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

# LeMARLA **L**arge Language Model-**E**nhanced **M**ulti-**A**gent **R**einforcement **L**earning for **A**utonomous enterprise workflow orchestration via graph semantic routing. LeMARLA couples a large language model, a heterogeneous graph attention network and a multi-agent policy under centralized training with decentralized execution. It turns a natural-language business instruction into an executable DAG of subtasks, routes each subtask to the employee best able to handle it, and adapts its plan online whenever a subtask fails. <p align="center"><img src="https://img.shields.io/badge/python-3.10+-blue" /> <img src="https://img.shields.io/badge/pytorch-2.1+-red" /> <img src="https://img.shields.io/badge/license-MIT-green" /></p> --- ## Highlights * Task decomposition: LLaMA-2-13B-chat generates the ordered subtask list and its precedence structure. A learned dependency scorer (Eq. 3) rewrites edges based on subtask semantics rather than surface heuristics.* Semantic routing: a three-relation heterogeneous graph (task, employee, skill) is embedded with type-aware attention (Eq. 4-7). The Eq. 8 routing score fuses cosine similarity of task and employee embeddings with a load-balance penalty.* Multi-agent policy: independent PPO actors observe local workload signals, share a centralized critic during training, and act with a semantic prior injected into the softmax (Eq. 9). Trained with the clipped surrogate of Eq. 12.* Online feedback: failure signals from the execution engine trigger targeted re-decomposition of the failed subgraph without recomputing completed subtasks.* Log-replay evaluation: a duration model and a failure model, fitted on the training split under log-normal likelihood, drive counterfactual outcomes on the withheld test split. Ten independent seeds, Holm-Bonferroni-corrected significance tests. --- ## Repository content Two archives sit at the root of the repository: ```lemarla-code.zip # Source code, configurations, tests, cached model outputslemarla-data.zip # Enterprise workflow dataset (11 xlsx tables + docs)``` Downloading a single ZIP is enough for most users: `lemarla-code.zip` alreadyships the full dataset under `data/lemarla_dataset/`. `lemarla-data.zip` isprovided for those who only need the tables. ### Code archive layout ```lemarla-code/├── README.md├── requirements.txt├── configs/default.yaml # Every hyperparameter (learning rates, seeds, splits)├── data/lemarla_dataset/ # 11 xlsx tables of the operational log├── src/ # 10 modules mapped to Eq. 1-13├── scripts/ # prepare_data / train / evaluate / plot / run_all├── tests/test_smoke.py # 12 unit checks over every module└── outputs/ # Tables, figures (PDF + PNG) and checkpoints``` ### Data archive layout ```lemarla-data/├── README.md├── 01_departments.xlsx├── 02_job_categories.xlsx├── 03_skill_taxonomy.xlsx├── 04_subtask_category_taxonomy.xlsx├── 05_employee_master_table.xlsx├── 06_employee_skill_certifications.xlsx├── 07_collaborates_edges.xlsx├── 08_workflow_instances_raw.xlsx├── 09_subtask_execution_log.xlsx├── 10_dependency_edges.xlsx└── 11_annotated_reference_set.xlsx``` The `data/lemarla_dataset/README.md` inside the archive documents everycolumn and lists the row counts (326 employees, 48 skills, 21,435 rawinstances, 8,742 retained, 41,537 subtasks, 29,086 dependency edges, and a300-instance / 1,521-subtask / 1,063-edge annotated reference set). --- ## Installation ```bash# 1. Clone the repository and unpack the code archivegit clone https://github.com/<owner>/lemarla.gitcd lemarlaunzip lemarla-code.zip -d . # 2. Create a virtual environment and install the pinned requirementspython -m venv .venvsource .venv/bin/activate # Linux / macOSpip install -r lemarla-code/requirements.txt``` Requirements (from `lemarla-code/requirements.txt`): ```torch==2.1.0 torchvision==0.16.0numpy>=1.24.0 pandas>=2.0.0scikit-learn>=1.3.0 scipy>=1.11.0matplotlib>=3.7.0 seaborn>=0.12.0networkx>=3.1 openpyxl>=3.1.0tqdm>=4.65.0 PyYAML>=6.0sentence-transformers>=2.2.0ortools>=9.10.0transformers>=4.38.0torch-geometric>=2.5.0``` The pipeline runs on both GPU (recommended for full training) and CPU(sufficient for evaluation, plotting and unit tests). --- ## Quick start ```bashcd lemarla-code # Sanity: eleven module-level checkspython tests/test_smoke.py # Reproduce every experiment end to endpython scripts/run_all_experiments.py # includes MAPPO trainingpython scripts/run_all_experiments.py --skip-train # tables + figures only # Or run each stage individuallypython scripts/prepare_data.py # writes outputs/tables/table3_cascade.jsonpython