# From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step

**Authors:** Yuntian Deng, Yejin Choi, Stuart Shieber  
**arXiv:** [2405.14838](https://arxiv.org/abs/2405.14838)  
**Date:** 2024-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.14838v1/

## Key points

- Authors Yuntian Deng, Yejin Choi, and Stuart Shieber explore the use of language models for reasoning tasks
- Emphasis on generating explicit chain-of-thought (CoT) steps for high accuracy in final outputs
- Proposal of a novel method to train models to internalize CoT steps by gradually removing intermediate steps through fine-tuning
- Approach allows for simplified reasoning processes while maintaining high performance levels
- GPT-2 Small model achieves impressive accuracy rates of up to 99% on 9-by-9 multiplication problems using this method
- Effectiveness demonstrated on larger language models like Mistral 7B achieving over 50% accuracy on GSM8K without producing any intermediate steps
- Study highlights potential for improving AI systems' reasoning processes through internalization of chain-of-thought steps

## Summary

In their paper titled "From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step," authors Yuntian Deng, Yejin Choi, and Stuart Shieber explore the use of language models for reasoning tasks. They emphasize the importance of generating explicit chain-of-thought (CoT) steps in order to achieve high accuracy in final outputs. The researchers propose a novel method for training models to internalize these CoT steps, gradually removing intermediate steps through fine-tuning. This approach allows for simplified reasoning processes while maintaining high performance levels. Remarkably, even a GPT-2 Small model can achieve impressive accuracy rates of up to 99% on 9-by-9 multiplication problems using this method. The researchers also demonstrate the effectiveness of their approach on larger language models such as Mistral 7B. These advanced models can achieve over 50% accuracy on GSM8K without producing any intermediate steps, showcasing the potential of internalizing CoT steps in enhancing reasoning capabilities within language models. Overall, this study highlights the promising potential for improving AI systems' reasoning processes through the internalization of chain-of-thought steps.