Safer Large Reasoning Models: Safety Decision Before Chain-of-Thought Substantially Improves Alignment

arXiv·medium signal

ArXiv paper 2603.17368 proposes promoting safety policy evaluation to occur before chain-of-thought generation rather than after — finding that models that reason first and apply safety second can be manipulated through the reasoning trace itself. The reordering substantially improves safety capabilities without degrading benchmark performance. Directly relevant as frontier reasoning models (o3, claude-opus-4-6) become the default for agentic deployments.