ExpeL: LLM Agents Are Experiential Learners

Source: arXiv Authors: Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, Gao Huang Date: 2023-08-20 Primary category: cs.LG All categories: cs.LG, cs.AI, cs.CL

Abstract

ExpeL gathers experience from prior tasks, extracts natural-language knowledge, and recalls those lessons at inference time to improve future behavior. It is a bridge between pure reflection methods and later systems that compile experience into durable workflows or skills. The main architectural signal is that trajectory-local lessons can become reusable procedural guidance.