1:42会议进行中,随时更新任务
同时留意开电脑、记笔记和喝水,再按新指令切换为观察大笑。
观看完整演示The video could not load. Open the video directly ↗
1 Tsinghua University
2 Peking University
3 ByteDance
4 Beihang University
5 University of Wisconsin–MadisonWATCH · LISTEN · INTERACT
Live conversations and online videos.
01 / IN THE REAL WORLD
Live ALIVE with speech recognition and speech synthesis.
1:42同时留意开电脑、记笔记和喝水,再按新指令切换为观察大笑。
观看完整演示
0:54沿途识别椅子并报告颜色,收到取消指令后停止提醒。
观看完整演示
0:43在办公室巡看时,同时追踪沙发、饮水机、雨伞和笔记本电脑。
观看完整演示
0:26先请求打开冰箱,再根据新画面确认里面是否有可乐。
观看完整演示
0:35文字看不清时,请用户靠近,再识别瓶身上的 Vitamin C。
观看完整演示
0:29面对多个显示器,先澄清指代,再判断屏幕是亮的还是灭的。
观看完整演示
0:37Asks to keep watching, then reads the word when the whiteboard comes into view.
Watch demo
0:41屏幕关闭时请求展示内容,在解锁后读出文档标题。
观看完整演示02 / ACROSS ONLINE VIDEOS
ALIVE follows events, coordinates requests, and seeks feedback.
0:29同时追踪五类校园镜头,每类首次出现时提醒一次。
观看完整演示
0:13保留待办提醒,回答插入的问题,再在目标事件出现时继续通知。
观看完整演示
0:47Switches from watching handle removal to watching installation when the user changes the request.
Watch demo
0:11区分靠近与实际越过,在每处新采光顶上方通过时提醒。
观看完整演示
0:12等待取包动作完成,再检查更衣柜里是否还有遗留物品。
观看完整演示
0:14在计时器收好后检查桌面,提醒带走剩下的散页笔记。
观看完整演示
0:20Requests a flip to reveal the underside, then describes the browning and grill marks.
Watch demo
0:14Asks for the basket to be lifted out of the oil, then describes the visible fries.
Watch demo
0:30Asks permission to keep watching, then checks the top rack when the dishwasher becomes visible.
Watch demo
0:12柜门关闭时请求打开,在获得新画面后描述柜内物品。
观看完整演示
0:24Monitors both removal of the old filter and installation of the new one, with a reminder for each step.
Watch demo
0:16Asks the user to select the 1-Day Ventra Ticket, then reads its price from the updated screen.
Watch demoTHE NEXT STEP
Streaming models can answer questions and monitor events. ALIVE goes further: coordinating changing requests and seeking the feedback needed to proceed.
Streaming video LLMs can continuously observe the world, yet often handle user requests in isolation. We introduce ALIVE, an interactive streaming video model supporting proactive monitoring, streaming task coordination, and feedback-seeking interaction. We construct ALIVE-SFT-288K and ALIVE-RL-3K for two-stage training. Contrastive supervised fine-tuning teaches when and how to respond, while reinforcement learning with Behavior-Aligned Credit Assignment combines trajectory-level and local advantages to refine decisions, responses, and silence across multi-turn interactions. We also introduce FSI-Bench to evaluate feedback seeking and subsequent task completion. ALIVE achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench, demonstrating strong streaming understanding and interactive capabilities.
LEARNING TO INTERACT
Two stages teach a single model to choose an appropriate strategy, answer accurately, and remain silent when a response is unwarranted.
Similar visual observations can call for different behaviors. Controlled contrasts teach the role of available evidence, active requests, and user feedback.
Trajectory-level and local advantages refine behavior across multi-turn interactions, aligning credit with the requirements of each interaction.
DATA & EVALUATION
288,436 interaction examples for supervised fine-tuning.
3,523 interaction examples for reinforcement learning.
Four tasks test strategy selection and subsequent task completion.
FIVE BENCHMARKS
The highest overall scores among the compared models on four public benchmarks and FSI-Bench.
| Model | OVO-Bench | StreamingBench | ProactiveVideoQA | IPIBench | FSI-Bench |
|---|---|---|---|---|---|
| ALIVE | 68.9 | 76.0 | 51.6 | 40.3 | 52.8 |
| JoyAI-VL-Interaction | 60.6 | 67.4 | 28.8 | 20.2 | 17.4 |
| AURA | 65.3 | 73.1 | 32.0 | 21.6 | 8.3 |
| VideoChat3 | 58.8 | 68.7 | 37.6 | 12.0 | 5.9 |
OVO-Bench, StreamingBench and IPIBench: overall accuracy (%). ProactiveVideoQA: PAUC at ω = 0.5. FSI-Bench: four-task macro accuracy (%) under the Explicit Decision Protocol (EDP). Bars share a zero baseline and a 0–80 scale. These are the paper's reported comparisons, not a live leaderboard.

Continuously perceive. Sustain meaningful interaction.