Skip to content

Latest commit

 

History

History
21 lines (13 loc) · 1.02 KB

File metadata and controls

21 lines (13 loc) · 1.02 KB

January 2026

tl;dr: A curriculum learning experience of iteratively absorbing CoT into language model itself.

Overall impression

Stepwise Internalization is a method designed to achieve implicit chain-of-thought reasoning by gradually removing intermediate reasoning steps during training, first tokens first absorbed.

This work inspired later more influential work such as Coconut.

Key ideas

  • The primary difference between implicit CoT and No CoT lies in the use of intermediate reasoning steps as supervision during training.
  • There is a trade off between num of CoT tokens remaining and the accuracy. Completely absorb into the model sometimes does not work.

Technical details

Notes