# ROMA: Real-time Omni-Multimodal Assistant with Interactive Streaming Understanding

📄 Paper🌐 Project Page🤗 Model📊 Dataset

--- ## 🔥 ROMA's streaming understanding capabilities

> It supports proactive tasks, including event alerts and narration, alongside reactive question answering. --- ## 📖 Introduction **ROMA** is a **real-time omni-multimodal assistant** for unified **streaming audio-video understanding**. Unlike traditional VideoLLMs that only respond to explicit queries, ROMA supports both: - **Reactive interaction** (question answering) - **Proactive interaction** (event alerts and narration) ROMA processes continuous streams as **synchronized multimodal units**, aligning dense audio signals with discrete video frames. A lightweight **Speak Head** decouples **response timing** from **content generation**, enabling the model to autonomously decide **when to speak** in streaming environments. The model is trained with a **two-stage streaming curriculum** and evaluated on a **unified benchmark suite covering 12 streaming tasks**. --- ## 🚀 Highlights - **Unified Reactive + Proactive Streaming Interaction** - **Streaming Audio-Video Understanding** - **Lightweight Speak Head for Response Timing** - **Two-Stage Streaming Curriculum Training** - **Evaluation Across 12 Benchmarks** --- ## 📊 Performance ROMA achieves **state-of-the-art performance on proactive streaming tasks** while remaining competitive on traditional reactive QA benchmarks. ### Proactive

### Reactive

### Omni-Modal Reactive

--- ## 🧠 Model | Model | Modalities | Capability | |------|-----------|-----------| | ROMA | Audio + Video + Text | Streaming Multimodal Understanding | Model weights are available at: 👉 https://huggingface.co/EurekaTian/ROMA --- ## 📊 Dataset The **ROMA Proactive Streaming Dataset** is released at: 👉 https://huggingface.co/datasets/EurekaTian/ROMA_proactive Dataset statistics: | Subset | Task | Samples | |------|------|------| | Event-Driven Alert | Proactive Monitoring | **27K** | | Real-Time Narration | Streaming Captioning | **109K** | | **Total** | | **136,193** | Source datasets include: - DiDeMo - OOPS - Charades-STA - COIN - YouCook2 - ActivityNet These datasets are reformulated into **streaming interaction formats** for training proactive multimodal assistants. --- ## 📂 Repository Structure ``` ROMA ├── data │ └── test_mix_data.json # example dataset format ├── eval # evaluation scripts ├── sh │ └── train.sh # training entry ├── requirements.txt # dependencies └── README.md ``` --- ## 🏋️ Training Install dependencies: ```bash pip install -r requirements.txt ``` Run training: ```bash bash sh/train.sh ``` Example data format: ```bash data/test_mix_data.json ``` --- ## 📈 Evaluation All evaluation scripts are provided in: ```bash eval/ ``` The evaluation covers: - Proactive Alert - Real-Time Narration - Reactive QA --- ## 📚 Citation If you find this work useful, please cite: ```bibtex @article{tian2026roma, title={ROMA: Real-time Omni-Multimodal Assistant with Interactive Streaming Understanding}, author={Tian, Xueyun and Li, Wei and Xu, Bingbing and Dong, Heng and Wang, Yuanzhuo and Shen, Huawei}, journal={arXiv preprint arXiv:2601.10323}, year={2026} } ``` --- ## ⭐ Acknowledgement If you find this repository helpful, please consider giving it a ⭐