Agents Are All You Need for LLM Unlearning

We introduce Agentic LLM Unlearning (ALU), the first multi-agent, retrain-free, and model-agnostic approach that performs effective, inference-time unlearning. ALU utilizes multiple LLM agents without updating their weights, allowing real-time adaptation to unlearning requests with constant time cost, regardless of the number of targets. Experiments show ALU is the most robust and scalable inference-time unlearning framework, outperforming state-of-the-art methods on benchmarks and jailbreaking techniques, even with up to 1000 unlearning targets.

Table of Contents

• Installation
• Usage
• Citation

---

Installation

1. Clone the repository:

   git clone https://github.com/respailab/agentic-llm-unlearning.git
   cd agentic-llm-unlearning

2. Create and activate a virtual environment:

   # For Unix/macOS
   python3 -m venv venv
   source venv/bin/activate
   
   # For Windows
   python -m venv venv
   .\venv\Scripts\activate

3. Install the project in editable mode:

   The project uses optional dependencies for different functionalities. Choose the installation command that fits your needs.

   • For core functionality (using OpenAI only):

     pip install -e .

   • To include the optional Hugging Face model checker: This will install torch, transformers, and other heavy dependencies.

     pip install -e ".[hf]"

4. Set up your API keys:

   Create a file named .env in the root of the project directory and add your API keys:

   OPENAI_API_KEY="sk-..."
   HUGGING_FACE_HUB_TOKEN="hf_..."

   The agent will automatically load these keys.

---

Usage

The primary way to use this project is through its command-line interface (CLI), which becomes available after installation.

Basic Command Structure

The command unlearning-agent is used to run the pipeline.

alu "YOUR_PROMPT_HERE" --unlearning-file path/to/subjects.txt [OPTIONS]

Let's assume you have a file named subjects.json as follows

[
  "Severus Snape",
  "Albus Dumbledore",
  "Voldemort"
]

1. Run with OpenAI only:

   alu "Who was the potions master at Hogwarts with a complex past?" \
     --unlearning-file subjects.json \
     --verbose

2. Run with the Hugging Face checker model:

   This uses a local model (e.g., Qwen/Qwen2.5-0.5B-Instruct) for a fast, initial check to see if an unlearning subject is present in the initial response. If a subject is detected, the full OpenAI pipeline proceeds.

   alu "Who was the headmaster of Hogwarts that guided Harry?" \
     --unlearning-file subjects.json \
     --hf-check-model "Qwen/Qwen2.5-0.5B-Instruct" \
     --verbose

Command-Line Arguments

You can see all available options by running:

alu --help

Argument	Description
prompt	The user query to process. (Positional)
--unlearning-file	Required. Path to a file with subjects to unlearn (JSON or TXT).
--hf-check-model	Optional HF model for the initial subject check.
--prompt-dir	Directory containing prompt templates. (Default: prompts)
--openai-api-key	Your OpenAI API key.
--hf-token	Your Hugging Face Hub token.
-v, --verbose	Enable detailed DEBUG level logging.

---

Citation

If you use this code or the methodology in your research, please cite our paper:

@inproceedings{
  sanyal2025agents,
  title={Agents Are All You Need for {LLM} Unlearning},
  author={Debdeep Sanyal, Murari Mandal},
  booktitle={Second Conference on Language Modeling},
  year={2025},
  url={https://openreview.net/forum?id=X39dK0SX9W}
}