---
title: "Notes to Self: Can LLMs Benefit from Experiential Abstractions?"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.20372v1"
permalink: "https://feed.astralor.com/items/itm_e95f8e8fac39d7bc07921004"
published: "2026-07-22T17:02:34.000Z"
score: 76
channel: "ai"
tags: ["Large Language Models","Reinforcement Learning","Inference-time Retrieval","Mathematical Reasoning","Experiential Abstraction"]
---

# Notes to Self: Can LLMs Benefit from Experiential Abstractions?

## AI commentary

> This paper caught my eye for elegantly transplanting the human cognitive concept of "experience review" onto LLMs. Compared to brute-force scaling, this "self-reminder" approach is far more elegant. It's especially relevant for researchers focused on LLM reasoning optimization and agentic self-iteration. Of course, the results are limited to math and logic tasks—extending to commonsense or open-domain would be more convincing.

## At a glance

This paper investigates whether LLMs can benefit from experiential abstractions—reusable strategies and cautionary reminders—extracted from their own solution traces on the MATH training set, building a retrievable library via a stronger teacher or the LLMs themselves. Two usage modes are explored: inference-time retrieval and reinforcement learning with abstraction-augmented prompts. Results show significant improvements on mathematical and logical reasoning benchmarks, self-extracted abstractions matching teacher-extracted ones, and cross-dataset/model transferability, suggesting LLMs can extract and apply experiential abstractions akin to human experience distillation.

## Key points

- Extracts natural-language abstractions (e.g., strategies and cautionary reminders) from LLM solution traces to build a retrievable library
- Supports two modes: inference-time retrieval and RL-based prompt augmentation
- Achieves significant gains on math and logical reasoning benchmarks, self-extracted matching teacher-extracted
- Framework transfers to other datasets and models, demonstrating generalization

## Full text

Humans distill experience into reusable abstractions, e.g., strategies and cautionary reminders, and apply them to gradually solve problems more effectively. We study whether Large Language Models (LLMs) can similarly benefit from such experiential abstractions. From LLMs' solution traces on the MATH training set, a stronger teacher or the LLMs themselves extract natural-language abstractions into a retrievable library. We explore two usage modes: (1) inference-time retrieval and (2) reinforcement learning (RL) with abstraction-augmented training prompts. Experiential abstractions improve LLM performance on mathematical and logical reasoning benchmarks. Self-extracted abstractions match teacher-extracted ones, and our abstraction usage framework can transfer to other datasets and models. These findings suggest LLMs can extract and apply experiential abstractions much as humans leverage distilled experience.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_e95f8e8fac39d7bc07921004