README.md

metadata

title: ReToken Visual Retrieval
emoji: 🔍
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
short_description: VLM with learned retrieval token for visual retrieval
python_version: '3.12'
startup_duration_timeout: 30m

ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

This Space demonstrates ReToken-Qwen3VL-8B, a vision-language model augmented with a learned retrieval token for improved visual retrieval. The model is based on Qwen3-VL-8B and was trained with a single retrieval token that reads out the retrieval signal from the attention value space.

Usage

Upload an image and ask a question about it. The model will answer using its fine-tuned visual understanding capabilities.

References

• Paper (arXiv:2607.28627)
• GitHub Repository
• Model on Hugging Face