Causal-Copilot: An Autonomous Causal Analysis Agent

Last updated on Dec 10, 2025

Autonomous Causal Analysis with Causal-Copilot

The Challenge

Causal analysis is fundamental to scientific discovery and evidence-based decision-making. However, despite rapid advances in causal learning methods, a significant gap remains between theoretical sophistication and practical applicability. Domain experts often cannot leverage these powerful tools due to:

Algorithmic complexity: 20+ methods with distinct assumptions and hyperparameters
Steep learning curves: Requiring expertise in both causal theory and implementation
Configuration challenges: Selecting appropriate methods for specific data characteristics

Our Solution

Causal-Copilot is an LLM-powered autonomous agent that democratizes causal analysis by automating the entire analytical pipeline. Users simply upload their data and describe their analysis goals in natural language—Causal-Copilot handles the rest.

End-to-end workflow: from natural language query to comprehensive causal analysis report

System Architecture

Causal-Copilot is built on a modular architecture with five core components:

Module	Function
User Interaction	Natural language query parsing, domain knowledge integration, interactive feedback loop
Preprocessing	Data cleaning, schema extraction, statistical diagnostics (linearity, stationarity, heterogeneity)
Algorithm Selection	LLM-guided filtering, ranking based on data characteristics, hyperparameter configuration
Postprocessing	Bootstrap confidence evaluation, LLM-guided graph refinement, user revision loop
Report Generation	Causal graph visualization, result interpretation, LaTeX report compilation

Supported Algorithms

Causal-Copilot integrates 20+ state-of-the-art algorithms across three categories:

Causal Discovery

Family	Algorithms	Data Type
Constraint-based	PC, FCI, CD-NOD, PCMCI	Tabular & Time Series
Score-based	GES, FGES, XGES, GRaSP	Tabular
Continuous Optimization	NOTEARS, GOLEM, CALM, CORL, DYNOTEARS	Tabular & Time Series
LiNGAM Family	ICA-LiNGAM, DirectLiNGAM, VAR-LiNGAM	Tabular & Time Series
MB-based	InterIAMB, IAMBnPC, HITON-MB, BAMB	Tabular
Granger Causality	Linear & Nonlinear Granger	Time Series

Causal Inference

Method	Description
Double Machine Learning	LinearDML, SparseLinearDML, CausalForestDML
Doubly Robust Learning	LinearDRL, SparseLinearDRL, ForestDRL
Instrumental Variables	DRIV Family (Linear, Sparse, Forest)
Matching Methods	PSM, CEM
Counterfactual	DoWhy-based counterfactual estimation

Auxiliary Analysis

Feature Importance: SHAP-based attribution
Anomaly Attribution: Causal structure-based root cause analysis

Performance

Causal-Copilot consistently outperforms individual algorithms across diverse scenarios:

Tabular Data (F1 Score)

Scenario	Causal-Copilot	PC	FCI	GES	DirectLiNGAM
Default (p=10, n=1000)	0.89	0.92	0.91	0.92	0.22
Dense Graph (p=0.5)	0.65	0.41	0.44	0.40	0.26
Large Scale (p=50)	0.94	0.70	0.79	N/A	0.23
Super Large (p=100)	0.90	0.68	0.74	N/A	N/A
Extreme Large (p=500)	0.60	N/A	N/A	N/A	N/A
Non-Gaussian Noise	0.97	0.84	0.85	0.86	0.57
Heterogeneous Domains	0.77	0.51	0.62	0.40	0.23
Measurement Error	0.86	0.69	0.80	0.79	0.28

Time Series Data (F1 Score)

Scenario	Causal-Copilot	PCMCI	DYNOTEARS	VARLiNGAM
Small (p=5, l=3)	0.98	0.92	0.97	0.97
Large Lag (l=20)	0.85	0.84	0.77	0.77
Very Large (p=100)	0.18	N/A	N/A	0.12

Interactive Demo

Web-based interface for interactive causal analysis

Try it yourself on Hugging Face Space.

Sample Output

Example PDF report: dataset statistics, causal graph visualization, and analysis results

Quick Start

# Command-line usage
python main.py --data_file data.csv --apikey YOUR_KEY --initial_query "Discover causal relationships"

# Web interface deployment
python Gradio/demo.py

Citation

@article{causalcopilot2025,
  title={Causal-Copilot: An Autonomous Causal Analysis Agent},
  author={Wang, Xinyue and Zhou, Kun and Wu, Wenyi and others},
  journal={arXiv preprint arXiv:2504.13263},
  year={2025}
}