REVERSE-VLM: Vision-Language Model with
REtrospective VERification and SElf-correction

   

---

Welcome to the official repository for our paper: Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025). Explore our project page here for an interactive overview!

Authors: Tsung-Han Wu, Heekyung Lee, Jiaxin Ge, Joseph E. Gonzalez, Trevor Darrell, David M. Chan (UC Berkeley & POSTECH)

🔗 Model Checkpoints:

• 🤗 tsunghanwu/reverse_llava_v15
• 🤗 tsunghanwu/reverse_llava_more
• 🤗 tsunghanwu/reverse_qwen25_vl

📦 Dataset:

• 🧾 REVERSE Visual Instruct 1.3M

📄 Change Log:

• [04/17/2025]: REVERSE is now live on HuggingFace and GitHub! Explore checkpoints, dataset, and full paper from our project site.
• [05/29/2025]: REVERSE supports Qwen2.5-VL and is also effective. Check it out!
• [09/18/2025]: REVERSE is accepted to NeurIPS 2025!!!

---

🔧 Installation Guide

1. Clone this repository

git clone https://github.com/tsunghan-wu/reverse_vlm
cd reverse_vlm

2. Set up the environment

• For LLaVA:

  conda create -n reverse python=3.10 -y
  conda activate reverse
  pip install --upgrade pip  # enable PEP 660 support
  pip install -e .
  pip install -e ".[train]"
  pip install flash-attn --no-build-isolation --no-cache-dir

• For Qwen series, please follow the installation guideline in Qwen2-VL-Finetune.

📈 Evaluation

• Download model checkpoints:

  • 🤗 reverse_llava_v15 (LLaVA-v1.5-7B style model)
  • 🤗 reverse_llava_more (LLaVA with LLama3.1-8B-Instruct style model)
  • 🤗 tsunghanwu/reverse_qwen25_vl (Qwen2.5-VL-3B-Instruct style model)

• Download required evaluation files from Google Drive
  → Unzip and place them into playground/data/eval. Then, follow the included instructions to download additional assets.

• Run evaluations with:

bash scripts/eval/*.sh

We conduct 100-round bootstrapped evaluation. Reported numbers should closely match those in the paper.

🚀 Training

1. Data Preparation

• Download QA pairs from: 🤗 tsunghanwu/reverse-instruct-1.3m

• Organize datasets under playground/data/ using the following structure (following LLaVA’s layout):

Show Data Structure

playground/data/
├── coco
│   ├── annotations
│   ├── test2017
│   ├── train2017
│   └── val2017
├── gqa
│   └── images
├── ocr_vqa
│   └── images
├── share_textvqa
│   └── images
├── textvqa
│   └── train_images
└── vg
    ├── VG_100K
    └── VG_100K_2
    

2. Model Setup and Training

• Add special tokens to the base LLM:

python3 scripts/add_new_token_to_llava.py
python3 scripts/add_new_token_to_qwen.py

Supported settings:

• LoRA finetuning for LLaVA-series

  • lmsys/vicuna-7b-v1.5, with mm_projector weights from LLaVA-v1.5-7B's projector
  • meta-llama/Llama-3.1-8B-Instruct, with mm_projector weights from LLaVA-MORE-8B's projector

• Direct finetuning for the Qwen2.5-VL model

  • Qwen/Qwen2.5-VL-3B-Instruct
  • To ensure the apple-to-apple comparison, we fine-tune the released Qwen2.5-VL-3B model using both the LLaVA-FT setup and our REVERSE recipe, applying both on the same 100k subset. This allows us to directly compare the impact of our training/inference recipe against the basic training/inference baseline under consistent conditions as the Qwen2.5-VL's instruction tuning data is not publicly available.

• Launch Training: bash scripts/train/*.sh

3. Merge LoRA Weights (for LLaVA series only)

After training, merge the LoRA adapter weights into the base model:

CUDA_VISIBLE_DEVICES=5 python3 scripts/merge_lora_weights.py --model-path <your lora path> --model-base <the base llm path with new tokens> --save-model-path <final model path>

⚠️ Notes:

• Set GPU_SETTINGS and MASTER_PORT appropriately when using DeepSpeed.
• Naming matters:
  • Your LoRA directory should contain llava_lora
  • The final merged model path should contain llava
    This is required due to how LLaVA loads models internally — otherwise, it may fail silently or load incorrectly.

🙏 Acknowledgements

We are grateful for the foundational code provided by LLaVA, LLaVA-More, and Fine-tuning Qwen2-VL Series. Utilizing their resources implies agreement with their respective licenses. Our project benefits greatly from these contributions, and we acknowledge their significant impact on our work.

📚 Citation

If you use our work or our implementation in this repo or find them helpful, please consider giving a citation.

@article{wu2025reverse,
  title={Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling},
  author={Wu, Tsung-Han and Lee, Heekyung and Ge, Jiaxin and Gonzalez, Joseph E and Darrell, Trevor and Chan, David M},
  journal={arXiv preprint arXiv:2504.13169},
  year={2025}
}