🧠 Causal-CoT: Causal Chain-of-Thought for Validated Reasoning

Official implementation of Causal-CoT, a framework that integrates causal graph construction, augmentation, and verification into the Chain-of-Thought (CoT) paradigm.
Paper: Causal-CoT: Causal Chain-of-Thought for Validated Reasoning (under review at ICLR 2026)

---

📜 Overview

Chain-of-Thought (CoT) prompting enables large language models (LLMs) to produce step-by-step reasoning. However, generated rationales are often unfaithful or logically inconsistent.
Causal-CoT mitigates these problems by turning linear CoT traces into structured, verifiable causal graphs through a three-stage pipeline:

1. DAG-guided CoT — construct an initial Directed Acyclic Graph (DAG) from the premise and hypothesis.
2. Reflection & Augmentation — enrich the DAG by prompting for missing mediators, confounders, or contextual variables.
3. Causal Verification — estimate conditional probabilities (via LLM prompts), apply do-calculus, and verify causal effects quantitatively.

This pipeline improves reasoning fidelity, interpretability, and stability across mathematics, commonsense, and causal reasoning benchmarks.

---

🧩 Key Features

• Graph-structured reasoning: map CoT steps to DAG nodes/edges.
• Do-calculus verification: quantitative causal effect estimation and hypothesis acceptance.
• Prompt-based augmentation: elicit missing premises, mediators, or confounders from the LLM.
• IR backends (optional): integrate web / knowledge graph / RAG evidence for uncertain edges.
• Modular & extensible: clear separation of DAG construction, augmentation, and verification.

---

🔖 Repository snapshot

Causal-CoT/
├── causal_cot/                # Core pipeline implementation
│   ├── stable_run.py          # Entry script for experiments
│   ├── dag_construction.py    # Stage I: DAG-guided CoT
│   ├── augmentation.py        # Stage II: Reflection & Augmentation
│   └── verification.py        # Stage III: Causal Verification
├── causalnet/                 # Example benchmark datasets (CSV)
├── docs/                      # Documentation & figures (e.g., pipeline image)
├── requirements.txt
├── LICENSE
└── README.md

Note: actual filenames inside causal_cot/ may vary slightly depending on the implementation; the above shows the recommended logical layout.

---

⚙️ Quick start (exact steps)

# 1. Clone repo
git clone https://github.com/AuroraHashcat/Causal-CoT.git

# 2. Enter repository
cd Causal-CoT

# 3. Install dependencies
pip install -r requirements.txt

# 4. Run the example
cd causal_cot
python stable_run.py --mode 1 --input-file ../causalnet/causalnet_llama-8b.csv

• --mode 1 — example execution mode used for demonstration (adjust per project code comments).
• --input-file — path to the dataset CSV (here we use the CausalNet example file).

Outputs (logs, graphs, DAG visualizations, and result summaries) are written to the configured outputs/ or results/ directory (see stable_run.py for exact paths).

---

📋 Usage / Arguments (example)

stable_run.py accepts (at least) these common arguments:

--mode             Execution mode (int). Example: 1 = run full pipeline on dataset.
--input-file       Path to CSV/JSON dataset used by the experiment.
--model            LLM model identifier (optional; depends on your environment).
--output-dir       Directory for results/plots (optional).
--temperature      LLM temperature for prompting (optional).
--seed             Random seed for reproducibility (optional).

See the script's top help for full, up-to-date options: python stable_run.py --help

---

🔬 Reproducibility & experiments

This repository aims to reproduce the main pipeline and representative results from the paper:

• Datasets evaluated: MATH, CausalNet (CNET), COPA, CSQA, GPQA, STRATEGYQA, HellaSwag.
• Typical evaluation: reformulate multiple-choice as binary causal judgments, construct/augment DAGs, estimate conditional probabilities via LLM prompting, apply do-calculus formulas (ATE / NDE / NIE / TE) and threshold (τ) to accept or reject causal links.
• Example reported improvements (averaged over models in the paper):

Domain	Dataset	CoT (%)	Causal-CoT (%)	Δ
Math	MATH	49.3	52.0	+2.7
Causal	CausalNet	61.2	66.0	+4.8
Commonsense	GPQA	38.0	58.7	+20.7

Exact reproduction requires the same LLMs, seeds, and (optional) retrieval backends; see stable_run.py and experiment config for details.

---

🛠 Implementation notes

• DAG construction: decompose premise/hypothesis into atomic statements and extract stated relations to form initial DAG (G₀).
• Augmentation: run internal prompting (or IR retrieval + fusion) to add mediators, confounders, and colliders, producing (G_c).
• Verification: use targeted prompts to obtain verbal likelihoods (e.g., “very unlikely” → calibrated probability), map them to numeric probabilities (via calibration table or Beta priors), then compute causal effects.
• Probabilities & calibration: verbal-to-probability mapping uses calibrated buckets (e.g., very unlikely → 0.01–0.1, possible → 0.3–0.7).

---

🔍 Troubleshooting

• Missing dependencies / import errors: ensure pip install -r requirements.txt.
• LLM API access: configure your environment variables or API keys as described in stable_run.py.
• Dataset path errors: verify relative path for --input-file (e.g., ../causalnet/...).
• Slow runs: causal verification adds latency — use smaller datasets or debug mode for testing.

---

🧾 Citation

If you use this code or the ideas in your work, please cite:

@inproceedings{causalcot2026,
  title     = {Causal-CoT: Causal Chain-of-Thought for Validated Reasoning},
  author    = {Anonymous},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

---

🔐 License

This repository is released under the MIT License. See LICENSE for details.

---

👥 Contributors

• AuroraHashcat — implementation, experiments, and integration.

---

📂 Acknowledgements & reproducibility statement

The approach builds upon Chain-of-Thought prompting and causal inference foundations. Supplementary datasets, scripts, and configurations are included in the repository under docs/ and causalnet/.

---

📬 Contact / Issues

If you find bugs or wish to reproduce specific results, please open an issue or pull request at:
https://github.com/AuroraHashcat/Causal-CoT