Training Large Language Models to Reason in a Continuous Latent Space

arxiv:2412.06769

Original source · Paper JSON

Reasoning and the "physics" of language models · 2024-12-09

Cites (6) · Cited by (19)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3

Contents

• Training Large Language Models to Reason in a Continuous Latent Space
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Coconut : Chain of Continuous Thought
• 4 Continuous Space Enables Latent Tree Search
• 4.1 Experimental Setup
• 4.2 Overall Results
• 4.3 Interpreting the Latent Reasoning as Tree Search
• 4.4 Why is Latent Space Better for Planning?
• 5 Empirical Results with Coconut
• 5.1 Experimental Setup
• 5.2 Baselines and Variants of Coconut
• 5.3 Results and Discussion
• 6 Conclusion
• Appendix A Datasets
• A.1 Examples
• A.2 Construction of ProsQA
• A.3 Statistics
• Appendix B Clock-Time Reasoning Efficiency Metric
• Appendix C More Discussion
• C.1 Using More Continuous Thoughts
• C.2 Coconut with Larger Models
• Instructions for reporting errors