Active Causal Experimentalist (ACE)

Learning Intervention Strategies via Direct Preference Optimization

---

Overview

ACE (Active Causal Experimentalist) learns experimental design strategies for causal discovery using Direct Preference Optimization (DPO). By learning from pairwise comparisons between candidate interventions rather than non-stationary reward magnitudes, ACE achieves stable training and strategic intervention selection.

Key Results:

• 70-71% improvement over baselines at equal intervention budgets
• Statistical significance: p<0.001, Cohen's d ≈ 2.0
• Scales to 15-node systems while maintaining competitive performance
• Multi-domain validation: Synthetic benchmarks, physics simulations, economic data

---

Repository Structure

ACE/
├── paper/                    # LaTeX source for ICML 2026 submission
│   ├── paper.tex
│   └── references.bib
├── ace_experiments.py        # Main ACE implementation (5-node)
├── baselines.py             # Baseline methods (Random, PPO, Max-Variance, etc.)
├── experiments/             # Domain-specific experiments
│   ├── complex_scm.py       # 15-node complex SCM
│   ├── duffing_oscillators.py
│   ├── phillips_curve.py
│   └── run_ace_complex_full.py  # Full ACE for 15-node
├── jobs/                    # SLURM job scripts for HPC
├── scripts/                 # Analysis and utility scripts
├── tests/                   # Comprehensive test suite
├── results/                 # Experimental results
└── docs/                    # Technical documentation, development docs, and archives

---

Quick Start

Installation

# Clone repository
git clone https://github.com/PatrickAllenCooper/ACE.git
cd ACE

# Create conda environment
conda env create -f environment.yml
conda activate ace

# Or install with pip
pip install -r requirements.txt

Run ACE on 5-Node Benchmark

# Single seed
python ace_experiments.py --episodes 200 --seed 42

# Multi-seed validation (N=5)
bash jobs/workflows/run_ace_only.sh

Run Ablation Studies

# Diversity ablation
python ace_experiments.py --no_diversity_reward --episodes 100 --seed 42

# Submit all ablations to HPC
bash jobs/workflows/submit_ablations_verified.sh

Run Complex 15-Node SCM

# Local
python experiments/run_ace_complex_full.py --seed 42 --episodes 300

# HPC
sbatch jobs/run_ace_complex_single_seed.sh

---

Key Components

ACE Architecture

• Policy: Qwen2.5-1.5B language model for intervention generation
• Training: Direct Preference Optimization (DPO) from pairwise comparisons
• Reward: Information gain + node importance + exploration diversity
• Lookahead: K=4 candidates evaluated on cloned learners
• Stability: Reference policy updates, observational training

Baseline Methods

• Random: Uniform random intervention selection
• Round-Robin: Systematic node coverage
• Max-Variance: Uncertainty sampling via MC Dropout
• PPO: Learned policy with value-based RL

---

Experimental Results

Experiment	N	ACE Result	Baseline	Improvement
5-Node SCM	5	0.92 ± 0.73	2.03-2.10	70-71%
15-Node SCM	1	4.54	4.51-4.71	Competitive
Lookahead Ablation	5	0.61	2.10 ± 0.11	71%
Diversity Ablation	2	2.82 ± 0.22	-	+206% degrad.

Statistical Validation:

• Paired t-tests with Bonferroni correction (α = 0.0125)
• Large effect sizes (Cohen's d: 2.0-2.5)
• Multiple independent seeds for robustness

---

Documentation

• docs/: Main documentation directory containing technical specs and archived reports
• docs/submission/: Paper submission preparation and reports (e.g., SUBMISSION_CHECKLIST.md)
• docs/development/: Development status, guidance, and experimental planning docs
• docs/FINAL_PAPER_RESULTS_SUMMARY.md: Complete experimental results

---

Citation

@inproceedings{ace2026,
  title={Active Causal Experimentalist (ACE): Learning Intervention Strategies via Direct Preference Optimization},
  author={Anonymous},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}

---

License

MIT License - See LICENSE file for details.

---

Supplementary Materials

See supplementary_materials/ for:

• Complete code for all experiments
• Detailed setup instructions
• Result analysis scripts
• Additional documentation

Packaged archive: ace_supplementary_materials.tar.gz

---

Contact

For questions about the code or experiments, please open an issue on GitHub.

Paper Submission: ICML 2026 (Submitted February 2026)