LLM-empowered graph attention-based return decomposition method
The Computer Journal · 18.08.2026
DOI: 10.1093/comjnl/bxag082 ↗Kurz gefasst
Fachartikel in der Zeitschrift The Computer Journal, online erschienen am 18.08.2026. Verfasst von Zhou Xin und 3 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben.
Abstract (Originalsprache)
The application of reinforcement learning in real-world scenarios is often confronted with the problem of sparse and delayed rewards, which poses significant challenges to the temporal credit assignment of intelligent agents. Although existing reward redistribution methods have made some progress in addressing the credit assignment issue, they still suffer from reward model attribution bias caused by redundant state information, making it difficult to further improve the accuracy of credit assignment. To solve the above problems, we propose a novel framework called LLM-Empowered Graph Attention Return Decomposition (LEGRD) method. Guided by the information bottleneck theory, this method employs modular prompts to instruct large language models to automatically construct abstract graph structures aligned with task semantics, thereby achieving the filtering of redundant information in raw states and the structured modeling of key relationships. On this basis, a return decomposition model based on graph attention networks is designed, which utilizes an adaptive attention mechanism to distinguish the contribution differences of different relationships in the graph to the reward, thus realizing the accurate decomposition from sparse episodic rewards to dense single-step rewards. Experimental results on multiple MuJoCo benchmark tasks show that LEGRD significantly improves sample efficiency. In terms of final policy performance, LEGRD achieves a maximum performance improvement of 24.47% compared with the state-of-the-art return decomposition methods, and $\sim$24.76% compared with the baseline algorithm relying on manual reward design. This fully validates the effectiveness of the proposed framework in alleviating the sparse and delayed reward problem.
Zitieren
APA
Xin, Z., Lei, C., Xiliang, C., & Xiaoyan, W. (2026). LLM-empowered graph attention-based return decomposition method. The Computer Journal. https://doi.org/10.1093/comjnl/bxag082
Vancouver
Xin Z, Lei C, Xiliang C, Xiaoyan W. LLM-empowered graph attention-based return decomposition method. The Computer Journal. 2026. doi:10.1093/comjnl/bxag082
BibTeX
@article{Xin2026,
author = {Xin, Zhou and Lei, Cao and Xiliang, Chen and Xiaoyan, Wang},
title = {LLM-empowered graph attention-based return decomposition method},
journal = {The Computer Journal},
year = {2026},
doi = {10.1093/comjnl/bxag082}
}
EUVIMED