Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges

Nayoung Lee, Ziyang Cai, Avi Schwarzschild, Kangwook Lee, Dimitris Papailiopoulos
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:32930-32964, 2025.

Abstract

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, our method 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 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 expand model capabilities while preserving architectural simplicity.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-lee25d, title = {Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges}, author = {Lee, Nayoung and Cai, Ziyang and Schwarzschild, Avi and Lee, Kangwook and Papailiopoulos, Dimitris}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {32930--32964}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/lee25d/lee25d.pdf}, url = {https://proceedings.mlr.press/v267/lee25d.html}, abstract = {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, our method 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 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 expand model capabilities while preserving architectural simplicity.} }
Endnote
%0 Conference Paper %T Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges %A Nayoung Lee %A Ziyang Cai %A Avi Schwarzschild %A Kangwook Lee %A Dimitris Papailiopoulos %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-lee25d %I PMLR %P 32930--32964 %U https://proceedings.mlr.press/v267/lee25d.html %V 267 %X 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, our method 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 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 expand model capabilities while preserving architectural simplicity.
APA
Lee, N., Cai, Z., Schwarzschild, A., Lee, K. & Papailiopoulos, D.. (2025). Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:32930-32964 Available from https://proceedings.mlr.press/v267/lee25d.html.

Related Material