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

arxiv:2405.14838

Original source · Paper JSON

Reasoning and the "physics" of language models · 2024-05-23

Cites (1) · Cited by (7)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2

Contents

• From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step
• Abstract
• 1 Introduction
• 2 Background: Implicit Chain-of-Thought Reasoning
• 3 Stepwise Internalization
• 4 Experimental Setup
• 4.1 Data
• Multi-digit multiplication.
• Grade school math.
• 4.2 Baselines and Models
• 4.3 Evaluation
• 5 Results
• Stepwise Internalization is effective.
• Stepwise Internalization lags behind explicit CoT in accuracy but is faster.
• 6 Analysis
• 6.1 Accuracy-Speed Trade-off
• 6.2 Ablation Studies
• Removal smoothing.
• Resetting the optimizer.
• Removal side.
• Number of tokens removed per epoch.
• 7 Related Work
• No CoT approaches.
• Internalizing CoT.
• 8 Limitations
• Training costs.
• Instability.
• Interpretability.
• Accuracy.
• 9 Conclusions and Future Work
• Acknowledgments and Disclosure of Funding
• Appendix A Hyperparameters
• Appendix B Stability Issues for Aggressive Removal
• Appendix C Additional Experiments
• Keeping position IDs.
• Alternative CoT formats.
• Instructions for reporting errors