LAMDA: Large Language Model as Decision Analyst

LAMDA employs large language models (LLMs) to generate influence diagrams.

This repository contains

• the implementation of LAMDA (in src/)
• the data used in the experiments (in data/)
• the scripts of minimal examples (in scripts/)
• the code for the experiments (in experiments/) and analysis of results (in analysis/)
• the experiments outputs (in output/) and results

Setup the environment

pip install -r requirements.txt

Notice:

• To properly install pycid, Python version 3.9 is suggested.
• pycid==0.8.2 depends on an earlier version of pgmpy. You can install pgmpy==0.1.26 after pycid is installed.

Run the experiments

To run the experiments, you need to set the API keys in APIKEYS/apiKeys.py. For example, you can set the OpenAI API key by replacing yourOpenaiApiKey with your actual API key.

openaiKey = "yourOpenaiApiKey"

Then you can run the experiments by running the following scripts:

• experiment/graphExtractExperiment.py: the influence diagram graph geration experiment
• experiment/probabilityExtractExperiment.py: the conditional probability distribution assignment experiment
• experiment/decisionExperiment.py: the decision-making experiment

Run the minimal examples

To get familiar with the functionality of LAMDA, you can run the following scripts:

• createInfluenceDiagram.py: create an influence diagram from JSON files
• extractGraph.py: extract the graph from the text
• extractProbability.py: load existing graph, extract the probability from the text, and solve the influence diagram for optimal policy
• testDecisionMaker.py: test the decision makers used in the experiments (Vanilla, Cot, Sc, Dellma, Aid)

Cite this work

If you find this work useful, please cite:

@article{hong2026lamda,
  title={LAMDA: Large Language Model as Decision Analyst},
  author={Hong, Yifan and Qin, Sizhong and Wang, Chen},
  journal={Decision Analysis},
  year={2026},
  publisher={INFORMS}
}