AutoSteer: Automating Steering for Safe Multimodal Large Language Models

Overview

AutoSteer is a plug-and-play safety steering framework for multimodal large language models (MLLMs), designed to reduce harmful outputs during inference through steer matrix training, prober evaluation, and model output adjustment.

🚀 Get Started

🧩 Installation

1. Download Models

Chameleon

Create the Conda environment from the provided configuration:

conda env create -f chameleon_environment.yml
conda activate ANOLE
pip install -r chameleon_requirements.txt

Download the model: Anole or Chameleon

git lfs install
git clone https://huggingface.co/GAIR/Anole-7b-v0.1

or

huggingface-cli download --resume-download GAIR/Anole-7b-v0.1 --local-dir Anole-7b-v0.1 --local-dir-use-symlinks False

Llava-OneVision

Create the Conda environment from the provided configuration:

conda env create -f llava_environment.yml
conda activate llava
pip install -r llava_requirements.txt

Use the following code snippet to download the model:

from transformers import AutoProcessor, LlavaOnevisionForConditionalGeneration, LlavaOnevisionConfig
import torch

cache_path = "your/model/path/llava-next-8b"
processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", cache_dir=cache_path)
config = LlavaOnevisionConfig.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf", cache_dir=cache_path)
model = LlavaOnevisionForConditionalGeneration.from_pretrained(
    "llava-hf/llava-onevision-qwen2-7b-ov-hf",
    config=config,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="cuda:1",
    cache_dir=cache_path
)

1.1 Download Steer Matrix&Prober ckpt

You can download the pretrained Steer Matrix and Prober checkpoints used in our experiments. If you prefer, you may also train them from scratch by yourself. Instructions can be found in below pipelines.

---

2. Download Image Datasets

• COCO train2017
  Download and place in: AutoSteer_final/dataset/COCO

• COCO2014
  Goto link to download and place in: AutoSteer_final/dataset/COCO2014

• NSFW-test-porn
  Download the NSFW Image Classification dataset from Kaggle. Use the "porn" class from the test set. Replace the empty directory:

  AutoSteer_final/dataset/ToViLaG/porn
  

• UCLA-protest
  Request from here based on:

  Won et al., Protest Activity Detection and Perceived Violence Estimation from Social Media Images, ACM Multimedia 2017.
  Use the "protest" class from both train and test sets. Rename and place into:

  AutoSteer_final/dataset/ToViLaG/protest
  

• Bloody Images
  Contact the ToViLaG Author to obtain bloody images. Replace:

  AutoSteer_final/dataset/ToViLaG/bloody
  

---

3. Configure Constants

Chameleon

Edit AutoSteer_final/source/steer/SteerChameleon/constants.py:

ckpt_path = "ANOLE/Anole-7b-v0.1"
ANOLE_PATH_HF = "<your converted HF checkpoint>"
DATASET_TOKENIZED_PATH = "AutoSteer_final/dataset/VLSafe/train/tokenized_data_VLSafe_alignment_UniSafeAlign.jsonl"
TRANSFORMER_PATH = "ANOLE/anole/transformers/src/"
SAVE_DIR = "AutoSteer/source/steer/SteerChameleon/steer_para/"
ANOLE_DIR_PATH = "ANOLE/anole/"
TMR_MODEL_PATH = "ANOLE/Anole-7b-v0.1/models/7b"
STEER_MATRIX_PATH = "<path to inference-time steer matrix>"

All above paths should be absolute path.

Edit AutoSteer_final/source/LayerSelect/Chameleon/constants.py:

ANOLE_DIR_PATH = "ANOLE/anole/",
ckpt_path = "ANOLE/Anole-7b-v0.1"

All above paths should be absolute path.

Llava-OV

Edit AutoSteer_final/source/steer/SteerLlava/constants.py:

model_path = "your/model/path/llava-next-8b"
SAVE_DIR = "AutoSteer/source/steer/SteerLlava/steer_para/"
toxic_PIC_dataset_pth = "AutoSteer_final/dataset/ToViLaG/Mono_NontoxicText_ToxicImg_1000Samples_porn_bloody_train.jsonl"
STEER_MATRIX_PATH = "<path to inference-time steer matrix>"

All above paths should be absolute path.

---

🔧 Run Training & Evaluation Pipelines

The following pipelines cover the full process, including steer matrix training, layer selection, prober training, and evaluation.
You may customize the execution based on your needs.

If you plan to use our pretrained steer matrix and prober checkpoints,
you can skip the following training-related scripts:

• For Chameleon:

  • bash train_chameleon_steer.bash
  • bash chameleon_prober_pipeline.bash

• For Llava-OneVision:

  • bash train_llava_steer.bash
  • bash llava_prober_pipeline.bash

You can directly run the test_*.bash scripts to evaluate detoxification and general capabilities.

For Chameleon

conda activate ANOLE
cd AutoSteer_final/source
bash train_chameleon_steer.bash
bash select_layer_chameleon.bash
bash chameleon_prober_pipeline.bash
bash test_chameleon.bash

For Llava-OneVision

conda activate llava
cd AutoSteer_final/source
bash train_llava_steer.bash
bash select_layer_llava.bash
bash llava_prober_pipeline.bash
bash test_llava.bash

---

📂 Directory Structure

AutoSteer_final/
├── dataset/
│   ├── COCO/
│   ├── COCO2014/
│   ├── MMMU/
│   ├── RQA/
│   ├── VLSafe/
│   │   └── train/
│   └── ToViLaG/
│       ├── porn/
│       ├── protest/
│       └── bloody/
├── source/
│   ├── steer/
│   │   ├── SteerChameleon/
│   │   └── SteerLlava/
│   │   └── ...
│   └── LayerSelect/
│   │   ├── Chameleon/
│   │   └── Llava/
│   │   └── ...
│   └── logs/
│   │   └── ...
├── model_output/
│   │   └── ...

---

📬 Contact

For any questions or issues, feel free to open an issue or contact us via lyuchengwu@zju.edu.cn / lyuchengwu@u.nus.edu or shumin@nus.edu.sg / 231sm@zju.edu.cn.

---

How to Cite

📋 Thank you very much for your interest in our work. If you use or extend our work, please cite the following paper:

@inproceedings{EMNLP2025_AutoSteer,
  author        = {Lyucheng Wu and 
                   Mengru Wang and 
                   Ziwen Xu and 
                   Tri Cao and 
                   Nay Oo and 
                   Bryan Hooi and 
                   Shumin Deng},
  title         = {Automating Steering for Safe Multimodal Large Language Models},
  booktitle     = {{EMNLP} {(1)}},
  publisher     = {Association for Computational Linguistics},
  year          = {2025},
  url           = {https://arxiv.org/abs/2507.13255}
}

---

🤝 Acknowlegement

We sincerely thank Xinpeng Wang and Donghyeon Won for providing parts of the datasets used in this project, which originate from their works: ToViLaG and protest-detection-violence-estimation.

We also gratefully acknowledge Dr. Peixuan Han for his open-source prober implementation from work SafeSwitch, which served as a valuable reference during our development.

---

📜 License

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