Is the Reversal Curse a Binding Problem? Uncovering Limitations of Transformers from a Basic Generalization Failure
arxiv:2504.01928
Reasoning and the "physics" of language models · 2025-04-02
Complete cleaned Markdown · Source Markdown including bibliography
Read in parts
Contents
- Is the Reversal Curse a Binding Problem? Uncovering Limitations of Transformers from a Basic Generalization Failure
- Abstract
- 1 Introduction
- 2 Learning Reversal at the Concept Level
- 3 The Binding Problem Underlying Surface-level Predictions
- 3.1 Inconsistency of Concept Representations
- 3.2 Entanglements of Concept Representations
- 3.3 Experiments
- 4 Arithmetic Reasoning via Parametric Forward-Chaining
- 5 Related Work
- 6 Discussion & Conclusion
- Acknowledgement
- Ethics Statement
- Reproducibility Statement
- Appendix A Hyperparameters and Training Details
- Appendix B Additional Discussion & Results on JEPA
- JEPA Framework.
- Experimental Stability.
- Appendix C Arithmetic Reasoning
- C.1 Learning Basic Arithmetic Deductions
- C.2 Scaling the Reasoning Challenge
- Appendix D Use of Large Language Models
- Instructions for reporting errors