InferenceEvolve

Towards Automated Causal Effect Estimators through Self-Evolving AI

Can Wang, Hongyu Zhao, Yiqun T. Chen*

96
Experiments
4
Datasets
2
LLMs
100
Iterations each
Projection
Dataset

Code Evolution Viewer

Explore how the best program evolves across checkpoints. Select a run, then click checkpoints to see code changes.

Dataset
Model
Lambda
Split
Rep

Results Explorer

Evolution trajectories showing best-so-far combined score across iterations. Select experimental conditions to filter.

Dataset
Split
Lambda
Model

LLM Judge Panel

Three independent LLM judges assess each evolved estimator for novelty, similarity to known methods, and complexity.

Algorithm Family Distribution

Similarity Score Distribution

Baseline Comparison

How evolved estimators compare against classical baselines (BART, Causal Forest, Backdoor adjustment).

Mode
DatasetMetricBest EvolvedBaselineWin RateBest Improvement

Quick Start

Get started with running your own causal inference evolution experiments.

# Clone the repository
git clone https://github.com/yiqunchen/causal-agent.git
cd causal-agent

# Install dependencies (requires uv)
uv sync

# Configure API keys
export OPENAI_API_KEY="..."
export GOOGLE_API_KEY="..."

# Run an evolution experiment
uv run python main.py \
  --dataset ihdp \
  --model gpt-5 \
  --lambda-weight 0.2 \
  --max-iterations 100 \
  --train-test-split 50:50