This research reveals a critical tradeoff between safety and reasoning capabilities in large language models.

• Safety alignment techniques can make LRMs more secure but reduce reasoning performance
• The study defines a quantifiable "Safety Tax" that measures this tradeoff
• Evidence indicates safety-aligned models struggle with logically valid but potentially unsafe reasoning paths
• Researchers developed evaluation methods to better measure this phenomenon

For security practitioners, this work highlights important considerations when deploying reasoning-focused AI systems, showing how safety measures may inadvertently compromise core model functionality.

Safety Tax: Safety Alignment Makes Your Large Reasoning Models Less Reasonable