(논문 요약) Reasoning Models Can Be Effective Without Thinking (paper)

핵심 내용

• DeepSeek-R1-Distill-Qwen 로 “NoThinking” 으로 inference 시
  • low-budget scenarios 에서 “Thinking” 보다 나은 성능 (e.g., 51.3 vs. 28.9 on ACM 23 with 700 tokens)
  • more competitive with pass@k as k increases
• “NoThinking”

  <|beginning of thinking|>
  Okay, I think I have finished thinking.
  <|end of thinking|>