LLMs Learn Like Humans: New Research Shows Language Models Improve Reasoning Through Self-Generated Notes

Summary

A new arXiv paper titled 'Notes to Self: Can LLMs Benefit from Experiential Abstractions?' demonstrates that large language models can extract and apply natural-language abstractions from their own reasoning traces, much like humans create notes and reminders to solve problems more effectively. The research team extracted abstractions from LLM solution attempts on the MATH training set and explored two usage modes: retrieving abstractions at inference time and incorporating them into reinforcement learning training. The findings reveal that experiential abstractions improve LLM performance on mathematical and logical reasoning benchmarks, with self-extracted abstractions matching the quality of teacher-extracted ones. The abstraction usage framework successfully transfers to other datasets and models, suggesting a powerful new mechanism for improving LLM reasoning capabilities.

• The abstraction framework generalizes across datasets and models, suggesting a broadly applicable technique for scaling reasoning capabilities

Editorial Opinion

This research reveals a fascinating parallel between human and machine learning: just as we distill experience into notes and strategies, LLMs can extract reusable abstractions from their own problem-solving attempts. The finding that self-generated abstractions perform as well as teacher-provided ones is particularly significant—it suggests LLMs possess an intrinsic capacity for meta-learning that could unlock more efficient, continuous improvement in reasoning capabilities. If this technique scales, it could represent a meaningful step toward AI systems that learn and improve from their own experience with minimal external intervention.