---
license: cc-by-nc-4.0
pipeline_tag: video-text-to-text
---

# ST-Evidence-7B

We propose **Evidence-Backed Video Question Answering** (E-VQA), a task where multimodal models are designed to jointly produce a semantic **textual answer** and associated spatiotemporal evidence. This evidence includes **temporal segments** and **dense, tracked object segmentation masklets**. A masklet is defined as a temporal sequence of object segmentation masks tracked over time.

This model is presented in the paper [Evidence-Backed Video Question Answering](https://huggingface.co/papers/2607.11862) (ECCV 2026).

ST-Evidence-7B was fine-tuned from UniPixel, which is built upon Qwen2.5-VL and SAM 2.1. UniPixel is a unified model designed to handle both video question answering and mask generation.

This was released for research purposes only, in support of the academic paper *Evidence-Backed Video Question Answering*.

## Resources

- **Repository:** [SalesforceAIResearch/EVQA](https://github.com/SalesforceAIResearch/EVQA)
- **Paper:** [Evidence-Backed Video Question Answering](https://huggingface.co/papers/2607.11862)
- **Benchmark (ST-Evidence):** [Salesforce/ST-Evidence-Bench](https://huggingface.co/datasets/Salesforce/ST-Evidence-Bench)
- **SFT Dataset (ST-Evidence-Instruct):** [Salesforce/ST-Evidence-Instruct](https://huggingface.co/datasets/Salesforce/ST-Evidence-Instruct)

## License

CC-BY-NC 4.0

## Citation

If you find this work useful for your research, please cite our paper:

```bibtex
@inproceedings{wang2026evidence,
  title={Evidence-Backed Video Question Answering},
  author={Wang, Shijie and Zhou, Honglu and Wang, Ziyang and Xu, Ran and Xiong, Caiming and Savarese, Silvio and Sun, Chen and Niebles, Juan Carlos},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}
```