Update README.md

README.md

@@ -3,6 +3,24 @@ license: mit
 language:
 - en
 ---
+# TISDiSS: Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation
+
+Official implementation of **TISDiSS**, a scalable framework for discriminative source separation that enables flexible model scaling at both training and inference time.
+
+## 🏆 Highlights
+
+- **State-of-the-art Performance**: Achieves SOTA results on WSJ0-2mix, WHAMR!, and Libri2Mix datasets
+- **Dynamic Inference**: Adjustable Reconstruction block repeat times (N_re) at inference stage for performance-efficiency trade-offs without retraining
+- **Effective Training Strategy for Low-Latency Separation**: Training with more inference repetitions consistently improves shallow-inference performance, offering a practical solution for low-latency separation
+
+## 📄 Paper
+
+**arXiv**: [https://arxiv.org/abs/2509.15666](https://arxiv.org/abs/2509.15666)
+
+**Huggingface Paper**: [https://huggingface.co/papers/2509.15666](https://huggingface.co/papers/2509.15666)
+
+**Status**: Submitted to ICASSP 2026
+
 ## 🚀 Quick Start
 
 ```bash
@@ -111,4 +129,17 @@ You can modify the script to test different N_re values:
 for re in 3 6 8; do
     # Your inference commands here
 done
+```
+
+## 📚 Citation
+
+If you find this work useful in your research, please consider citing:
+
+```bibtex
+@article{feng2025tisdiss,
+  title={TISDiSS: A Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation},
+  author={Feng, Yongsheng and Xu, Yuetonghui and Luo, Jiehui and Liu, Hongjia and Li, Xiaobing and Yu, Feng and Li, Wei},
+  journal={arXiv preprint arXiv:2509.15666},
+  year={2025}
+}
 ```