Saved in:
Bibliographic Details
Main Authors: Ye, Shicheng, Yu, Chao, Ke, Kaiqiang, Xu, Chengdong, Wei, Yinqi
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.12810
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914039797383168
author Ye, Shicheng
Yu, Chao
Ke, Kaiqiang
Xu, Chengdong
Wei, Yinqi
author_facet Ye, Shicheng
Yu, Chao
Ke, Kaiqiang
Xu, Chengdong
Wei, Yinqi
contents Large language model (LLM)-based agents have shown strong potential in multi-task scenarios, owing to their ability to transfer knowledge across diverse tasks. However, existing approaches often treat prior experiences and knowledge as monolithic units, leading to inefficient and coarse-grained knowledge transfer. In this work, we propose a novel hierarchical memory architecture that enables fine-grained knowledge transfer by decoupling high-level planning memory from low-level execution memory. To construct and refine these hierarchical memories, we introduce Hierarchical Hindsight Reflection (H$^2$R), a mechanism that distills reusable and hierarchical knowledge from past agent-environment interactions. At test time, H$^2$R performs retrievals of high-level and low-level memories separately, allowing LLM-based agents to efficiently access and utilize task-relevant knowledge for new tasks.Experimental results across two benchmarks demonstrate that H$^2$R can improve generalization and decision-making performance, outperforming prior baselines such as Expel.
format Preprint
id arxiv_https___arxiv_org_abs_2509_12810
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle H$^2$R: Hierarchical Hindsight Reflection for Multi-Task LLM Agents
Ye, Shicheng
Yu, Chao
Ke, Kaiqiang
Xu, Chengdong
Wei, Yinqi
Artificial Intelligence
Large language model (LLM)-based agents have shown strong potential in multi-task scenarios, owing to their ability to transfer knowledge across diverse tasks. However, existing approaches often treat prior experiences and knowledge as monolithic units, leading to inefficient and coarse-grained knowledge transfer. In this work, we propose a novel hierarchical memory architecture that enables fine-grained knowledge transfer by decoupling high-level planning memory from low-level execution memory. To construct and refine these hierarchical memories, we introduce Hierarchical Hindsight Reflection (H$^2$R), a mechanism that distills reusable and hierarchical knowledge from past agent-environment interactions. At test time, H$^2$R performs retrievals of high-level and low-level memories separately, allowing LLM-based agents to efficiently access and utilize task-relevant knowledge for new tasks.Experimental results across two benchmarks demonstrate that H$^2$R can improve generalization and decision-making performance, outperforming prior baselines such as Expel.
title H$^2$R: Hierarchical Hindsight Reflection for Multi-Task LLM Agents
topic Artificial Intelligence
url https://arxiv.org/abs/2509.12810