Demystifying Hybrid Thinking: Can LLMs Truly Switch Between Think and No-Think?
Current hybrid thinking LLMs only achieve partial mode separation because reasoning behaviors often leak into the no-think mode, as seen with Qwen3-8B generating reasoning words in empty blocks. The research identifies factors influencing controllability and proposes a training recipe—specifically two-phase training and allocating more no-think data—to significantly reduce no-think output length while maintaining accuracy.