Search papers, labs, and topics across Lattice.
CRAFT, a novel red-teaming alignment framework, is introduced to enhance the robustness of large reasoning models against jailbreak attacks by aligning them to generate safety-aware reasoning traces. This is achieved by optimizing objectives defined over the hidden state space, integrating contrastive representation learning with reinforcement learning to differentiate between safe and unsafe reasoning trajectories. Experiments on Qwen3-4B-Thinking and R1-Distill-Llama-8B show that CRAFT significantly outperforms existing defenses like IPO and SafeKey, improving reasoning safety by 79.0% and final-response safety by 87.7% over the base models.
By aligning hidden representations, CRAFT achieves a remarkable 79% improvement in reasoning safety, suggesting that latent-space interventions are a potent defense against jailbreaks.
We propose CRAFT, a red-teaming alignment framework that leverages model reasoning capabilities and hidden representations to improve robustness against jailbreak attacks. Unlike prior defenses that operate primarily at the output level, CRAFT aligns large reasoning models to generate safety-aware reasoning traces by explicitly optimizing objectives defined over the hidden state space. Methodologically, CRAFT integrates contrastive representation learning with reinforcement learning to separate safe and unsafe reasoning trajectories, yielding a latent-space geometry that supports robust, reasoning-level safety alignment. Theoretically, we show that incorporating latent-textual consistency into GRPO eliminates superficially aligned policies by ruling them out as local optima. Empirically, we evaluate CRAFT on multiple safety benchmarks using two strong reasoning models, Qwen3-4B-Thinking and R1-Distill-Llama-8B, where it consistently outperforms state-of-the-art defenses such as IPO and SafeKey. Notably, CRAFT delivers an average 79.0% improvement in reasoning safety and 87.7% improvement in final-response safety over the base models, demonstrating the effectiveness of hidden-space reasoning alignment.
