Tag
The paper presents a halt vector method to internally control thinking length in reasoning models like DeepSeek-R1-Distill-Qwen-7B, reducing unnecessary reasoning while preserving accuracy.