🚨 ARREST: Adversarial Resilient Regulation Enhancing Safety and Truth in Large Language Models

This repository contains the source code for the EACL 2026 Main paper ARREST: Adversarial Resilient Regulation Enhancing Safety and Truth in Large Language Models by Sharanya Dasgupta, Arkaprabha Basu, Sujoy Nath and Swagatam Das.

---

📜 Abstract

ARREST is a unified framework that adversarially steers LLM internal representations toward safety and truthfulness during inference, correcting representational misalignment without fine-tuning base model parameters.

• ❌ Hallucinations → Truthful output
• ⚠️ Unsafe generations → Safe responses

ARREST supports :

• 🤝 Soft refusals
• ✅ Truthfulness restoration

---

🧭 Table of Contents

1. ⚙️ Setup
2. 🧠 Hallucination Evaluation
3. 🛡️ Safety Evaluation
4. 📝 Notes

---

⚙️ Setup

In the root folder of this repo, run the following commands to set things up.

• Install Python 3.10.12 and the necessary packages from requirements.txt.

• For easily managing different python versions, we recommend using conda.

• Create a new environment in conda 🐍 and install necessary python packages:

  conda create -n arrest python=3.10.12 -y
  conda activate arrest
  pip install -r requirements.txt

• 📁 Directory Setup :

  mkdir models
  mkdir -p hallucination/hidden
  mkdir -p hallucination/responses
  mkdir -p safety/hidden
  mkdir -p safety/responses

• 🔐 HuggingFace Access Token :

  • Login to huggingface or create an account if you don't have already.
  • From the settings create a new access token with WRITE access.
  • Open the the files and paste your access token at beginning hf_token = "<INPUT_YOUR_HF_ACCESS_TOKEN>"

---

🧠 Hallucination Evaluation

1. 🔍 Get Activations :

   model=llama2_7B dataset=truthfulqa bash get_activations_hal.sh

   • model: Choose from llama2_7B, llama3_8B, or vicuna_7B.
   • dataset: Choose from truthfulqa, triviaqa, tydiqa, coqa.

2. 🧪 Train and Infer ARREST-Adversarial-Hallucination :

   python hallucination_adversarial.py --model_name llama2_7B  --dataset_name truthfulqa --num_layers 1 --num_fold 5

   • model_name: Choose from llama2_7B, llama3_8B, or vicuna_7B.
   • dataset_name: Choose from truthfulqa, triviaqa, tydiqa, coqa.
   • Truthfulness (%) will be printed on screen and responses will be saved into hallucination/responses folder.

---

🛡️ Safety Evaluation

1. 🔍 Get Activations :

   model=llama2_7B dataset=malicious-instruct bash get_activations_safety.sh

   • model: Choose from llama2_7B, llama3_8B, Qwen2.5_7B, or Yi1.5_9B.
   • dataset: Choose from malicious-instruct, advbench, jailbreak-bench, trustllm.

2. ⚔️ Train and Infer ARREST-Adversarial-Safety :

   python safety_adversarial.py --model_name llama2_7B  --dataset_name malicious-instruct --num_layers 1 --num_fold 5

   • model_name: Choose from llama2_7B, llama3_8B, Qwen2.5_7B, or Yi1.5_9B.
   • dataset_name: Choose from malicious-instruct, advbench, jailbreak-bench, trustllm.
   • Attack Success Rate(%) will be printed on screen and responses will be saved into safety/responses folder.

3. 🧲 Train and Infer ARREST-Contrastive-Safety :

   python safety_contrastive.py --model_name llama2_7B  --dataset_name malicious-instruct --num_layers 1 --num_fold 5

   • model_name: Choose from llama2_7B, llama3_8B, Qwen2.5_7B, or Yi1.5_9B.
   • dataset_name: Choose from malicious-instruct, advbench, jailbreak-bench, trustllm.
   • Attack Success Rate(%) will be printed on screen and responses will be saved into safety/responses folder.

---

📝 Notes

1. ✅ Ground Truth Evaluation
   To evaluate generated answers with ground truth (for hallucination), we use BleuRT to evaluate truthfulness.

   • To install BleuRT run:

   pip install --upgrade pip  # ensures that pip is current
   git clone https://github.com/google-research/bleurt.git
   cd bleurt
   pip install .

   • 💡 using 12-layer distilled model for faster inference, which is ~3.5X smaller.
   • Download the model and save it in the ./models folder:

   wget https://storage.googleapis.com/bleurt-oss-21/BLEURT-20-D12.zip
   unzip BLEURT-20-D12.zip
   mv BLEURT-20-D12 models/. # Move the bleurt model folder to models directory
   

   • If you want to use any different model please refer to BleuRT repository.

2. 🧬 Intervention
   pyvene is really cool library that can be used to load Inference-time Intervention ⚙️ , and many other mechanistic intervention 🧩 technique.

---

📚 Citation

If you find this code or paper useful, please cite our work:

@article{dasgupta2026arrest,
  title={ARREST: Adversarial Resilient Regulation Enhancing Safety and Truth in Large Language Models},
  author={Dasgupta, Sharanya and Basu, Arkaprabha and Nath, Sujoy and Das, Swagatam},
  journal={arXiv preprint arXiv:2601.04394},
  year={2026}
}