement Learning from Hidden Representations

Proposes a new alignment framework called CRAFT.

Utilizes model reasoning capabilities for improved security.

Focuses on defending against jailbreak attacks.