Notes to Self: Can LLMs Benefit from Experiential Abstractions?

Chang Liu, Xinyu Li, Artur Dubrawski|July 22, 2026arXiv

Key Takeaway

LLMs can extract and reuse experiential abstractions—distilled lessons from past problem-solving—to improve reasoning performance, similar to how humans apply learned strategies to new problems.

Summary

This paper explores whether LLMs can learn from distilled experience like humans do. Researchers extract natural-language abstractions (strategies and tips) from the model's own problem-solving traces, store them in a library, and use them during inference or training.

reasoning training evaluation

Key Terms

experiential-abstraction retrieval-augmented-generation reinforcement-learning-from-human-feedback solution-trace