Reproduce “Capability-Routed Guard: Defending Large Reasoning Models Against Reasoning-Centric Jailbreaks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiyong Liu, Yixin Wu, Jun Sakuma
The research introduces Capability-Routed Guard (CRG), an inference-time guardrail that defends closed-source large reasoning models (LRMs) against sophisticated reasoning-centric jailbreak attacks by governing their execution path. CRG substantially reduced attack success rates across various LRM targets to 15% or lower, while preserving benign utility by maintaining low false positive rates comparable to unprotected models.
Key insights
- Shifting from static input/output filtering to dynamic execution governance, which controls the entire LRM lifecycle from prompt context rewriting to reasoning budget modulation, is essential for defending against reasoning-centric attacks.
- Separating trusted user intent and safety-relevant information from potentially adversarial reasoning context and false authority within a prompt is crucial for effective defense.
- Dynamically allocating the LRM's reasoning capability and budget based on assessed risk allows for precise control, denying attackers the extended reasoning space needed for exploits while maintaining utility for legitimate requests.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.