The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

arxiv:2309.12288

Original source · Paper JSON

Reasoning and the "physics" of language models · 2023-09-21

Cites (1) · Cited by (5)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

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

Contents

• The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”
• Abstract
• 1 Introduction
• 1.1 Contributions: Evidence for the Reversal Curse
• 2 Experiments and results
• 2.1 Experiment 1: Reversing descriptions of fictitious celebrities
• 2.1.1 Dataset and finetuning
• 2.1.2 Results
• 2.2 Experiment 2: The Reversal Curse for real-world knowledge
• 2.3 Experiment 3: Reversing instructions
• 2.3.1 Dataset and finetuning
• 2.3.2 Results
• 3 Related work
• The Reversal Curse in LLMs trained from scratch
• Studying the Reversal Curse with influence functions
• Mechanisms explaining factual recall
• Knowledge editing in LLMs
• Inconsistencies in language model statements
• Forward vs backward recall in humans
• 4 Discussion and future work
• 4.1 Future Work
• Studying other types of relations
• Finding reversal failures via entity-linking
• Analyzing the practical impact of the Reversal Curse
• Contributions and Acknowledgments
• Appendix A Reproducibility
• Appendix B Additional details for Experiment 1
• B.1 Dataset
• B.2 GPT-3-350M hyperparameter sweep
• B.3 Scaling experiment
• B.4 Llama-7b hyperparameter sweep
• B.5 Statistical analysis of log-probabilities
• B.6 In-context results
• B.7 Ablation with larger dataset
• B.8 Ablation using prompt tuning
• B.8.1 Dataset
• B.8.2 Training details
• B.8.3 Results
• Appendix C Additional details for Experiment 2
• C.1 Few-shot prompts
• C.2 Personally identifiable information
• Appendix D Experiment 3: Reversing instructions
• D.1 Llama-1 sweep
• Appendix E Compute costs
• Appendix F Relationship between our work and Grosse et al. 2023
• Appendix G Forward vs backward recall in humans
• Instructions for reporting errors