Efficient Reasoning via Thought Compression for Language-Guided Segmentation

ArXiv 2026 | [Paper] | [Code]

WISE is a framework for language-guided/referring/reasoning segmentation, which:

• 🚀 Accelerates inference by 5x through token reduction.

• 🧠 Compresses reasoning via self-distilled concise rationales.

• 🏆 Achieves zero-shot SOTA without verbose bottlenecks.

Installation

git clone https://github.com/mrazhou/WSIE.git
cd WSIE

conda create -n wise python=3.11
conda activate wise

pip install torch==2.5.1 torchvision==0.20.1
pip install -e .
pip install sam2 matplotlib

Training

bash training_scripts/run_qwen2_5_3b_refCOCOg.sh

Merge Checkpoint (optional)

python3 training_scripts/model_merger.py --local_dir [path_to_your_actor_checkpoint]

Evaluation

bash evaluation_scripts/eval_all.sh [path_to_your_actor_checkpoint]/actor

Note: The current code has been organized to some extent. Feel free to open an issue or contact me via email for updates and maintenance.

Results

Citation

If you find this repository helpful, please consider citing our paper:

@inproceedings{zhou2026efficient,
  title={Efficient Reasoning via Thought Compression for Language-Guided Segmentation},
  author={Zhou, Qing and Zhang, Shiyu and Jia, Yuyu and Gao, Junyu and Ni, Weiping and Wu, Junzheng and Wang, Qi},
  booktitle={arXiv preprint arXiv:2604.02040},
  year={2026},
}

and the Seg-Zero paper:

@article{liu2025segzero,
  title        = {Seg-Zero: Reasoning-Chain Guided  Segmentation via Cognitive Reinforcement},
  author       = {Liu, Yuqi and Peng, Bohao and Zhong, Zhisheng and Yue, Zihao and Lu, Fanbin and Yu, Bei and Jia, Jiaya},
  journal      = {arXiv preprint arXiv:2503.06520},
  year         = {2025}
}

Acknowledgments

Thanks very much to Seg-Zero, Qwen2.5-VL and SAM2 for their great work.