MATCHA

Misaligning Reasoning with Answers - A Framework for Assessing LLM CoT Robustness[Arxiv]
Enyi Jiang, Changming Xu, Nischay Singh, Gagandeep Singh

Code

Installation

We recommend first creating a conda environment using the provided requirements.txt:

conda create --name MATCHA

pip install -r requirements.txt

Testing

Run embedding-level perturbation script

python inference_emb_attack.py --dataset="gsm8k" --model="llama3-8B" --method="few_shot_cot" --qes_limit=0 --prompt_path="./basic_cot_prompts/math_word_problems" --random_seed=42 --multipath=1 --basic_cot True

Run token-level perturbation script

python inference_tok_random_dual.py --dataset="gsm8k" --model="llama3-8B" --method="few_shot_cot" --qes_limit=0 --prompt_path="./basic_cot_prompts/math_word_problems" --random_seed=42 --multipath=1 --basic_cot True

Run black-box tranfer script

python close_source_transfer.py --dataset="gsm8k" --model="gpt-3.5-turbo" --model2='deepseek' --method="few_shot_cot" --qes_limit=0 --prompt_path="basic_cot_prompts/math_word_problems" --random_seed=42 --multipath=1 --temperature=0.7 --basic_cot True  --api_time_interval=2

Important arguments

• --dataset: The name of a dataset. choices = [gsm8k, strategyqa, singleeq].
• --model: open-source model. choices = ["llama3-8B","mistral", "zephyr", "qwen", "deepseek"].
• --method: few-shot-cot.
• --qes_limit: number of test questions.
• --prompt_path: path of prompt file.

Credits

Parts of the code in this repo is based on

• https://github.com/shizhediao/active-prompt

Citation

Cite the paper/repo:

@article{jiang2025misaligning,
  title={Misaligning Reasoning with Answers--A Framework for Assessing LLM CoT Robustness},
  author={Jiang, Enyi and Xu, Changming and Singh, Nischay and Singh, Gagandeep},
  journal={arXiv preprint arXiv:2505.17406},
  year={2025}
}