Notes to Self: Can LLMs Benefit from Experiential Abstractions?
A retrievable library of “notes to self” improved LLM results on math and logic benchmarks.
The paper builds those notes from solution traces on the MATH training set, using either a stronger teacher model or the same model itself. It tests two paths: pulling relevant abstractions at inference time, and using abstraction-augmented prompts during reinforcement learning. Self-extracted notes performed about as well as teacher-extracted ones, and the setup transferred across other datasets and models. ArXiv · AI/CL/LG's note
The paper builds those notes from solution traces on the MATH training set, using either a stronger teacher model or the same model itself. It tests two paths: pulling relevant abstractions at inference time, and using abstraction-augmented prompts during reinforcement learning. Self-extracted notes performed about as well as teacher-extracted ones, and the setup transferred across other datasets and models. ArXiv · AI/CL/LG's note
score 5