---
license: mit
tags:
- computer-vision
- multi-task-learning
- vision-foundation-models
- mixture-of-experts
- icml-2026
datasets:
- imagenet-1k
- pascal-context
- nyud-v2
---

# PRISM Checkpoints

This repository hosts model-only checkpoints for **PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization**.

Paper: https://arxiv.org/abs/2606.03444

Code: https://github.com/robotyingtang/PRISM-VFM

## Files

```text
checkpoints/
  prism_stage1_vit_b.pth
  prism_stage1_vit_l.pth
  prism_stage2_pascal_vit_b.pth
  prism_stage2_pascal_vit_l.pth
  prism_stage2_nyud_vit_b.pth
configs/
  prism_stage1_vit_b.yml
  prism_stage1_vit_l.yml
  prism_stage2_pascal_vit_b.yml
  prism_stage2_pascal_vit_l.yml
  prism_stage2_nyud_vit_b.yml
```

## Usage

```bash
hf download robotyingtang/PRISM-VFM checkpoints/prism_stage2_pascal_vit_b.pth --local-dir pretrain/prism
```

```bash
python test_condition_moe.py \
  --exp prism_s2_pascal \
  --config_path configs/s2_prism/pascal_s2.yml \
  --checkpoint pretrain/prism/checkpoints/prism_stage2_pascal_vit_b.pth \
  --results_dir results \
  --evaluate
```

## Citation

```bibtex
@inproceedings{tang2026prism,
  title={PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization},
  author={Ying Tang and Dong Li and Youjia Zhang and Zikai Song and Junqing Yu and Wei Yang},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  year={2026}
}
```