ALIVE: Making Streaming Video Models Interactive
| Name | Last commit message | Last commit date |
|---|---|---|
| docs | ||
| Web_Inference | ||
| .gitignore 108 B | ||
| LICENSE 11.1 KB | ||
| README.md 7.0 KB | ||
| THIRD_PARTY_NOTICES.md 909 B |
ALIVE: Making Streaming Video Models Interactive
🤗 Hugging Face Daily Papers · #2 Paper of the day
Jinzhao Li1,3,* · Yuan Xu2,3,* · Jinxuan Wu4 · Yukun Wang1 · Yijia Lei1 · Hanyu Ji2 · Yinuo Chen1 · Zining Huang1
Chi Zhang1 · Hang Ye2 · Yiwu Zhong2 · Honglei Yan3 · Panwang Pan3 · Yizhou Wang2 · Yin Li5 · Miao Liu1,†
1 Tsinghua University 2 Peking University 3 ByteDance
4 Beihang University 5 University of Wisconsin–Madison
* Equal contribution † Corresponding author
An interactive streaming video model for proactive monitoring,
streaming task coordination, and feedback-seeking interaction.
Demo
▶ Watch the full demo · 4:38 · English + 中文 · Download MP4
More demos: Real-world live conversations ↗ · Web-video demos ↗
Overview
Streaming video LLMs can continuously observe the world, yet often handle user requests in isolation. ALIVE turns them into interactive assistants with three capabilities:
| Capability | What ALIVE does |
|---|---|
| 🔵 Proactive Monitoring | Monitor unfolding events and respond when the requested condition is met. |
| 🟣 Streaming Task Coordination | Maintain multiple requests and update active tasks as instructions change. |
| 🟠 Feedback-Seeking Interaction | Ask for clarification, request a concrete user action, or propose continued monitoring when more evidence is needed. |
Quick Start
Run the ALIVE browser demo locally with streaming speech recognition and speech synthesis on two GPUs: one for ALIVE and one shared by ASR/TTS (Linux, Python 3.12).
| Component | Hugging Face checkpoint |
|---|---|
| ALIVE | ALIVE-Team/ALIVE (ALIVE-8B) |
| ASR | Qwen/Qwen3-ASR-1.7B |
| TTS | Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice |
setup.py installs the isolated environments and downloads the ASR/TTS checkpoints; download ALIVE-8B separately:
git clone https://github.com/lijinzhao30/ALIVE.git
cd ALIVE/Web_Inference
python3.12 scripts/setup.py
huggingface-cli download ALIVE-Team/ALIVE --include "ALIVE-8B/*" --local-dir models/ALIVE
cp config.example.json config.json # set "model_path": "models/ALIVE/ALIVE-8B"
python3.12 start.py # open https://localhost:8090
See the web inference guide for details.
Data
We build ALIVE-SFT-288K for supervised fine-tuning and ALIVE-RL-3K for reinforcement learning.
Method
ALIVE is trained in two stages:
- Contrastive supervised fine-tuning teaches when and how to respond.
- Reinforcement learning with Behavior-Aligned Credit Assignment (BACA) refines decisions, responses, and silence across multi-turn interactions.
Benchmark
- We introduce FSI-Bench to evaluate whether streaming video LLMs, when the current context is insufficient to resolve a request, can identify the missing prerequisite, actively seek the required feedback, and complete the task.
- ALIVE achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench, demonstrating strong streaming understanding and interactive capabilities.
TODO
- Paper
- ALIVE-8B checkpoint
- Local web inference
- Benchmark and evaluation code
- Training data
Citation
@article{li2026alive,
title={ALIVE: Making Streaming Video Models Interactive},
author={Li, Jinzhao and Xu, Yuan and Wu, Jinxuan and Wang, Yukun and Lei, Yijia and Ji, Hanyu and Chen, Yinuo and Huang, Zining and Zhang, Chi and Ye, Hang and others},
journal={arXiv preprint arXiv:2611.21366},
year={2026}
}
License
Code is released under the Apache License 2.0. See third-party notices and asset credits.