scripts/train.py # writes outputs/checkpoints/{simulator.joblib, lemarla.pt}python scripts/evaluate.py # writes six xlsx tables under outputs/tables/python scripts/plot_all_figures.py # writes fourteen figures under outputs/figures/``` --- ## Reproduced results Ten independent seeds, 95% CIs computed with t-critical 2.262, Holm-Bonferronicorrected p-values. | Method | TCR (%) | E2EL (min) | LBI (1/min) | RA@1 (%) | RA@3 (%) | DL (ms) ||--------------------|----------------------|----------------------|----------------------|----------------------|----------------------|---------|| Rule-Based | 78.4 [77.8, 79.0] | 52.1 [51.0, 53.2] | 0.58 [0.57, 0.59] | 38.5 [37.8, 39.2] | 58.7 [58.3, 59.1] | 1.2 || Vanilla-MARL | 83.2 [82.6, 83.8] | 45.3 [44.3, 46.3] | 0.63 [0.62, 0.64] | 40.2 [39.6, 40.8] | 56.9 [56.1, 57.7] | 38.7 || LLM-Only | 84.6 [83.9, 85.3] | 43.5 [41.8, 45.2] | 0.48 [0.46, 0.50] | 52.4 [51.6, 53.2] | 72.3 [71.4, 73.2] | 1250.0 || GNN-MARL | 85.8 [85.0, 86.6] | 38.2 [37.2, 39.2] | 0.62 [0.60, 0.64] | 48.6 [47.9, 49.3] | 69.8 [68.8, 70.8] | 42.3 || TDAG-Assign | 87.4 [86.4, 88.4] | 35.1 [34.2, 36.0] | 0.57 [0.56, 0.58] | 55.1 [54.4, 55.8] | 77.4 [76.8, 78.0] | 852.0 || TDAG-Assign-LB | 87.1 [86.4, 87.8] | 35.4 [34.5, 36.3] | 0.66 [0.64, 0.68] | 54.8 [54.1, 55.5] | 77.0 [75.9, 78.1] | 863.0 || CP-RCPSP | 86.2 [85.7, 86.7] | 33.8 [32.7, 34.9] | 0.69 [0.67, 0.71] | 53.7 [53.0, 54.4] | 76.5 [75.7, 77.3] | 1783.0 || SOP-Orchestrator | 86.5 [85.9, 87.1] | 37.8 [36.9, 38.7] | 0.51 [0.49, 0.53] | 54.9 [54.2, 55.6] | 77.2 [76.2, 78.2] | 974.0 || **LeMARLA** | **91.7 [91.3, 92.1]**| **30.4 [29.0, 31.8]**| **0.71 [0.69, 0.73]**| **60.4 [59.7, 61.1]**| **82.6 [81.8, 83.4]**| **47.6**|| Historical | 88.5 [88.4, 88.6] | 41.2 [41.1, 41.3] | – | – | – | – | All eight baselines are separated from LeMARLA at Holm-p < 0.001 withCohen's d ≥ 3.9 on TCR. Log-replay simulator diagnostics on the 4,154-subtask validation split: ```Duration model: MAE = 1.68 min, RMSE = 2.09 min, R² = 0.813, KS = 0.031Failure model : AUC = 0.877, Brier = 0.075, ECE = 0.013``` --- ## Directory of outputs | Path | Content ||---------------------------------------------------------|--------------------------------------------|| `lemarla-code/outputs/tables/table3_cascade.json` | Preprocessing cascade || `lemarla-code/outputs/tables/Table_5_main_comparison.xlsx` | Ten-method comparison || `lemarla-code/outputs/tables/Table_6_decomposition_quality.xlsx` | Boundary / dependency P*, R*, F1* || `lemarla-code/outputs/tables/Table_7_ablation.xlsx` | Four-way ablation || `lemarla-code/outputs/tables/Table_8_scalability.xlsx` | 50-2000 employee sweep || `lemarla-code/outputs/tables/Table_9_backbones.xlsx` | LLaMA-2/3.1, Qwen2.5 backbones || `lemarla-code/outputs/tables/Table_10_relative_contribution.xlsx` | Percent contribution of each module|| `lemarla-code/outputs/figures/Figure_*.pdf` | 14 figures || `lemarla-code/outputs/figures_png/Figure_*.png` | Rasterised previews of every figure || `lemarla-code/outputs/checkpoints/simulator.joblib` | Trained duration + failure models | --- ## Correspondence between code and paper equations | Manuscript element | Code location ||-------------------------------------------|-----------------------------------------------------|| Eq. 1 Workflow orchestration objective | `src/mappo.py::joint_reward` || Eq. 2 Two-layer projection | `src/decomposition.py::TaskProjection` || Eq. 3 Dependency scoring | `src/decomposition.py::DependencyScorer` || Eq. 4 Type-specific projection | `src/het_gat.py::TypeSpecificProjection` || Eq. 5 Node-level attention | `src/het_gat.py::TypeAwareGATLayer` || Eq. 6 ELU-activated aggregation | `src/het_gat.py::TypeAwareGATLayer.forward` || Eq. 7 Semantic-level fusion | `src/het_gat.py::SemanticLevelAttention` || Eq. 8 Routing score | `src/het_gat.py::routing_score` || Eq. 9 Softmax policy with prior | `src/mappo.py::MAPPOPolicy.action_distribution` || Eq. 10 Joint reward decomposition | `src/mappo.py::MAPPOPolicy.joint_reward` || Eq. 11 Advantage from centralized critic | `src/mappo.py::MAPPOPolicy.advantage` || Eq. 12 Clipped PPO surrogate | `src/mappo.py::MAPPOPolicy.update` || Eq. 13 RA@k routing accuracy | `src/metrics.py::routing_accuracy_at_k` || Table 3 Preprocessing cascade | `scripts/prepare_data.py` || Table 4 Baselines (× 8) | `src/baselines.py::build_all_baselines` || Tables 5-10 | `scripts/evaluate.py` || Figures 1-14 | `scripts/plot_all_figures.py` | --- ## License Released under the MIT License. See `LICENSE` for the full text. --- ## Contact Please open an issue in this repository for questions, suggestions or bugreports.