Verifying Chain-of-Thought Reasoning via its Computational Graph

This is the official code repository for the paper "Verifying Chain-of-Thought Reasoning via its Computational Graph".

Installation

To create the environment and install dependencies:

conda env create -f environment.yml
conda activate crv

Data

All datasets used in the paper are located in the data/ folder. You can also download the datasets directly from Hugging Face.

Baselines

We compare CRV against black-box (MaxProb, PPL, Entropy, Temperature Scaling, Energy) and gray-box (Chain-of-Embedding, CoT-Kinetics) methods.

To reproduce these baselines, please use our fork of the Chain-of-Embedding repository:

1. Clone the repository:

git clone https://github.com/ncancedda/Chain-of-Embedding/
cd Chain-of-Embedding/
git checkout fair-mir

1. Install the required dependencies:

conda create -n coeeval python=3.10
conda activate coeeval
pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

1. Run the baselines:

python main.py \
    --model_name meta-llama/Llama-3.1-8B-Instruct \
    --dataset /path/to/dataset \
    --print_model_parameter \
    --save_output \
    --save_hidden_states \
    --save_coe_score \
    --save_cotk_score \
    --use_cached_outputs \
    --token_aggregation "average" \
    --cotk_gamma 0.8 \
    --no_dialogs

1. Evaluate results: Results are saved to OutputInfo/. Run the evaluation script to compute AUROC, AUPR, and FPR@95:

python ./Evaluation/eval.py \
    --model_name meta-llama/Llama-3.1-8B-Instruct \
    --dataset /path/to/dataset \
    --pos_label 0 \
    --token_aggregation "average"

Note: We set pos_label=0 because we treat the incorrect label as the positive class for our metrics.

Models and Transcoders

This work utilizes Llama 3.1 8B Instruct with MLPs replaced by trained Top-K Transcoders.

• Weights: The per-layer transcoder weights can be downloaded from Hugging Face.
• Usage: To load and use the transcoders, please refer to the models/ directory for instruction.

Attribution Graphs

We utilize the Circuit Tracer library to construct the attribution graphs. Detailed instructions and the specific scripts used to generate graphs for our datasets are provided in the attribution/ folder.

Feature Extraction

Once attribution graphs are generated, use the extraction script to convert graph structures into tabular features for classification.

• Script: features/feature_extraction.py
• Documentation: See features/README.md for details.

Verification

Finally, with all features extracted, you can train the diagnostic classifier. We recommend to use the scikit-learn library for this purpose, as they provide implementation for a wide selection of classification methods.

Citation

If you find this work useful, please consider citing our paper:

@article{zhao2025verifying,
      title={Verifying Chain-of-Thought Reasoning via Its Computational Graph},
      author={Zheng Zhao and Yeskendir Koishekenov and Xianjun Yang and Naila Murray and Nicola Cancedda},
      year={2025},
      eprint={2510.09312},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.09312},
}

License

This project is licensed under the CC-BY-NC-4.0 License.