Self-Guard: Defending Large Reasoning Models via enhanced self-reflection

TL;DR

Self-Guard is a lightweight framework that enhances safety compliance in large reasoning models by leveraging self-reflection and internal safety activation, effectively reducing risks like manipulation and leakage without sacrificing utility.

Contribution

It introduces a novel, representational-level safety reinforcement method combining safety prompting and activation steering to improve model safety awareness and compliance.

Findings

Effective in bridging the awareness-compliance gap

Generalizes well across unseen risks and model scales

Maintains model utility while enhancing safety

Abstract

The emergence of Large Reasoning Models (LRMs) introduces a new paradigm of explicit reasoning, enabling remarkable advances yet posing unique risks such as reasoning manipulation and information leakage. To mitigate these risks, current alignment strategies predominantly rely on heavy post-training paradigms or external interventions. However, these approaches are often computationally intensive and fail to address the inherent awareness-compliance gap, a critical misalignment where models recognize potential risks yet prioritize following user instructions due to their sycophantic tendencies. To address these limitations, we propose Self-Guard, a lightweight safety defense framework that reinforces safety compliance at the representational level. Self-Guard operates through two principal stages: (1) safety-oriented prompting, which activates the model's latent safety awareness to evoke spontaneous reflection, and (2) safety activation steering, which extracts the resulting directional shift in the hidden state space and amplifies it to ensure that safety compliance prevails over sycophancy during inference. Experiments demonstrate that Self-Guard effectively bridges the awareness-compliance gap, achieving robust safety performance without compromising model utility. Furthermore, Self-Guard exhibits strong generalization across diverse unseen risks and varying model scales, offering a cost-efficient solution for LRM safety alignment.