(논문 요약) Training Large Language Models to Reason in a Continuous Latent Space (Paper)

핵심 내용

• Input embedding 을 생성 (continuous space)
  

• 학습시 step 한 단계를 $c$ 단계의 continuous token(s) 으로 대체
  

실험 결과

• iCoT: internalized CoT
• Pause token: special tokens inserted between the question and answer
• w/o curriculum: directly train questions and answers in the final stage
• w/o thought: no continuous latent thought generated
• pause as thought: tokens to replace the continuous thoughts