Notes to Self: Can LLMs Benefit from Experiential Abstractions?
- ID: 1b481872
- 原文链接: https://arxiv.org/abs/2607.20372
- PDF: https://arxiv.org/pdf/2607.20372
- 作者: Chang Liu, Xinyu Li, Artur Dubrawski
- 日期: 2026-07-22
- 分类: cs.CL
- 标签: llm-agents, memory, reasoning, experiential-learning, math
- 质量评分: 4/5
- 抓取时间: 2026-07-24T12:19:10+08:00
中文解读
这篇论文把经验总结显式做成 LLM 可检索的自然语言抽象库：从 MATH 解题轨迹中抽取策略/提醒，再在推理时检索，或放进强化学习训练提示摘要称这种 experiential abstractions 能提升数学与逻辑推理，自提取效果接近强教师提取，并可迁移到其他数据集和模型对 agent memory / reflection 系统有参考价值：记忆不只是保存轨迹，而是压缩成可复用的解题抽象
为什么值得关注
- 把 agent memory / reflection 从保存完整轨迹推进到可检索的经验抽象，适合作为长期记忆与推理增强系统的设计参考。
English Summary
Humans distill experience into reusable abstractions, e.g., strategies and cautionary reminders, and apply them to gradually solve problems more effectively. The paper tests whether LLMs can do the same by extracting natural-language abstractions from solution traces on MATH into a retrievable library, using them either at inference time or as RL training prompt augmentation. The abstract reports gains on mathematical and logical reasoning benchmarks, self-extracted abstractions matching teacher-extracted ones, and transfer across datasets and models.
Abstract
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.
Metadata
- arXiv ID: 2607.20372
- Primary category: cs.CL
- Categories: cs.CL
- Source: OpenCLI arXiv metadata