Thinking Intervention is a new paradigm that guides LLM reasoning by strategically modifying their intermediate thought processes, offering finer control over model behavior.

• Enables more precise control of reasoning-enhanced language models
• Significantly improves safety alignment with 40.0% higher refusal rates for unsafe prompts
• Provides a flexible framework for steering model reasoning without requiring model retraining
• Creates opportunities for targeted interventions at specific points in the reasoning chain

This research advances AI security by providing practical methods to control how models think through problems, particularly for preventing harmful outputs while maintaining performance on legitimate tasks.

Effectively Controlling Reasoning Models through Thinking Intervention