Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization

arxiv:2405.15071

Original source · Paper JSON

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

Cites (4) · Cited by (2)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4

Contents

• Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization
• Abstract
• 1 Introduction
• 2 General Setup
• 3 Composition—Delayed Generalization without Systematicity
• 3.1 Setup
• 3.2 Results
• 3.3 Analyzing the inner workings of the model throughout grokking
• 4 Comparison—Systematic Generalization via Parallel Circuit
• 5 The Power of Parametric Memory for Complex Reasoning
• 6 Related Work
• 7 Conclusion
• Limitations
• Acknowledgement
• Appendix A Training Details
• Appendix B Effect of Model Scale
• Appendix C Effect of Tokenizations
• Appendix D More Details on Circuit Analysis
• D.1 Composition
• D.2 Comparison
• Appendix E Additional Results
• E.1 Weight decay
• E.2 Transformer with parameter sharing
• E.3 Comparison task across different inferred/atomic ratio
• Appendix F Complex Reasoning Task
• Instructions for reporting errors