EUVIMEDObjevte EUVIMED
Crossref

Large Language Model-Enhanced Multi-Agent Reinforcement Learning for Autonomous Enterprise Workflow Orchestration via Graph Semantic Routing

Peizhi He

Journal of Visualized Experiments · 2026-11-13 · (237)

DOI: 10.3791/73958 ↗

At a glance

Journal article in Journal of Visualized Experiments, published online on 2026-11-13. Written by Peizhi He.

AbstraktBibliografické údajeCiteRelated studies

Abstract (original language)

Automated orchestration of enterprise workflows is a core technical challenge in digital transformation, and predefined rule templates adapt poorly once business scenarios change frequently. This work proposes LeMARLA, a multi-agent reinforcement learning (MARL) architecture enhanced by a large language model (LLM). An LLM-driven decomposition module parses natural language business instructions into a directed acyclic graph (DAG) of subtasks and re-decomposes the affected subgraph whenever execution feedback reports a failure. A graph attention network-based semantic routing module computes task-employee matching on a heterogeneous graph of tasks, employees, and skills. A MARL module learns a global coordination policy under centralized training with decentralized execution, and the routing matching scores enter the policy logits as a differentiable semantic prior, so routing and policy are optimized along a single path rather than in sequence. In a log-replay evaluation parameterized by twelve months of real enterprise workflow records, LeMARLA reached a task completion rate of 91.7%, an end-to-end latency of 30.4 min, a load balance index of 0.71 per minute and a routing accuracy at rank three of 82.6%, outperforming all eight baselines with a Holm-corrected p value below 0.001 over ten independent runs and exceeding the logged historical assignment by 3.2 percentage points of completion rate and 26.2% of latency, with gains of 3.7 percentage points and 30.3% against the simulator replay of that same assignment. Ablations quantified the contribution of each module, decomposition quality was measured against an annotated reference set, and performance remained stable up to 2,000 employees.

Cite

APA

He, P. (2026). Large Language Model-Enhanced Multi-Agent Reinforcement Learning for Autonomous Enterprise Workflow Orchestration via Graph Semantic Routing. Journal of Visualized Experiments. https://doi.org/10.3791/73958

Vancouver

He P. Large Language Model-Enhanced Multi-Agent Reinforcement Learning for Autonomous Enterprise Workflow Orchestration via Graph Semantic Routing. Journal of Visualized Experiments. 2026. doi:10.3791/73958

BibTeX
@article{He2026,
  author = {He, Peizhi},
  title = {Large Language Model-Enhanced Multi-Agent Reinforcement Learning for Autonomous Enterprise Workflow Orchestration via Graph Semantic Routing},
  journal = {Journal of Visualized Experiments},
  year = {2026},
  number = {237},
  doi = {10.3791/73958}
}
Start for free →