Large language models often struggle with length generalization and solving complex problem instances beyond their training distribution. We present a self-improvement approach where models iteratively generate and learn from their own solutions, progressively tackling harder problems while maintaining a standard transformer architecture. Across diverse tasks including arithmetic, string manipulation, and maze solving, self-improving enables models to solve problems far beyond their initial training distribution-for instance, generalizing from 10-digit to 100-digit addition without apparent saturation. We observe that in some cases filtering for correct self-generated examples leads to exponential improvements in out-of-distribution performance across training rounds. Additionally, starting from pretrained models significantly accelerates this self-improvement process for several tasks. Our results demonstrate how controlled weak-to-strong curricula can systematically teach a model logical extrapolation without any changes to the positional embeddings, or the model architecture.
@article{arxiv.2502.01612,
title = {Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges},
author = {Nayoung Lee and Ziyang Cai and Avi Schwarzschild and Kangwook Lee and Dimitris Papailiopoulos},
journal= {arXiv preprint arXiv:2502.01612},
year = {2025}
}