Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models

Researchers have detailed a method to improve the safety of large reasoning models by activating an inherent capability they call Latent Safety Awareness, reducing the success rate of harmful attacks without relying on external human annotation. The approach, described in a paper submitted to arXiv on June 15, targets a persistent vulnerability in Large Reasoning Models (LRMs). While these models perform well on complex tasks, they remain susceptible to sophisticated jailbreaks and direct harmful queries [1]. Prior safety alignment methods have depended heavily on manual data labeling by humans [1]. The new work instead leverages what the authors term Latent Safety Awareness — the observation that LRMs can identify safety risks when presented with original queries alongside their own reasoning trajectories [1]. To exploit this, the researchers first use Supervised Fine-Tuning (SFT) to induce safe tags that trigger safety analysis for unsafe queries while preserving standard responses for general queries. They then apply Direct Preference Optimization (DPO) to improve the correctness and stability of that safety guidance [1]. All training responses are generated by the models being optimized [1]. Experiments on the DeepSeek-R1-Distill-Llama-8B model showed the Attack Success Rate dropped by an average of 24.65% on harmful benchmarks and 36.72% on jailbreak benchmarks [1]. The method, called Safe Trigger, had almost no negative impact on general performance or user experience, according to the paper [1]. The research appears on arXiv, an open-access repository for electronic preprints that has hosted more than two million articles since its founding in 1991 and now receives roughly 24,000 submissions per month [8]. The paper is accompanied by experimental tools developed through arXivLabs, a framework that allows community collaborators to build features such as citation explorers and code finders directly on the platform [6][7]. arXivLabs was formalized in 2020 to enable third-party innovation while upholding values of openness, community, and user data privacy [6]. The project is currently on a hiatus for new proposals while the arXiv development team focuses on modernizing its infrastructure and moving systems to the cloud [5].